* * *

In the second week of October 2025, MGX announced its participation in a $40 billion deal to buy Aligned Data Centers — the largest data centre transaction in history. Around the same time, the Abu Dhabi sovereign-backed investor was completing its $6.6 billion co-lead of OpenAI’s funding round at a $157 billion valuation, joining xAI’s $6 billion round, anchoring a $7 billion commitment to the Stargate AI infrastructure project, and reportedly preparing to join Anthropic’s $3–5 billion round at a $170 billion valuation.

In the same month, Wamda reported that the typical UAE founder closing a $20–40 million Series B was assembling the round from “smallish cheques from many investors rather than larger cheques from fewer and more strategic investors.” That is not a paraphrase. It is a direct quote from Khaled Talhouni, Managing Partner of Nuwa Capital, one of the largest dedicated regional VC firms, describing what it actually looks like to raise a UAE Series B in 2025.

These two events were celebrated locally in the same week, in the same publications, by the same people. They tell opposite stories about the same capital base.

The UAE’s startup capital base is structurally barbelled. There is more seed capital than the local demand for it. There is more sovereign capital than almost any economy its size could imagine. Between them is a $20–80 million vacuum — the stage at which regional companies turn into globally relevant ones — and the vacuum is the single most consequential structural feature of the ecosystem today.

This is not a story about anyone failing at their job. MGX is doing exactly what it was designed to do, brilliantly. Hub71 is doing exactly what it was designed to do, brilliantly. The early-stage funds — BECO, Wamda, Shorooq, Global Ventures, Nuwa, COTU — are doing what they were sized to do, professionally. The barbell exists because no one has been designed to fill the middle. Part 1 of this series argued that the UAE has confused being a Gateway with being a Factory. The missing middle is the specific reason the Factory hasn’t emerged: there is no domestic capital pool deep enough or patient enough to take a UAE-built company from regional product-market fit to global scale. So they leave.

* * *

The barbell, by the numbers

Three layers of capital. Three different scales. Three different theories of what the UAE is trying to build.

At the top is sovereign and quasi-sovereign capital. MGX launched in March 2024 with a target of $100 billion in AUM, is reportedly raising a further $25 billion structured fund, and has deployed across some of the largest AI deals in history: OpenAI, xAI, Anthropic, Mistral, Stargate, the BlackRock/Microsoft Global AI Infrastructure Partnership, and the $40 billion Aligned acquisition. Mubadala Capital has executed over 100 venture investments globally. ADQ, ADIA, and Khazna add further hundreds of billions of deployable capital. By any reasonable measure, the UAE has built one of the most sophisticated and globally relevant sovereign-investor stacks in the world.

At the bottom is seed and early-stage capital. Oraseya Capital was ranked the UAE’s most active investor for the second consecutive year by MAGNiTT. Dedicated regional VC firms — BECO, Wamda, Shorooq, Global Ventures, Nuwa, COTU, Plus VC — collectively manage roughly $2–3 billion. Hub71 received 3,100+ applications in 2024. In the first nine months of 2024, seed and pre-Series A rounds made up 42% of all UAE deal activity. The $1–5 million ticket band dominates regional deal flow. The bottom of the funnel is, if anything, overheated.

In between sits the empty middle. The $20–80 million ticket band where Series B and Series C rounds live in mature ecosystems. The 2024 UAE Series B ledger contains Huspy ($59M led by Balderton, a European firm), NymCard ($33M led by Mubadala, the rare UAE-led mid-market deal), and a handful of others. Merit Incentives closed a $28M Series B in February 2025. Astra Tech’s $500 million from Citi in late 2024 was debt, not equity. The two MENA mega-deals in H1 2025 were Tabby ($160M, Saudi-headquartered) and Ninja ($250M, Saudi-headquartered).

No publicly disclosed UAE-headquartered private technology company has been verified to have closed an equity Series C or Series D round in 2024 or 2025. Tabby’s Series E in February 2025 happened after the company had already relocated to Riyadh. That gap — not the headline funding total, not the unicorn count, not the StartupBlink rank — is the load-bearing fact about the UAE ecosystem today.

Even the optimistic data confirms the shape. MAGNiTT noted that the share of MENA Series A and B rounds exceeding $20 million jumped from 10% in H1 2024 to 42% in H1 2025. That is real progress on the upper-middle, but from a low base, and the absolute count of UAE-headquartered companies graduating to Series C remained in single digits. In Q3 2025, MENA funding reached $4.5 billion — of which $3.2 billion went to Saudi Arabia, $1.2 billion to the UAE. The middle is shifting south.

* * *

What UAE-based investors say when asked directly

The cleanest evidence that the missing middle is real comes from inside the ecosystem itself. UAE-based investors — people whose business depends on the ecosystem working — are increasingly on the record about it.

Khaled Talhouni, Managing Partner of Nuwa Capital (Dubai and Riyadh, ~$300 million AUM, 53 portfolio companies across the GCC and Egypt), in an Inc. Arabia interview in September 2024:

“Although there has been an increase in early-stage funds, we still see a shortage of capital in the ecosystem, especially when it comes to Series B+. Once they reach that stage, most founders struggle to raise and end up patching together the round — smallish cheques from many investors rather than larger cheques from fewer and more strategic investors.”

Amer Alaily, General Partner of BECO Capital’s newly launched $250 million Growth Fund (closed September 2025, targeting Series B to pre-IPO with average tickets of $20 million), at the close of the fund:

“Companies in the Gulf are achieving institutional scale, yet face limited access to dedicated growth capital. This fund gives us the flexibility to partner with the strongest emerging companies and support them through critical scaling phases toward potential exits.”

Note that BECO is the most-cited UAE VC of the last decade — the firm that backed Careem, Property Finder, and Kitopi at their earliest stages. When BECO’s growth-fund GP publicly states that Gulf companies “face limited access to dedicated growth capital,” that is not a market complaint. It is a market diagnosis from inside the market.

Mudassir Sheikha, Co-Founder and CEO of Careem — the only $3 billion+ UAE-built exit — at the Atlantic Council in November 2020, identifying the three high-level challenges founders should expect: access to funding, access to talent, and regional fragmentation. On the funding challenge specifically, in remarks paraphrased at the time:

“Companies still struggle to raise scale capital of around $10–20 million, which is essential for them to expand regionally. Series B and C challenges must be met before we see more exit options for businesses.”

Sheikha said this five years ago. The 2025 data suggests the diagnosis has not been resolved — it has migrated upward, from a $10–20 million problem to a $20–80 million one.

Fadi Ghandour, Executive Chairman of Wamda Capital and one of the most senior figures in MENA venture, has been on the record for years about why scaling is harder than it should be:

“Even though businesses are scaling, manoeuvring and finding ways to expand their presence across markets, it is so much more difficult than in the US. The entrepreneur has to worry about things they shouldn’t have to worry about. Unless we are able to open up and ease the movement of businesses from one country to another, you are not going to be able to get the quick scaling of companies.”

These are not outside critics. Talhouni runs one of the most active early-stage funds in the GCC. Alaily runs BECO’s growth vehicle, just launched specifically to address the gap he is describing. Sheikha built and sold the most successful UAE-built startup in history. Ghandour anchors the longest-running regional venture franchise. When the four most credible voices in the ecosystem all describe the same problem in the same terms, it is no longer a hypothesis. It is the consensus diagnosis.

* * *

What the sovereign capital is actually doing

Critics of the UAE’s strategy sometimes treat the outward deployment of sovereign capital as evidence of neglect of the domestic ecosystem. That is the wrong framing. MGX, Mubadala, ADQ, and ADIA are executing a third, legitimate strategy alongside the Gateway and the Factory. Call it the Portfolio strategy: purchase minority stakes in the global frontier of AI and infrastructure to buy national exposure to the next decade of technology.

On its own terms, the Portfolio strategy has worked spectacularly. MGX’s 2024–2025 deployment record is among the strongest of any sovereign investor in the world. It is positioned in OpenAI, xAI, Anthropic, Mistral, Aligned Data Centers, Databricks, Stargate, the global AI infrastructure partnership with BlackRock and Microsoft, and the TikTok USDS joint venture with Oracle and Silver Lake. Mubadala Capital led the $1.4 billion Crusoe funding round, taking its valuation past $10 billion in under a year. By any reasonable test of sovereign investment performance — access to top-tier deals, terms achieved, returns on capital, geopolitical positioning — the UAE is executing the Portfolio strategy at a level few peers can match.

The Portfolio strategy is not Factory infrastructure for UAE-built startups. It was never designed to be. MGX itself describes its mission as “enabling the AI fabric of the global economy” and building “one of the world’s most advanced AI ecosystems locally while partnering with global leaders.” The local-versus-global balance is roughly 5–10% local to 90–95% global on the publicly disclosed deal flow. That allocation is a deliberate choice, justified on portfolio-construction grounds. The criticism is not of the choice. It is of the conflation.

Sovereign capital deployed into US frontier AI does not fund a UAE Series B founder’s next round. It does not build the operating-talent base required to run a thousand-person product company from Dubai. It does not improve the GCC cross-border regulatory friction that doubles the cost of every regional expansion. These are different problems requiring different capital architecture. The Portfolio strategy is excellent at what it is. It is not, and was not meant to be, the missing middle.

The Saudi comparison clarifies the point. Saudi Arabia’s Public Investment Fund deploys at MGX-comparable scale globally — PIF has stakes in Lucid, Newcastle United, Uber, Nintendo, EA, and dozens of US infrastructure plays. But Saudi has also built, in parallel, a dedicated domestic capital architecture: SVC (Saudi Venture Capital), launched 2018, has now backed 59 VC funds and supported 900+ startups via a fund-of-funds structure. STV runs a $500 million growth fund with a domestic mandate. Sanabil (the PIF tech-investment arm) anchors regional growth rounds. Hassana, the pension institution that co-led Tabby’s Series E, is one of several Saudi institutionals that explicitly target regional scale-ups. Riyadh has built both the global Portfolio stack and the domestic Factory architecture. The UAE has built the first at a higher level and the second at a lower one.

* * *

Why this matters: the Saudi gravity is now structural

In Part 1 of this series, Tabby’s 2023 relocation to Riyadh was treated as the most visible single data point about the UAE’s scaling gap. The relocation looks even more rational when the capital architecture is mapped properly.

Tabby’s February 2025 Series E was co-led by Blue Pool Capital (Hong Kong) and Hassana Investment Company (Saudi PIF affiliate), with participation from Wellington Management (US) and STV (Saudi-mandated growth fund). Of the four lead and participating investors, three had Saudi or US institutional anchors. None was a UAE-domiciled growth fund of comparable depth. Tabby could have raised the round from outside the GCC entirely. It chose Riyadh because the largest concentration of customers, the most credible IPO venue, and the deepest pool of patient regional growth capital all sat in the same place — and that place was no longer Dubai.

Other UAE-built companies are now facing the same calculus. The Q3 2025 funding data shows the migration in real time: $3.2 billion flowed to Saudi Arabia, $1.2 billion to the UAE. Saudi mega-rounds (Tamara’s $2.4 billion debt facility, Hala’s $157 million Series B, Lendo’s $50 million debt facility) dominated the upper end. The UAE retained more deal count, but lost dollar weight at exactly the stage where companies cross into globally relevant scale.

This is not a critique of Saudi. Saudi built a domestic Factory by design, over a decade, with sovereign capital that was explicitly structured to flow inward as well as outward. The UAE built a magnificent Gateway and a sophisticated Portfolio. The Factory was never explicitly funded. The result is now visible in the monthly Wamda and MAGNiTT reports.

* * *

What a fix would look like

Three things need to change for the UAE’s capital architecture to support the Factory the marketing already describes. Each is a legitimate policy question. None requires the country to abandon the Gateway or the Portfolio strategies that have worked so well.

First, a domestic Series B/C fund-of-funds at scale. Saudi’s SVC has demonstrated that a sovereign-anchored fund-of-funds, deployed patiently over a decade, can transform a region’s mid-market capital depth. SVC has backed 59 VC funds and supported 900+ startups since 2018, with a clear domestic mandate. The UAE has no exact equivalent. The Dubai Future District Fund and Mohammed bin Rashid Innovation Fund operate at a smaller scale and a narrower mandate. A UAE fund-of-funds at $2–4 billion, deploying $200–400 million per year into domestic-mandate growth funds (BECO Growth, a successor Nuwa growth vehicle, and three or four new entrants), would close a meaningful share of the gap within a single fund cycle. The capital exists. The architecture does not.

Second, GCC regulatory integration that actually reduces cost. Ghandour’s point about “things entrepreneurs shouldn’t have to worry about” translates concretely into compliance overhead. A UAE company expanding to Saudi Arabia, Kuwait, Bahrain, Oman, and Qatar runs five separate regulatory projects — five licensing regimes, five payment infrastructures, five data-residency regimes, five sets of competition law. The cost of each is small individually and crippling cumulatively. A serious GCC single-market push on payments, e-commerce, fintech licensing, and data residency would do more for the Factory than any direct capital injection. This is a multilateral problem; it would need political will, not money. The UAE has the regional capital it would take to lead the conversation.

Third, an operator-talent pipeline from the existing portfolio. The single most replicable thing about Silicon Valley is the recycling of senior operators from one generation of successful companies to the next. The “Careem mafia” — the 600+ former Careem employees who have either launched startups or joined ventures as co-founders — is the most concrete UAE example of this dynamic. Multiplying that requires more UAE-built Careems, which requires capital that can take companies to $200 million in revenue. The talent pipeline and the capital architecture are the same problem viewed from different angles.

None of this is novel. Talhouni and Sheikha have been describing variants of these three requirements for half a decade. What is new in 2025 is that the gap between rhetoric and architecture has now produced its first major scaling-out event — Tabby — and the Q3 2025 funding data shows the trend continuing. The case for moving on the architecture, not just the marketing, has become harder to defer.

The BECO Growth Fund close in September 2025 is a useful first marker. A $250 million dedicated UAE-and-Saudi growth fund, with $20 million average ticket sizes, anchored by an established regional name. It is the first domestic vehicle structured explicitly for the missing middle. One fund of that size does not close the gap. Ten of them, anchored by a sovereign-backed fund-of-funds programme, might.

* * *

The question now is one of architecture

Part 1 of this series argued that the UAE has built a world-class Gateway and confused it with a Factory. Part 2 has tried to name the specific structural reason the Factory has not yet emerged: a capital stack with abundant capital at both ends and a vacuum in the middle, where regional companies turn into global ones.

The fix is not money. The UAE has more money than almost any economy of its size needs. The fix is allocation. A redirection of even 1–2% of MGX-scale annual sovereign deployment into a domestic-mandate fund-of-funds programme would represent the largest single-year change in regional venture capital architecture since the launch of SVC in 2018. That would not require abandoning the Portfolio strategy. It would require treating the Factory as an explicit policy goal rather than an implicit consequence of the Gateway.

The UAE has more capital than it needs. It has more founders than it needs. It has more sophisticated investors than it needs. What it does not yet have is the bridge between them — a domestic, patient, growth-stage capital base that turns regional product-market fit into global scale. The country that wants to be called the world’s startup capital has to build that bridge. Until it does, more of its best companies will keep crossing the one that already exists, which runs from Dubai to Riyadh.

Tabby’s relocation was the prologue. The Q3 2025 funding data is the first chapter. The next decade will be written by which of the country’s remaining unicorn candidates choose to stay — and that choice will be determined less by tax regimes or visa policies, which are already optimised, and more by where the Series C exists. On present architecture, the answer to that question is increasingly: not here.

That is the gap the UAE’s next phase of policy will have to close. The Gateway is built. The Portfolio is built. The Factory has not yet been started. The capital for it exists. The architecture for it has not yet been drawn.

* * *

This concludes the two-part series. Part 1 examined why the UAE’s Gateway has not yet become a Factory. Part 2 examined the capital architecture that would have to change for it to do so. Both pieces are diagnostic. The policy question is now the country’s to answer.

That is the diagnosis. It explains why the AI signal in the 2023-2026 labor data is so muted: there is no fresh shock, just an old one accelerating. But it does not answer the harder question. Is AI categorically different from prior general-purpose technologies, in a way that intensifies the bifurcation rather than continuing it?

This is the question on which the entire structural argument turns. If AI is just one more accelerant on a slope that was already steep, the policy response is conventional — retraining at scale, sectoral cushioning, wait for new equilibrium jobs to appear, the way they did after every prior transition. If AI is categorically different, the policy response has to be different, because the historical mechanism that produced new jobs is the mechanism under attack.

I will argue that AI is categorically different. Not because it is uniquely powerful — electricity was uniquely powerful too — but because it is the first general-purpose technology in history whose operating personnel is itself.

The structural feature that no prior GPT shared

Bresnahan and Trajtenberg’s standard definition of a general-purpose technology has three criteria: pervasiveness across the economy, scope for continuous technical improvement, and the generation of innovational complementarities. Electricity met all three. Internal combustion met all three. The computer met all three. AI meets all three.

What makes AI different is not on this list. It is a fourth feature the prior GPTs did not have.

Every prior general-purpose technology arrived with a labor-creating channel attached. Electrification spawned electrical engineers, electricians, line workers, and the entire grid-operation labor market. Internal combustion spawned mechanics, drivers, logisticians, the auto-repair industry. The computer spawned programmers, IT staff, network administrators, data analysts, an entire profession of system architects who did not exist before. Each technology required humans to operate it — at scale, in specialized roles, at wages that supported a middle class.

The reinstatement channel that David Autor and his coauthors describe operates through exactly this mechanism. New technology displaces some tasks; the same technology reinstates labor by creating new tasks where humans have comparative advantage. The advantage is the operating role: humans understand the technology in ways the technology itself cannot, at least until the next GPT arrives.

AI is the first general-purpose technology in history whose operating personnel is itself.

The claim sounds rhetorical. It is not. AI systems are now training other AI systems. Reinforcement learning from AI feedback has replaced significant portions of reinforcement learning from human feedback in the most recent frontier models. Agentic systems write code, run code, debug code, deploy code, monitor code, and roll back code, with diminishing human oversight at each step. The integration, fine-tuning, evaluation, and red-teaming layers that briefly created new categories of work — prompt engineer, RLHF labeler, AI integration specialist — are themselves being automated by the systems they were designed to support. Each generation of foundation model absorbs more of the operating personnel that the prior generation required humans to supply.

This is the disanalogy with Leontief in 1952. Leontief was wrong about computers obviating clerical labor because the computer required programmers, and the programmer was a new kind of cognitive worker the computer itself could not be. The complementary skills humans developed (programming, system design, data analysis) grew faster than the computer’s capability frontier. The complementarity gap stayed open for sixty years.

AI closes the complementarity gap in software, on a quarterly cycle. The skills humans develop to complement AI — prompting, evaluation, deployment, oversight — are themselves software, automatable on the same compute substrate. If the AI capability frontier advances faster than the human-skill-acquisition frontier, the comparative-advantage region available to labor shrinks monotonically. This is not a metaphor. It is a structural feature of an automation technology whose output and operating personnel are the same kind of thing.

Anton Korinek and Donghyun Suh formalize this in a 2024 paper that the AI policy community has not absorbed nearly enough. Their model: if automation proceeds sufficiently slowly, then there is always enough work for humans, and wages may rise forever. By contrast, if the complexity of tasks that humans can perform is bounded and full automation is reached, then wages collapse. But declines may occur even before if large-scale automation outpaces capital accumulation and makes labor too abundant.

The last sentence is the under-appreciated point. Wage collapse does not require full automation. It requires automation faster than capital accumulation absorbs the displaced. That is precisely the structural environment of 2026: AI capability advancing on a quarterly cycle, capital stock (data centers, GPUs, robots) expanding on a multi-year cycle, displaced labor with no obvious destination sector. The mathematical signature of this is real wages stagnating or declining for displaced cognitive workers while compute and energy prices rise. Stagflation of factor prices, not goods.

This is the categorical difference. AI does not just substitute for human labor at faster speed. It substitutes for the meta-skill — the capacity for novel problem-solving — that made human adaptation to prior automation waves possible. And it does so on a substrate (software) that closes its own complementarity gap.

What the candidate next rungs actually look like

Suppose the structural argument is correct. The cognitive frontier is under attack from a self-complementing technology. The historical reinstatement channel is weakening. Where do displaced workers go?

The honest answer requires applying a disciplined test. For any sector to be a real next rung — not just a residual employer — it must satisfy five conditions: humans must retain comparative advantage at it under further AI capability scaling; it must be scalable to absorb tens of millions of workers, not a thin elite; it must pay at levels comparable to displaced cognitive work; productivity gains must expand rather than contract employment in it; and it must remain AI-incomplete for at least 10-20 years.

Eight candidates are seriously in play. None pass all five conditions.

Skilled trades — plumbing, HVAC, electrical, construction — are the strongest candidate. Robotics lags AI by 10-20 years, and the Moravec paradox (sensorimotor skills are harder than abstract cognition) gives a real and durable buffer. The sector is large, the wages are moderate to high, and trades are credibly AI-incomplete for a decade or more. The problem is scalability for displaced cognitive workers. The trades cannot absorb tens of millions of laid-off paralegals, junior coders, and copywriters without massive retraining and severe wage compression. The blue-collar floor is real. It is not high enough to land displaced senior knowledge workers at their reservation wage.

Care and emotion work — eldercare, childcare, nursing, hospice, therapy — is the largest candidate by absolute scale. Health and long-term care employ about 10% of the OECD workforce. Demographic pressure in Japan, Germany, Korea, China, and the OECD generally makes the demand income-elastic. The problem is that outside credentialed roles (physicians, registered nurses), the wages are low and have been low for a generation. Care work cannot pay displaced cognitive workers what they lost.

Licensed accountability roles — physicians, attorneys, accountants, financial advisors, anyone whose signature carries legal weight — is the candidate I find most interesting and least durable. Roughly 22% of US workers held an occupational license as of the mid-2010s, up from about 5% in the 1950s. This is a real and substantial floor. But it is a floor that floats. Every “human must sign” rule is one regulatory decision away from being relaxed once the AI’s measured error rate falls below the human’s. This has already happened in radiology, where the FDA has approved hundreds of AI-enabled diagnostic tools. It is happening in legal e-discovery, accounting reconciliation, tax preparation. The political-economy floor buys ten to twenty years, not a hundred.

Wealth work and positional services — the gift wrappers, sommeliers, personal trainers, lifestyle managers, luxury hospitality staff that David Autor has called “wealth work” — is real, growing, and structurally inequality-amplifying. It scales with the size of the high-income consumer base, not with population. It is a rung, but it is not a level rung. It absorbs some displaced cognitive workers as servants of the cognitive workers who were not displaced.

Frontier research and creative work that AI itself opens up — the human in the loop on scientific discovery, the human writing the prompts that generate the next breakthrough — is genuinely AI-incomplete at the frontier. It is also, by definition, elite. It cannot absorb mass displacement.

Last-mile oversight of AI systems — the prompt engineers, the RLHF labelers, the evaluation specialists, the alignment researchers — is the rung the AI industry itself most often points to as labor’s destination. It is also, as discussed above, the rung being most actively automated by the very systems it was meant to supervise. Self-canceling.

Status competition and meaning-making — therapists, coaches, ritual specialists, cultural curators, religious leaders — is real and may be income-elastic in rich societies. It is small at the scale required to absorb displaced cognitive workers, and the most replicable forms of it (companion AI, therapy chatbots, generative content) are already being automated.

Unknown new categories we cannot yet name is the optimist’s last and most honest argument. About 60% of the jobs people held in 2018 in the United States did not exist as named occupations in 1940. The track record of new-category creation is robust enough that we should be humble about predicting its end. The problem is that the categories that have appeared since 1980 — social media manager, UX designer, data scientist, prompt engineer — have bifurcated to the extremes of the wage distribution and have not rebuilt a middle. The unknown-unknowns argument is unfalsifiable but the trend data inside it is going the wrong way.

The candidates do not pass independently. They pass only in combination: a heterogeneous portfolio of trades, care, licensed accountability, and wealth work that absorbs displaced cognitive workers via wage compression rather than rung-climbing. The mathematical signature of this absorption is exactly what part one’s mean-versus-median chart showed — a structural widening of the gap between mean and median income, with the middle hollowing and the extremes growing.

The next rung exists. It does not extend the ladder upward. It splays it sideways and downward.

The strongest version of the optimist case

This is the place where intellectual honesty requires more than caveat. The optimist counter-case is better than most readers know, and I want to give it room.

David Autor — the same economist whose 2024 paper documenting the post-1980 bifurcation is the empirical center of my entire argument — wrote a different paper that same year arguing the opposite case. The title is direct: “Applying AI to Rebuild Middle Class Jobs.” His claim is that generative AI is unique among recent technologies in that it democratizes elite expertise rather than codifying routine work. The previous wave of computerization (1980-2010) hollowed the middle by automating routine cognitive tasks. The AI wave can re-create a middle by letting non-expert workers do expert work: nurse practitioners doing diagnostic work, paralegals doing legal drafting, technicians doing engineering analysis, customer-service representatives resolving issues that previously required senior staff.

This is the exact opposite of the structural pessimism I just laid out. It says AI is not the ladder being pulled up — it is a new escalator up the ladder, accessible to people who never reached the top before.

The empirical evidence for Autor’s optimist case is not weak. A widely cited 2023 study of customer-support agents at a Fortune 500 software firm found a 14% average productivity gain from AI assistance, with a 34% improvement for novice workers and “minimal impact” on experienced and highly skilled ones — the productivity distribution compressing upward. A randomized controlled trial published in Science on 453 college-educated professionals doing writing tasks found the average time taken decreased by 40% and output quality rose by 18%, with gains largest for lower-ability workers. A study of GitHub Copilot on 95 professional programmers found developers completed an HTTP server task 55.8% faster than the control group, with effects favoring less experienced developers.

In each case the pattern is the same: AI augments lower-skilled workers more than higher-skilled ones, which compresses the wage distribution in favor of the middle. This is the augmentation channel Autor and Brynjolfsson have argued we should be steering AI toward. It is also the natural emergent behavior of current frontier models, which are trained to produce competent expert output and therefore most help the people closest to incompetent.

Why doesn’t this neutralize the structural argument?

Three reasons. First, the empirical studies cover narrow domains where the human floor remains expert oversight — call centers, business writing, code completion. They do not generalize to settings where AI handles the full workflow without supervision. The augmentation finding holds for AI-assisted humans; it does not hold for AI replacing humans, which is the structural pressure the pessimist case identifies. As enterprise AI moves from copilot mode to agent mode through 2026 and beyond, the augmentation studies become decreasingly relevant.

Second, the augmentation outcome depends on a specific policy choice — building and deploying AI in ways that complement rather than substitute for labor. Erik Brynjolfsson has called the alternative the “Turing Trap”: building AI that imitates rather than augments because the imitation economics are temporarily more attractive to capital. The incentives push toward imitation. No major economy has yet implemented the tax or regulatory framework that would change those incentives at the firm-decision level. The optimist case is conditional on a political-economy response that has not happened.

Third, even within Autor’s optimist scenario, the middle that gets rebuilt is not the same middle that was hollowed. The hollowed middle was middle-skill production and clerical work paying $50,000-$80,000 a year. The rebuilt middle is augmented-low-skill work paying $35,000-$50,000 a year. The wage compression that makes the optimist case work also means the middle re-forms at a lower wage floor than the one it replaced. This is consistent with what the empirical evidence shows: real wage compression, not real wage recovery.

The honest synthesis is that Autor’s optimist case is the strongest argument against the structural pessimism, that it is partially correct, and that it does not fully neutralize the structural case. Both Autors can be right. Generative AI can democratize expertise and attack the reinstatement channel. The middle can re-form and re-form at a lower wage. The bifurcation can continue and be partially offset by a new low-skill augmentation rung. None of these are mutually exclusive.

The bet the pessimist case asks you to make is that the augmentation channel is smaller than the displacement channel at the wage level required to support a middle-class life. The bet the optimist case asks you to make is the opposite. The empirical evidence through 2026 does not yet decide between them, but the direction of the post-1980 trend — bifurcation, not middle-rebuilding — is the prior I would not bet against without strong evidence to the contrary.

Where this argument is most likely to be wrong

Three places to watch for the structural case failing.

The first is capability stall. If AI capability gains slow — through compute-scaling exhaustion, data quality ceilings, alignment-induced caution, or just diminishing returns on the current paradigm — the displacement-reinstatement decomposition reverts toward 2000-2015 conditions and the structural argument is overstated. The pessimist case requires capability to continue advancing on something like a quarterly cycle. If that ends in 2027 or 2028, the rest of this analysis ends with it.

The second is robotics trajectory. The blue-collar floor depends on humanoid robots remaining expensive and unreliable while cognitive AI continues to scale. If Tesla, Figure, Unitree, or one of the Chinese humanoid manufacturers cracks unit economics by 2030 in a way that current capability projections do not anticipate, the physical-services rung collapses much faster than expected. The buffer Moravec’s paradox provided to the trades becomes much shorter than I have argued. Conversely, if humanoid robotics remains expensive for longer than expected, the structural argument understates the durability of the blue-collar floor and the trades become a stronger absorption mechanism than I have credited.

The third is political-economy response. The argument I have made assumes the augmentation-versus-imitation choice continues to default to imitation because of capital incentives. If the European Union AI Act, US worker-augmentation tax policy, or analogous interventions create binding incentives toward augmentation, the reinstatement channel reopens through deliberate institutional choice. Acemoglu and Johnson have made this argument forcefully in Power and Progress: progress depends on the choices we make about technology. The pessimist case requires that those choices not be made, or be made too late. Both are plausible and neither is certain.

The single highest-probability failure mode for my argument is the robotics one. The single highest-probability failure mode for the optimist case is that AI deployment moves faster from copilot mode to agent mode than the augmentation studies anticipate. Both deserve serious attention from anyone trying to construct a 10-25 year structural view.

The line the argument actually draws

The two-part argument I have made does not predict mass unemployment. It does not predict wage collapse. It does not predict that artificial general intelligence is imminent or that AI will replace all human labor within twenty years. Those predictions exist and they have made the AI debate worse than it needed to be.

What this argument does predict is more modest and more durable. The structural bifurcation of the US labor market that began in 1980 will continue. AI will accelerate it, not initiate it. The reinstatement channel that absorbed every prior wave of displacement will continue to thin, because AI attacks the channel itself rather than redirecting labor through it. New rungs will appear but they will not be vertical extensions of the ladder; they will be horizontal and downward extensions that absorb displaced cognitive workers via wage compression. The mean will continue to pull away from the median. The middle will continue to hollow.

The honest range of outcomes runs from “structural bifurcation continues at current pace, real wages for displaced cognitive workers fall by 10-15% relative to the high-skill professional class over the next decade” at the optimist end, to “automation outpaces capital accumulation, labor’s share of national income compresses meaningfully below current already-depressed levels, stagflation of factor prices emerges with cognitive-labor disinflation and compute-energy-robotics inflation” at the pessimist end. Both ends are inside the structural envelope the data supports. Neither end requires AGI. Neither end requires confident capability forecasts beyond 2027.

What sits outside the envelope, on the optimist side, is the standard reassurance that the ladder will rebuild itself. It has not been rebuilding for forty years. There is no specific mechanism by which AI would make it rebuild now, and the mechanism by which prior technologies rebuilt it — the labor-creating channel of an operating-personnel requirement — is the mechanism AI most directly attacks.

The escape valve is not closing because the technology is uniquely powerful. It is closing because the technology is, for the first time in history, its own operator.

That is the structural break. The ladder has a last rung. We are standing on it.

If this argument is right, the policy implications and the asset-allocation implications cascade. I will write about both, separately, in coming weeks.

Each of these is a fake of something LinkedIn was built to deliver. The post is a fake of human reflection. The selfie is a fake of professional presence. The award is a fake of recognition. The recruiter is a fake of opportunity. The job is a fake of a job.

That is the platform now. LinkedIn has become a marketplace in which every signal it was created to broker has been counterfeited at scale — and the algorithm rewards the counterfeit, because to an engagement-maximizing system, the difference between a real signal and a forged one is invisible.

What follows is a tour of the forgery, in five parts.

The jobs aren’t real

The most quantifiable form of decay on the platform is in the job listings themselves. A September 2025 analysis by ResumeUp.AI estimated that 27.4% of active US LinkedIn job postings are likely ghost jobs — listings posted with no real intent to hire the applicant. A MyPerfectResume survey of more than 750 recruiters found that 81% admit their employer posts ghost jobs. A separate 2024 Resume Builder survey, reported by CNBC, put the figure at 39% of hiring managers acknowledging the practice.

The motives are mostly mundane and occasionally sinister. Some companies post fake openings to maintain “growth” optics in front of investors. Others harvest resumes for future pipelines. The ugliest motive comes from the MyPerfectResume data: 62% of hiring managers admitted their companies post ghost jobs specifically to make current employees feel replaceable. The listing is not a recruiting tool. It is a management tool aimed at the people who already work there.

In tech, the rot runs deeper. Industry analysis applying BLS JOLTS sector data estimated that roughly 48% of open tech listings never result in a hire — meaning that close to half of what looks like demand on the platform is statistical mist.

The labor market itself now distorts. The Bureau of Labor Statistics reported 7.4 million open positions in June 2025 against only 5.2 million hires. That 2.2 million gap is what economists call “fluid vacancies” and what job seekers experience as silence. Policymakers read the openings number. The candidate reads the silence. Both are looking at the same chart.

The cost falls on people. Greenhouse’s 2024 State of Job Hunting report found that 72% of US job seekers say the hiring process negatively affects their mental health. A typical application takes 45 minutes of tailored work. When more than a quarter of those applications were never going to be answered, the math of effort against return collapses into something that looks like learned helplessness — but isn’t, because the candidates weren’t unqualified. The position simply wasn’t a position.

The profiles aren’t people

LinkedIn removed 200 million fake accounts in 2024 alone. Its own transparency reports show bot accounts on the platform rose from 21.5 million in H1 2019 to 83.4 million in H1 2025 — roughly quadrupling in six years, with growth of around 50% annually over the last three. Invalid traffic on the platform sits near 20%, meaning one in five interactions on a typical post comes from something that isn’t a person.

The qualitative version is more familiar than the numbers. Auto-DMs from “Senior Directors” you’ve never met. Connection requests from headshots that turn up on a reverse image search as a Vietnamese stock model. InMails that begin “Let’s explore synergies” or “Would love 15 minutes of your time” with no context of who you are or what you do. The pattern is automation of intent rather than automation of activity — outreach designed to register motion, not to communicate.

A flourishing arms race has now emerged in which AI-generated outreach from automated recruiter accounts meets AI-generated applications from automated candidate accounts. The bots are talking to the bots. Somewhere in the middle, a human writes a CV, applies to a ghost job through an ATS, and gets a rejection email written by a third AI.

The awards aren’t earned

The vanity-award industry is older than LinkedIn and has simply found in it a perfect substrate. The canonical case study is the American Biographical Institute, which from 1967 until its 2012 bankruptcy issued hundreds of “Man of the Year” and “Woman of the Year” awards at $195–$295 each — paid certificates dressed as recognition. The model never died. It migrated to LinkedIn-adjacent magazines like Finance Monthly Awards, CorporateLiveWire, and Top Choice Awards, all of which solicit nominees by email and then quietly invite them to underwrite their own win.

A boating-industry executive published an exchange that captures the dynamic with unusual clarity. After he declined to pay for an award nomination, the sales rep called back and offered him second place: “First place has been reserved by another magazine.” The category, the prize, and the runner-up slot were all for sale, in that order.

LinkedIn’s own native version of this is the “Top Voice” badge, awarded by a process the platform has never fully explained and which has produced, predictably, a thriving subculture of people who post daily inspirational platitudes in pursuit of a logo. The bar has fallen so far that the badge is now treated by many working operators as a near-negative signal — a marker not of expertise but of someone willing to spend hours per week performing for an algorithm.

The result is a feed in which everyone is, by their own description, “award-winning.” The word has been emptied. So has the audience’s ability to use it as a filter.

The posts aren’t written

LinkedIn now spends serious engineering resources fighting the content its own users produce. In January 2025 it published research on a 150-billion-parameter model called 360Brew, deployed gradually through that year, designed in part to detect and downrank AI-generated posts. The reported effect: AI-flagged posts receive roughly 30% less reach and 55% less engagement than human-written ones. LinkedIn’s editorial team trained the detector on specific tells, including the now-infamous “it’s not X, it’s Y” construction and recycled leadership platitudes.

The same engagement-maximizing logic that produces the AI-generated leadership post also produces the gym-selfie captioned “discipline is a mindset,” the wedding-photo carousel labeled “what marriage taught me about KPIs,” and the beach-pose shot with the caption “your network is your net worth.” The algorithm does not know the difference. Neither, increasingly, does the platform.

The canonical artifact remains Braden Wallake’s August 2022 post — the “crying CEO” selfie accompanied by the line “This will be the most vulnerable thing I’ll ever share,” published after he laid off staff. The backlash was instant and global, and yet the format it pioneered — public vulnerability as engagement bait — has only spread. Wallake was not an aberration. He was a template.

The mechanics trace back further still, to Josh Fechter, who in the late 2010s discovered that LinkedIn’s algorithm rewarded “read more” clicks. He invented broetry — one sentence per line, hook-driven, designed to bait the click — and the platform spent the next several years trying to suppress what he had taught it to reward. He was eventually banned. The style outlived him.

The credentials aren’t credentials

The smallest scam on the platform is also the most pervasive. LinkedIn’s “Education” field accepts whatever the user types. A two-week online certificate from Harvard Business School Online or MIT Professional Education becomes, on the profile, “studied at Harvard” or “MIT alumnus.” A weekend in Cambridge becomes a credential. The platform does not verify. It only stores.

The same logic governs job titles. “Founder & CEO” describes both a person running a $200 million company and a person who registered an LLC on a Friday afternoon to consult into their old employer. “Senior Advisor” describes both a former cabinet secretary and a friend who agreed to take a phone call. The platform standardizes the language of authority without verifying any of the claims under it.

This is, in the end, what makes the rest of the counterfeiting possible. If anyone can call themselves anything, then the recruiter, the founder, the award winner, and the thought leader are indistinguishable from their forgeries by inspection alone. The labor of separating real from fake is pushed entirely onto the reader, who has neither the time nor the tools to do it.

The case for staying

The honest counterargument is that none of this has actually killed LinkedIn, and may never. Microsoft reported $17.81 billion in LinkedIn revenue in FY2025, up 9% year over year. Premium subscriptions crossed $2 billion. Roughly a billion people still maintain profiles. Real hiring still happens. Real introductions are still made. The alternatives — Bluesky, sector-specific Slacks, the ruins of Twitter — have not reached the critical mass that would let any one of them replace LinkedIn’s function as a global professional identity layer.

All of that is true. It is also beside the point. The right question is not whether the platform is dead but whether the return on time invested in it has collapsed. By every available signal — the fake jobs, the fake profiles, the fake awards, the AI-generated feed, the unverified credentials — it has. Time spent on LinkedIn in 2026 produces meaningfully less useful information per hour than time spent on it in 2016. The platform is bigger. The signal is weaker.

That is what enshittification actually looks like in its terminal phase. Not the platform shutting down. The platform thriving financially while becoming useless to its users.

The business model is the decay

Cory Doctorow’s original framing of enshittification described a three-stage process: platforms first court users, then exploit those users to lock in business customers, then squeeze everyone to extract value for shareholders. LinkedIn has now visibly completed the cycle. The users were courted with free networking. The business customers — recruiters, advertisers, sales teams — were locked in with Talent Solutions, Sales Navigator, and Marketing Solutions, which together generate the bulk of that $17.81 billion. The third stage is now underway.

In October 2024 the Irish Data Protection Commission fined LinkedIn €310 million for processing user data for behavioral analysis and targeted advertising without a valid lawful basis under GDPR. The investigation had been open since 2018. A year after the fine, LinkedIn announced that starting November 3, 2025, member data — profiles, posts, resumes, endorsements, going back to 2003 — would by default be used to train Microsoft’s generative AI models. The setting was turned on without prior notice. Opting out would not retract data already ingested.

So here is the closed loop. The platform’s users wrote authentic, original content for fifteen years. That content trained the AI that now produces the slop they scroll past. The platform’s algorithm, built to maximize engagement, rewards the slop equally with the human work. The platform’s detector then downranks some of the slop, but cannot meaningfully distinguish a good human post from a good machine one. And the data your authentic post contributed to the training set has, in legal terms, already been ingested. You cannot get it back.

Your labor produced the noise that has made the platform you used worthless to you. That is not a decline. That is a business model.

What’s left

When every signal on a platform has been forged, the platform stops being a signal-broker and becomes something else. A directory, maybe. A weak verification layer. An advertising surface. A training set.

Most people will continue to use LinkedIn the way most people continue to use airports: as a necessary friction on the way to somewhere they actually want to be. The professional networks that will matter five years from now are forming somewhere else, in smaller rooms, with friction that filters. The signal is moving to where the counterfeiting hasn’t yet found a foothold.

If you can hear the difference between a real recommendation and a paid one, between a real recruiter and a bot, between a real award and an invoice, you already know what to do.

You stop counting impressions, and start counting the people who would still answer your call.

The standard story about AI and jobs treats November 2022 — the launch of ChatGPT — as the moment the cognitive frontier came under attack. Entry-level paralegal hiring fell. Junior copywriting markets contracted. Computer science graduates began posting on Reddit about sending 150 applications and getting four interviews. The narrative writes itself: AI is breaking the career ladder.

Here is the argument I want to make across two pieces: AI did not break the ladder. The ladder has been structurally bifurcating since around 1980, and AI is accelerating a forty-year trend rather than initiating a new one. That distinction matters because it tells you what to watch, what to dismiss, and what the real structural question is — which is not whether AI is taking jobs (the answer through 2026 is: at the margins, yes) but whether the historical mechanism by which new technologies created mass employment is still operating at all.

This first piece does the diagnostic work. It maps the ladder historically, names what changed in 1980, and shows why the cognitive frontier was already cracking before generative AI existed. The second piece will ask the harder question: given a self-complementing technology — the first general-purpose technology in history whose operating personnel is itself — what does the next rung even look like, if there is one?

The thesis, stated once, clearly: the cognitive frontier was the historical escape valve for displaced labor, the valve has been narrowing for forty years for reasons unrelated to AI, and AI is the first technology that attacks the valve itself rather than redirecting labor through it.

The ladder as it actually worked

Every economic-history textbook teaches the ladder: agriculture to manufacturing to services to knowledge work. The pattern is real. The timing is not what most people think.

Britain’s male agricultural employment share fell from roughly 60% in 1600 to 22% by 1841 and below 10% by 1901 — a two-and-a-half-century transition, not the compressed Industrial Revolution of folk memory. The United States replicated the pattern with a half-century lag: 60% in agriculture in 1850, 12% by 1950. Japan compressed it harder, going from 70% agricultural at the Meiji Restoration to 17% by 1970. Korea did it in two generations. China did it in one.

The thing every textbook gets right is that the destination sector mattered enormously. Lewis-style dual-sector transitions worked because agricultural surplus labor migrated to a sector with higher productivity and income-elastic demand. British weavers displaced by power looms eventually ended up in factories that paid more than agricultural labor had. American farmworkers displaced by tractors ended up in manufacturing that paid more than farming. Manufacturing workers displaced by automation in the 1970s and 1980s ended up in services that, for the educated subset, paid more than manufacturing.

The thing every textbook understates is that the destination existed because the technology created complementary tasks for humans. Electrification spawned electrical engineers, electricians, line workers, and the entire grid-operation labor market. Internal combustion spawned mechanics, drivers, logisticians. Computers spawned programmers, IT staff, network administrators, data analysts. Each general-purpose technology arrived with a labor-creating channel attached.

The economic historian David Autor and his coauthors have a careful term for this: reinstatement. Technology displaces some tasks; the same technology reinstates labor by creating new tasks where humans have comparative advantage. When the reinstatement effect equals or exceeds the displacement effect, total labor demand holds up. When it doesn’t, you get either lower wages or fewer jobs — depending on which way the labor market clears.

For most of the twentieth century, the reinstatement effect was robust. New work appeared roughly as fast as old work was automated. The ladder worked.

What changed around 1980

The empirical center of the argument I want to make is a paper most people outside academic economics have not read: Autor, Chin, Salomons and Seegmiller, “New Frontiers: The Origins and Content of New Work, 1940–2018,” published in the Quarterly Journal of Economics in 2024.

What they did is conceptually simple and methodologically painful. They built a map of how the actual content of jobs in the United States changed across eight decades, using detailed occupational descriptions, and tracked where new work emerged. The headline finding will sound dry until you sit with it: the locus of new-work creation shifted from middle-paid production and clerical occupations over 1940–1980 to high-paid professional occupations and secondarily to low-paid services since 1980.

In English: before 1980, when new jobs appeared, they appeared mostly in the middle of the wage distribution. Production work, clerical work, mid-skill technical work. After 1980, when new jobs appeared, they appeared at the ends of the distribution. High-paid professional and managerial work at the top; low-paid personal-service work at the bottom. The middle stopped generating new categories.

The pattern shows up cleanly in the employment data.

Employment share by occupation wage tercile, 1980–2020. Middle-wage occupations fell from 33.5% to 24% of US employment; both ends gained roughly equally.

This is the bifurcation. It is forty years old. It happened well before generative AI. It happened during the era of personal computers, the internet, and globalization — but those technologies didn’t break a previously-intact middle so much as they progressively eroded the channel through which the middle had been replenished.

The same paper makes a sharper claim that is the analytical heart of the matter. New work falls into two types: augmentation innovations, which expand what humans can do, and automation innovations, which substitute for what humans were doing. Their finding: augmentation innovations cause the emergence of new work and the growth of occupational employment, while automation innovations do not spur new work but do erode occupational employment.

The mix has been shifting toward automation for four decades. The reinstatement channel has been thinning.

The consequence shows up in income data. The ratio of mean to median household income — a clean mathematical signature of distribution mass moving from the middle to the top — has been widening since the late 1960s, with most of the divergence after 1980.

Caption: Mean and median US real household income, 1967–2024 (2024 dollars). The mean was 12% above the median in 1967; it is now 40% above. Source: US Census Bureau.

The income gap began widening before 1980 — the ratio was already at 1.21 by 1980, up from 1.12 in 1967 — and accelerated thereafter, hitting 1.40 by 2024. The occupational bifurcation (first chart) is the mechanism; the income divergence (second chart) is the consequence. They are the same phenomenon viewed from different angles.

If you pair this with the Acemoglu-Restrepo decomposition of US labor demand growth across the postwar period — they find that recent decades exhibit an acceleration in the displacement effect, especially in manufacturing, a weaker reinstatement effect, and slower productivity growth than in previous decades — you get a coherent picture. The ladder didn’t break in 2022. It started bending around 1980. The middle rungs began thinning. The top rung grew but became harder to reach. The bottom rung absorbed the spillover at wages that didn’t replace what was lost.

This is the structural setup AI walked into.

Why the diagnostic matters before you talk about AI

I want to dwell on this for a moment because it changes how the AI-and-jobs debate should be conducted.

If you believe the ladder was working fine until ChatGPT, then the policy debate is about cushioning a technological shock and waiting for new equilibrium jobs to appear, the way new equilibrium jobs appeared after the steam engine, the electric motor, and the personal computer. The historical analogy is reassuring. The policy response is modest — retraining, unemployment insurance, maybe some sectoral support.

If you believe the ladder has been structurally bifurcating since 1980, the question becomes different. You are no longer asking whether AI causes a temporary disruption. You are asking whether a technology that operates on cognitive tasks is hitting an already-narrowed reinstatement channel — and whether it widens or further narrows the channel.

The “this time is different” debate gets resolved differently in this framing. Critics of the AI-displacement thesis correctly point out that every prior technology was called categorically different by smart contemporaries, and most of those contemporaries were wrong. Keynes in 1930. Leontief in 1952. The Triple Revolution memo of 1964. Each predicted technological unemployment that did not materialize at the predicted scale.

But the rebuttal — previous predictions were wrong, therefore this one is also wrong — assumes the underlying conditions are stable. They are not. The reinstatement channel that absorbed Leontief’s clerical workers into new occupations was operating at full strength in 1952. It is operating at materially reduced strength in 2026. The same prediction made today is being made into a different structural environment.

What this looks like in the data right now

The 2023–2026 evidence is more ambiguous than either the AI-doom commentariat or the dismissive economist class will tell you.

On one side: a Stanford working paper from October 2025, using payroll data covering several million workers, found a 13% relative decline in employment for 22–25-year-olds in AI-exposed occupations since late 2022, with no equivalent decline for older workers in the same occupations. UK government data from 2024 showed 16–24-year-olds in computer programming employment down 44% in a single year. Entry-level professional hiring at large firms has visibly contracted.

On the other side: a Yale Budget Lab analysis through Q1 2026 finds no statistically significant AI effect on aggregate unemployment. A Danish study of LLM adoption found precise null effects on earnings and recorded hours. A widely cited MIT report on enterprise AI deployment found that 95% of corporate generative-AI pilots delivered no measurable profit impact. The Economic Innovation Group has argued — credibly — that most of what looks like AI displacement is actually the lagged effect of the sharpest monetary tightening cycle in four decades.

The honest read of the 2023–2026 data is that AI is producing concentrated displacement at the entry level of cognitive work without yet producing aggregate disruption. The signal is real but small. The bifurcation hypothesis predicts exactly this: the top of the cognitive ladder under direct attack, the middle continuing its long-running hollowing, the bottom absorbing some spillover at compressed wages.

The two-and-a-half years of evidence we have is not enough to settle the AI-specific question. It is enough to confirm that the broader bifurcation is continuing and that AI is, at minimum, not reversing it.

What this is not an argument for

This is not an argument that AI is causing mass unemployment. It is not. Through Q1 2026 there is no aggregate unemployment shock attributable to AI in any major economy. Even the most aggressive estimates of AI-attributed layoffs in the United States — roughly 55,000 of 1.2 million job cuts in the first eleven months of 2025, per Challenger Gray and Christmas — represent under 5% of the total. The remaining 95% is rate cycle, post-pandemic correction, federal hiring freeze, and immigration policy reversal.

This is also not an argument that the ladder model was always wrong. It worked for two centuries. It moved millions of British peasants into factories, then into clerical jobs, then into professional work. It moved American farmworkers into manufacturing into services. It is one of the most empirically robust patterns in economic history.

What this is an argument for is that the ladder has been structurally weakening for forty years, that the AI debate makes more sense as a continuation of that weakening than as a fresh shock, and that the question worth asking is no longer “will AI take jobs?” — it will take some, at the margins, at the entry level. The question worth asking is: given that the reinstatement channel has been thinning since 1980, what happens when a technology arrives that targets the cognitive frontier itself and operates on a faster diffusion cycle than any prior general-purpose technology?

That is the question for part two.

Where this leaves us

The historical ladder mapped: real, robust, multi-century. The structural break around 1980: real, well-documented, pre-AI. The bifurcation pattern since: empirically established, accelerating. The AI-era data: thin but directionally consistent with continued bifurcation.

What we have not yet addressed — and what makes the AI argument different from a generic “post-1980 polarization” argument — is whether AI is categorically different from prior general-purpose technologies in a way that intensifies the bifurcation, or whether it is just one more accelerant on a slope that was already steep.

The argument for categorical difference rests on a single structural feature: AI is the first general-purpose technology in history whose operating personnel is itself. Electricity needed electricians. Computers needed programmers. AI agents can, in principle and increasingly in practice, train, monitor, fine-tune, deploy and debug other AI agents. The historical reinstatement channel — humans gainfully employed operating the new technology — is the channel AI most directly attacks.

That is the structural argument. Whether it survives serious scrutiny, and what it implies for the next rung — if there is one — is what part two is for.

Part two next week: why AI is the first self-complementing general-purpose technology, what the candidate next rungs actually are, and the strongest version of the optimist counter-case (which is better than most readers think).

The diagnosis explains where we are. It does not explain why we’re stuck.

That is the work of Part 2. Stagflation is the regime that develops when the standard policy response to a downturn is unavailable. Whether a supply shock turns into a stagflation regime, or resolves like a normal recession, depends entirely on whether the central bank and the fiscal authority retain the room to respond. In 2001, both did. In 1973 and 1979, neither did. In 2026, neither does — and the reasons why they don’t are structural rather than cyclical. They will not change with the next data release or the next election.

Six interlocking constraints. Each one would matter on its own. Together they form a closed policy box.

Constraint 1: Productivity has stopped doing the work

The single most underrated piece of the stagflation diagnosis is what’s happening in productivity and unit labor costs. The 1970s stagflation was, fundamentally, a productivity story — wage gains outpacing output gains, embedding inflation directly into unit costs. The disinflation of the 1980s and 1990s, by contrast, was a productivity-led story — output per hour grew faster than compensation, unit labor costs fell, and the inflation impulse subsided.

The Q1 2026 productivity data, released by the BLS on May 7th, has a stagflation signature underneath the surface that few have flagged.

Headline productivity growth came in at 0.8% on the quarter (annualized), which sounds fine. But the composition is the story:

• Output per hour: +0.8% Q1, but only +0.7% in hours worked, with output at +1.5%

• Hourly compensation: +3.1% Q1 annualized

• Real hourly compensation: −0.5% Q1 — workers losing ground in real terms

• Unit labor costs: +2.3% Q1 — running above the Fed’s inflation target

• Labor’s share of nonfarm business output: 54.1% — the lowest reading since the BLS began tracking it in 1947

That last data point is genuinely historic. Workers are taking home a smaller share of output than at any point in the post-war era. The 65% labor share of the 1950s and 1960s — when stagflation was about to arrive — is gone. The economy is producing more per worker, but the gains are accruing to capital, not labor. That sounds like a margin-friendly outcome until you realize what it means for the inflation transmission: workers cannot absorb price increases through real wage gains, because real wages are negative. They can only absorb them through reduced consumption — which is precisely what the K-shape and credit-stress data showed in Part 1.

The deeper concern is what the productivity numbers hide. The Indeed Hiring Lab’s analysis of the Q1 release puts it cleanly: while labor productivity rose 2.9% year-over-year, total factor productivity — the cleaner measure of genuine efficiency gain — actually decelerated in 2025, from 1.5% to 0.8%. The productivity gains we are seeing come from capital investment, particularly AI capex, not from underlying efficiency improvement.

That’s an enormous distinction. Capital-driven productivity gains are sustainable only as long as the capital cycle continues. If the AI capex pulse decelerates — and the bond-market pressure on hyperscaler issuance is one signal that it might — the productivity tailwind that has been masking unit labor cost pressure will disappear with it.

This is the productivity setup the 1970s would recognize: real wages negative, labor share collapsing, unit labor costs rising, headline productivity propped up by capital spending rather than efficiency. The appearance of productivity health hides an underlying productivity problem that will surface the moment capex slows.

Constraint 2: Fiscal dominance has removed the second policy lever

In a normal recession, the federal government can stimulate. Tax cuts, spending increases, transfer payments — these have backstopped every postwar recovery, often more decisively than monetary policy. The 2001 recession resolved partly because Bush could deliver tax cuts and Iraq war spending into a budget that was running a 2.3% surplus. The 2008 recovery used $800 billion of ARRA stimulus into a deficit that, while elevated, did not yet threaten Treasury financing. The 2020 pandemic response was the largest fiscal expansion in peacetime history, and it broke the back of the COVID recession in 18 months.

That lever is not available in 2026.

Federal debt held by the public reached 101% of GDP this year — surpassing the previous all-time high of 106% set in 1946 in CBO’s projection path. By 2036, the same projection takes debt-to-GDP to 120%. The federal deficit ran 5.8% of GDP in 2026: a peacetime record, larger than the deficits the United States ran during World War I or the Vietnam War. Net interest costs sit at 19% of federal revenue today, on a path to 28% over the next three decades. All three major rating agencies have already downgraded U.S. sovereign credit, explicitly citing persistent deficits, the rising interest burden, and political gridlock.

Deutsche Bank named the regime directly in a March 2026 note:

US sovereign debt has hit levels where interest expense is becoming a primary driver of the deficit. In a ‘Fiscal Dominance’ regime, the Fed’s ability to aggressively hike rates to curb inflation is constrained, as doing so risks a fiscal or financial crisis. Such an environment often encourages higher-for-longer inflation.

Yellen, in her January Brookings remarks, came close to naming the same problem in slightly different words. The combination of persistent deficits, rising interest costs, and political inability to address either creates a configuration where the Treasury’s financing needs constrain the Fed’s monetary autonomy. The Fed can no longer hike rates aggressively without risking a fiscal crisis.

The implication is profound. Under fiscal dominance, the inflation rate is no longer a pure monetary variable. It is set partly by the bond market’s tolerance for further Treasury issuance, by the political constraint on revenue-raising or spending cuts, and by the Fed’s increasing reluctance to drive real rates high enough to crush demand because of what that would do to debt service. Each of those is a structural, not cyclical, constraint.

This is the second lock on the policy box. Even if the Fed wanted to break the stagflation regime through aggressive tightening, fiscal arithmetic forbids it. And the fiscal authority cannot stimulate through a downturn because the deficit is already at peacetime records and the bond market is watching.

Constraint 3: The Fed itself has just been compromised

Kevin Warsh was confirmed as Federal Reserve Chair on May 14th by a Senate vote of 54–45 — the most partisan Fed-chair confirmation in the institution’s history. His confirmation followed an extended pressure campaign from the White House against the previous chair, Jerome Powell, including a Department of Justice investigation of the central bank. Powell stayed on the Board of Governors but ceded the chair on a vote that was almost completely along party lines.

Warsh inherits the position at roughly 3.5–4.0% Fed Funds, with core PCE running near 3% and rising, inflation expectations breaking out on every major survey, and markets pricing approximately zero cuts for 2026 with non-trivial odds of a hike. He inherits an institution whose credibility is, for the first time in the post-Volcker era, a market variable in its own right.

The institutional damage may be larger than the personnel change. Markets are now pricing in a permanent risk premium for the possibility that future Fed chairs face similar pressure. Even if Warsh acts with complete competence and orthodoxy, he cannot deliver the credibility profile that Powell, Bernanke, Greenspan, or Volcker held. A central bank operates on accumulated reputation, and the Powell pressure campaign has expended a significant fraction of that reputation.

The 8–4 dissent at the April FOMC meeting — the largest committee split since October 1992 — is the market’s read of where the institution sits. Dissents at that scale historically precede a leadership change. Markets read the configuration as a fractured committee about to be led by a chair with limited political capital.

The result: the Fed has perhaps 100–150 basis points of credible easing room before something breaks in the bond market. That is not enough to handle a wealth-shock-driven economic slowdown, and it is dramatically less than the 550 basis points Greenspan deployed against the 2001 recession or the 500 basis points Bernanke deployed against 2008.

In any previous post-1980 economic downturn, the Fed could ease aggressively. This time, it can’t — or at least not credibly enough to matter. That is the single most consequential sentence in this whole picture. Everything else — the K-shape, the tariffs, the energy shock, the fiscal arithmetic — flows from or interacts with that one constraint.

Constraint 4: Housing — the missing recovery channel

Of all the differences between 2001 and 2026, the absence of the housing channel may be the most underrated. It is the lever Greenspan pulled to turn the dotcom recession from a tech-sector capex unwind into a 9-month, V-shaped recovery. It is not available in 2026.

Recall the 2001 mechanism. As Greenspan cut from 6.5% to 1.0%, mortgage rates collapsed from roughly 8% to below 6%. That triggered a massive refinancing wave: households extracted equity, monthly payments fell, the wealth effect for the middle class ran in reverse of the equity-market wealth effect. By 2003, mortgage equity withdrawal was running at over $700 billion annually — a stimulus larger than any explicit fiscal program. The housing boom that followed pulled the US out of recession and into the longest residential construction expansion in postwar history.

The 2026 setup is the inverse. Freddie Mac’s PMMS reports the 30-year fixed mortgage rate at 6.36% as of May 14, 2026, almost flat over the past year. The National Association of Realtors notes that middle-income buyers can currently afford only 21% of the homes available for sale — versus roughly 50% before the pandemic. Home prices have appreciated 16% since the start of 2022 despite mortgage rates moving from 3% to 7%. The affordability problem is structural.

This matters in three specific ways for the stagflation thesis:

First, the refinancing channel is closed. The majority of US mortgage holders have rates below 5%, locked in during the 2020–2021 window. They cannot refinance into current rates without making their housing situation worse. The Fed cannot deliver mortgage-rate stimulus by cutting because mortgage rates are not primarily a function of the Fed Funds rate — they are a function of the 10-year Treasury yield, and the 10-year is anchored by inflation expectations and Treasury issuance, neither of which the Fed can directly control without monetizing debt.

Second, household formation is collapsing for the cohort that would normally power housing-led recovery. The under-35 demographic that historically buys first homes during recoveries cannot afford to. They are the same cohort showing up in the student-loan default data from Part 1. They are not going to drive a housing recovery; they are going to delay household formation further.

Third, the regional divergence in housing is itself a K-shape story. The Northeast and Midwest — where prices are still rising — are largely the regions with established wealth. The South and West — where prices are softening as pandemic-era migration unwinds — are where new household formation was supposed to drive recovery. Realtor.com’s 2026 top-housing-markets list now reads Hartford, Rochester, Worcester — older, higher-cost regions. Texas and Florida have rolled over. The geography of housing matches the geography of inequality.

The implication for the stagflation regime is straightforward. Even if the Fed could cut rates aggressively — which it cannot — the housing channel that pulled the US out of 2001 would not respond. The middle class cannot lever up into a refinancing wave because rates remain elevated. The young cohort cannot enter the market because affordability is at multi-decade lows. The wealth effect from housing, which traditionally backstops the equity-driven wealth effect at the top, is structurally broken.

This is the fourth lock on the policy box. The Fed’s transmission mechanism to the real economy has been narrowed to one channel: equity markets. And that channel runs only in the direction of supporting top-decile consumption, not in the direction of broad-based recovery.

Constraint 5: The global picture provides no disinflationary cushion

Every postwar US inflation episode has resolved partly through a global disinflationary tailwind. The 1980s saw the China opening and the broader Asian export-led growth model that compressed global goods prices. The 1990s saw NAFTA, the WTO accession of major emerging markets, and the broader hyperglobalization that drove core goods CPI persistently negative through the 2000s. The 2010s saw the post-GFC deflationary overhang in Europe, Japan’s continued deflation, and China continuing to export disinflation through manufactured goods.

The 2026 backdrop is different on every front.

China is no longer exporting deflation through the US supply chain. The tariff regime has explicitly designed this out. Effective tariffs at 11% — the highest since 1943 — mean that even if China’s domestic prices fall, the prices US consumers pay rise. The disinflationary channel has been closed by policy. And the deeper issue is structural: deglobalization is a multi-year process, not a single shock. Supply chains are being reshored at higher cost. Trade routes are being rerouted around political fault lines. Each of these is a permanent, not transitory, addition to the goods-price level.

Europe is running its own energy-driven stagflation. The post-Ukraine energy reset means European industrial input costs are structurally higher than they were pre-2022. German manufacturing remains in protracted contraction. The ECB has the same supply-shock problem the Fed has — and less monetary credibility to address it. European demand is therefore not the disinflationary force on US tradables that it was through the 2010s.

Japan has finally exited deflation. After thirty years of acting as a global disinflationary anchor, Japan’s CPI is running above 2% and the BOJ is normalizing policy. The yen carry trade, which provided a persistent source of cheap liquidity to global asset markets, is being unwound. One of the deepest structural disinflationary forces in the global economy has reversed.

The dollar is providing no support. Despite the geopolitical backdrop — Iran war, safe-haven demand, US rate differentials — the DXY sits around 99, roughly 10% below its January 2025 peak above 109. The Cambridge Currencies framing from May puts the structure neatly: “a Fed that can’t cut (inflation too high) and can’t hike (growth too soft), while the rest of the world quietly moves reserves away from US assets.” The dollar’s reserve status is a slow trend, not an imminent collapse — its share of global FX reserves has declined from 71% in 2000 to 57% in Q3 2025 — but the direction is unambiguous. A weaker dollar adds to imported inflation. The currency cushion that worked in prior US inflation episodes is not working in this one.

The implication: there is no external disinflationary force coming to rescue the US inflation regime. China, Europe, Japan, and the dollar each provided a disinflationary tailwind in prior cycles. None of them are providing it now. The supply-side inflation must therefore be resolved domestically — through tightening that the Fed cannot deliver, or through demand destruction that crushes the K-shape from the top down.

Constraint 6: The commodity supercycle is a structural inflation floor

The final lock on the policy box is the one most market commentary has noticed but underweighted: a commodity supercycle is underway, and it is structurally inflationary.

Gold above $5,000 is the cleanest market price for this. Goldman Sachs has a $5,400 year-end target; some analysts are calling for $7,000. But the gold price is not really about gold — it is the market’s pre-existing verdict on monetary debasement, fiscal dominance, and the loss of central-bank credibility. The same forces showing up in the inflation expectation data are showing up in the gold price. They are the same regime.

Jeff Currie — who called the 2020 commodity supercycle while at Goldman and has the track record to back the framing — laid out the current setup in a May 2026 thread that has been widely shared:

The QCI Commodity Total Return Index is up 217% since October 2020. The S&P GSCI Total Return is up 205%. Gold is up 140%. The Nasdaq 100 trails at 130%. The S&P 500 is up 85%. Welcome to the most asymmetric trade in modern financial history. At $105 Brent for 2026, the Munificent 7 oil majors generate 15.5% free cash flow yield on a multiple near seven times. The Magnificent 7 technology names generate roughly 1.5% free cash flow yield on a multiple closer to 28 times. The Magnificent 7 is the bid for molecules, electrons, copper, water, gallium, and concrete.

The structural drivers Currie identifies are the same structural drivers underpinning the stagflation case: deglobalization, electrification, and what he calls “redistribution” — the synchronous fiscal bonanza underway in the United States, Germany, Japan, and China. Each is a multi-decade trend, not a cyclical pulse.

The supply side reinforces this. Refinery investment is at a 10-year low. Upstream oil and gas investment is down 35% from its 2015 peak. The top 20 mining companies are spending 40% less than they did at the 2012 cycle high. A copper mine takes 10–16 years from discovery to first production. The International Copper Study Group projects a 150,000-ton structural deficit for 2026; Wood Mackenzie projects a 304,000-ton refined copper deficit for 2025–26. The supply response cannot arrive on the timescale that would matter for the current inflation regime.

The World Bank’s April 2026 Commodity Markets Outlook is direct: overall commodity prices are forecast to rise 16% in 2026, driven by soaring energy and fertilizer prices and record-high prices for several key metals. Energy prices are projected to surge 24% to their highest level since Russia’s invasion of Ukraine in 2022.

This is not a cyclical commodity boom. It is the early-to-middle innings of a structural supercycle driven by:

• Decade of underinvestment in mining and energy extraction during the 2014–2025 “exploitation phase”

• Exploding AI-driven demand for copper, silver, uranium, and electricity

• Energy transition demand for industrial metals on a multi-decade horizon

• Geopolitical fragmentation creating supply concentration and strategic-asset behavior

• Synchronous fiscal expansion across the four largest economies

• Central bank gold accumulation at the fastest pace in 50 years, signaling FX-reserve rotation

The implication for the stagflation case is the cleanest of all six constraints. Commodities are roughly 6% of headline CPI directly and significantly more through indirect input costs across food, transportation, manufacturing, and services. A structural 16% annual rise in commodity prices, partially passed through to consumer prices, sets an inflation floor of perhaps 1.5–2.5% on top of whatever else the economy is doing. That floor is independent of monetary policy, demand conditions, or business cycle dynamics. It is, in the deepest sense, the supply-side bedrock of the stagflation regime.

Why the dot-com analogy gets the bubble right and the resolution wrong

Many strategists are reaching for a 2000–2002 comparison to think about the current setup. The surface analogies are real and worth respecting. Tech capital expenditure is the marginal driver of GDP, just as telecom and dotcom capex was in 1999–2000. Information processing equipment contributed 0.8 percentage points to Q1 2026 growth — a super-cycle contribution rather than a normal one. The Mag-7 sits in the mid-30s as a share of S&P market cap, comparable to tech’s peak share in March 2000. The hyperscalers’ investment-grade debt issuance rhymes with the late-1990s telecom bond binge that financed the fiber overbuild. And the wealth-effect-driven consumer is the same engine: household consumption held up by capital gains in a small set of equities.

At the bubble-mechanics level, the parallel is sound. At the resolution level, it falls apart — and the way it falls apart is precisely the recession-versus-stagflation distinction from Part 1.

The 2001 episode was a recession. It followed the textbook recession playbook. CPI averaged 3.4% in 2000 and ran below 2% through most of 2002–2004 — the bust was disinflationary. Demand collapsed, oil briefly fell, China entered the WTO and inaugurated the peak globalization disinflation cycle. Greenspan entered the 2001 recession with the Fed Funds rate at 6.5% and cut 550 basis points over 24 months, to 1.0% by June 2003. The U.S. ran a 2.3%-of-GDP surplus in 2000, so Bush could deliver tax cuts and Iraq war spending without consequence. Household debt-to-disposable-income was around 96% and consumer credit was healthy. Unemployment was 3.9% in March 2000 with strong payroll momentum.

Every one of those policy and structural cushions is absent in 2026. CPI is sticky at 3.8%; the bust will be inflationary, not disinflationary. Warsh has perhaps 150 basis points of credible easing room. The deficit is 5.8% of GDP, debt-to-GDP is 101%, three credit-rating downgrades are in the bag. Household debt is at $18.8 trillion record, with student-loan delinquency at 10.3% and the bottom of the K already in credit distress. The labor market is already flatlining before any wealth-effect shock arrives. The housing channel is closed. The global disinflationary cushion is gone. The commodity supercycle is providing a structural inflation floor. The Fed has been politically compromised.

If the AI capex cycle unwinds — and any of the catalysts that triggered the 2000 unwind could trigger this one — the resolution path looks nothing like 2001. The 2001 recession resolved through aggressive Fed easing into a disinflationary environment with fiscal cushion. None of those conditions exist now.

The closer historical analog for the resolution dynamics is 1973–75 or 1979–82, not 2000–02 — supply-driven inflation persisting through a growth downturn, fiscal constraints binding, central-bank credibility under pressure, and an equity drawdown that does not get bailed out by aggressive easing. The 1973–75 episode saw a 48% nominal S&P drawdown over 21 months, but the real return on the index did not fully recover until 1985 — a twelve-year wait, almost entirely because of inflation. The 1979–82 cycle saw double-digit unemployment, two recessions back-to-back, and a 27% real S&P drawdown before Volcker finally broke the inflation regime.

The dot-com analogy is correct as a warning about the entry point. It is dangerously wrong as a guide to what comes next — because what comes next is not a recession in the 2001 sense. It is something with a different name and a different duration.

What this all adds up to

Six locks. Together they form a closed box.

• Productivity is propped up by capital spending rather than efficiency, so unit labor costs are an embedded inflation source.

• Fiscal dominance removes the government’s ability to stimulate or absorb additional tightening.

• Compromised Fed independence caps credible easing at perhaps 150 basis points.

• Closed housing channel eliminates the standard middle-class wealth-effect transmission.

• No global disinflationary cushion — China, Europe, Japan, and the dollar each provided one in prior cycles; none provide one now.

• Commodity supercycle sets a structural inflation floor independent of monetary or fiscal policy.

Any one of these would make the current inflation harder to resolve than past episodes. The six together describe a regime where the standard recession playbook is not partly unavailable — it is fundamentally unavailable. There is no Greenspan move to make.

The policy box is closed. The question for portfolio construction is no longer when will the Fed cut and start a recovery. The question is: what should I own when the playbook everyone is positioned for is the wrong playbook?

That is the work of Part 3.

Part 3 — “Position for the Regime, Not the Recession” — drops next.

* * *

In 2023, the United Arab Emirates’ first homegrown fintech unicorn quietly moved its operational headquarters from Dubai to Riyadh.

Tabby was the poster child. Founded in Dubai in 2019. Funded by Dubai-based VCs and Abu Dhabi’s Mubadala. Valued at $3.3 billion at its February 2025 Series E. Profiled by every regional publication as proof that the Emirates could produce the global technology companies its policymakers had been promising. And then, at the moment it actually needed to scale, it left.

The Saudi rationale was clean. By Tabby’s own disclosure, more than 80% of its 15 million users were Saudi. The IPO would list on Tadawul, not the DFM. Hassana, the Saudi pension institution that co-led the Series E, sits at the centre of Saudi capital markets. The move was not a vote against Dubai’s regulators or free zones or visa authorities — all of which had become easier to work with every year for a decade. It was a recognition that the customers, the listing venue, and the strategic relationships that mattered for the next chapter all sat in Riyadh. The marquee scaling story of the UAE ecosystem had to leave the UAE to scale.

This is awkward, because the official ambition has rarely been higher. In September 2025 the authorities announced a target of two million companies by 2031, up from 1.2 million today, alongside ten UAE-origin unicorns within the same window. Thats one company for every six residents. Hub71 in Abu Dhabi reported $2.17 billion in cohort funding for 2024, up 45% year-on-year. The Dubai Chamber of Digital Economy says Dubai welcomes a new digital startup roughly every thirteen hours.

By almost any input measure, the UAE has built one of the world’s most efficient places to start, incorporate, fund, and staff a business.

By almost any output measure, it has not yet built a place that produces globally recognised technology brands.

The UAE is a Gateway, not a Factory. It is the best place in the world to move your company to if you want to access the region. It has not yet mastered the art of building companies that go out and conquer the world. The strategy is one of Capture — attracting talent, capital, and headquarters from elsewhere. The marketing is one of Creation. The gap between the two is the central credibility risk to the strategy.

These are different problems, and you cannot solve the second by doing more of the first.

The minister himself counts “around five” unicorns that have originated in the UAE since the ecosystem began. The most valuable — Tabby — relocated to Riyadh. Another — Property Finder — was taken private in 2024 in a debt-financed buyback. A third — Careem — was sold to Uber in 2020 and broken up thereafter. The list of UAE-built tech companies that scale globally, today, with majority revenue from non-MENA customers, is one that no one in the ecosystem can populate convincingly. If it existed, it would be in every Hub71 press release.

* * *

The Gateway works. The numbers prove it.

Tracxn registers 56,212 companies under the UAE startup label. StartupBlink ranks 2,028 active tech startups. The government puts the total company count at 1.2 million, of which roughly 94% are SMEs. These figures are not measuring the same thing. The gap between them is the story.

The largest numbers measure incorporations and relocations, not companies built in the venture sense. The UAE’s free zone system was designed to make registration trivially easy — and it works. DIFC, ADGM, Dubai Internet City, twofour54, IFZA, and roughly forty other zones. A residency visa, a flexi-desk, 100% foreign ownership, total cost often under twenty thousand dirhams. For the country, this is a feature: it lifts the economy off oil and makes the UAE a magnet for capital and talent. The free zone strategy may be the single most successful piece of economic policy in the Gulf in thirty years.

Telegram, the messaging platform with hundreds of millions of users, is headquartered in Dubai because Pavel Durov chose it as a jurisdiction. Binance, whatever its current regulatory status, has substantial UAE operations because the UAE was willing to license crypto when most jurisdictions weren’t. Hub71’s Cohort 15 brought in EpiBone (US healthtech) and Partanna Oasis (US-Bahamian climate tech) with their funding rounds already closed. These are wins. They are the Gateway working exactly as designed.

What the Gateway does not do — not yet, at least not in measurable volume — is produce technology companies that scale globally from the UAE.

Hub71 illustrates the point. Cohort 15, announced in September 2024, admitted 21 startups from 1,228 applications. They had collectively raised $134.9 million in funding before joining. Seventeen of the twenty-one were founded outside the UAE. Eighty percent of the cohort was headquartered outside the country before admission. Hub71 is, on the evidence of its own data, primarily a redomiciliation platform with an accelerator wrapped around it. It is exceptional at what it does. It is not, on present evidence, building the next Careem from inside the cohort. It is importing companies that were already built somewhere else.

The federal target works the same way. The UAE added 250,000 new companies in 2025 alone. Hitting two million by 2031 requires roughly 130,000 net new registrations a year. Easy, given that Dubai welcomes a new digital startup every thirteen hours. What is not required is that any of those two million companies sell anything to anyone outside the Gulf.

This matters because the rhetorical framing has run ahead of the data. The national campaign is The Emirates: The Startup Capital of the World. Dubai’s separate D33 economic agenda targets thirty unicorns by 2033. The federal goal is ten by 2031. Government press releases benchmark Dubai against Silicon Valley, London, and Singapore. The gap between those framings and an output of around five UAE-origin unicorns — mostly regional, the largest of which moved to Saudi Arabia — is the gap between Capture and Creation. Between Gateway and Factory.

It is one thing to be the world’s best place to register a tech company. It is another to be the world’s best place to build one. The UAE has confused these. Every other claim in this piece follows from that confusion.

* * *

Name one with majority non-MENA revenue

The fastest way to test whether an ecosystem builds globally relevant companies is to ask a simple question: how many of its largest, most-celebrated startups earn the majority of their revenue from customers outside the home region?

For Singapore, the list is long. Sea Limited operates across Southeast Asia and Latin America. Grab (founded in Malaysia, headquartered in Singapore since 2014) covers eight countries. Shein (China-origin, Singapore-headquartered) sells in 150 markets. Trax, NIUM, Bolttech, Airwallex are all built to operate beyond Singapore from day one — because Singapore’s 5.9 million people cannot sustain them otherwise. Singapore is both a Gateway and a Factory.

For Estonia, with 1.36 million people, the answer is almost universal. Skype was global from inception. Wise (Estonian-founded, London-listed) operates in over 80 countries. Bolt runs across 45+ countries. Pipedrive sold to Vista at $1.5 billion as a CRM with customers in 179 countries. Veriff, Gelato, and Starship Technologies — all majority non-Estonian revenue. The country has roughly ten unicorns from a population smaller than Sharjah’s, and substantially all of them earn the bulk of their revenue outside Estonia. Estonia is barely a Gateway at all. It is overwhelmingly a Factory.

For Finland: Oura raised at an $11 billion valuation in October 2025, selling globally. Wolt sold to DoorDash for €7 billion. Supercell makes games played on every continent. For the Netherlands: Mollie processes payments across Europe; Adyen lists Spotify, Uber, and Microsoft as customers. For Switzerland: SonarSource and Scandit sell to enterprises globally.

For the UAE, the candidates look like this.

Careem. Acquired by Uber for $3.1 billion in a deal announced in March 2019 and completed in January 2020. GCC plus Pakistan. Majority of revenue always regional. Sold to a global platform precisely because it had reached the regional ceiling.

Kitopi. Cloud kitchens. $165.7 million in revenue in 2024 across UAE, Saudi Arabia, Kuwait, and Bahrain. Briefly expanded to the US in 2019, exited that market during the pandemic. Approximately 90%+ of revenue from the GCC.

Property Finder. Taken private in a 2024 buyback financed by $90 million of debt from Francisco Partners. MENA-only. No published international expansion.

Dubizzle Group. Classifieds and proptech. Regional. Privately held.

Tabby. Already discussed. Saudi-headquartered since 2023. 80%+ Saudi revenue by the company’s own disclosure.

Huspy. The strongest counter-case — and worth examining carefully, because it tests the Gateway-versus-Factory distinction directly. Founded in Dubai in 2020, raised a $59 million Series B in July 2025 led by Balderton. The Spanish business reportedly grew “more than 20x year-on-year in 2024.” Balderton reports Huspy facilitates over $7 billion in annual transactions across Europe and the Middle East. This is genuine international growth, and on Huspy specifically the burden of proof now sits with the sceptic. But UAE revenue still appears to dominate and one case does not yet make a pattern.

Astra Tech. BOTIM operator. Claims 150 million users across 155 countries. Most are migrant workers using BOTIM for remittances on corridors anchored in the GCC. User count is not revenue, and remittance corridors originating in the UAE are not a non-MENA business.

That is the list. No publicly verifiable UAE-built tech company of scale currently appears to derive a majority of its revenue from customers outside MENA. The qualifier “publicly verifiable” is doing real work — most of these companies are private and do not disclose revenue geography. But the absence of any such company being named by founders, investors, or the government is itself a data point. If a Factory output existed, it would be the headline.

The contrast with Singapore and Estonia is structural, not personal. UAE founders are not failing at scaling. They are responding rationally to a market with a different gravity. The GCC is roughly 60 million people across six different regulatory regimes and six different payment infrastructures. The TAM is real but bounded. Singapore-based startups can reach 680 million people in ASEAN through largely compatible regulation. Estonian startups reach 450 million people in the EU under a single regulatory framework. The UAE’s home region is roughly a tenth of either. Different incentive gradients produce different companies.

The honest version of the UAE’s pitch would be: come here to access the region, or come here to redomicile what you have already built. Both are accurate. Both are valuable. Neither is “the world’s startup capital,” which implies a Factory the UAE has not yet built. Conflating Gateway with Factory has costs — to the credibility of the strategy, to the expectations of the founders who arrive, and to the LPs who back UAE-domiciled funds expecting global-scale outcomes that the underlying market does not yet support.

* * *

The funnel doesn’t lie

If the qualitative version of the argument is which UAE companies scale globally (very few), the quantitative version is how many even get the chance.

MAGNiTT’s 10-Year Funnel Analysis of Startup Funding in MENA, Southeast Asia and Africa (October 2025) found that between 2014 and 2024, just over 7% of early-stage MENA startups advanced to a Series A. Only 0.6% reached late-stage rounds. Southeast Asia, by comparison: more than 17% to Series A. The UAE sits roughly at the MENA mean.

For every hundred startups that raise a pre-seed or seed round in the UAE, ninety-three never raise a Series A. For every hundred that raise an A, the great majority never raise a B. By Series C, survivor counts are in single digits. Most of the funnel collapses between $1 million in seed funding and $10 million in Series A. The Factory is leaking at every stage.

Farah El Nahlawi, MAGNiTT’s research department manager, on the same data: “This steep drop-off highlights a clear scale-up gap. The issue isn’t with founding activity — early-stage participation is strong — but rather with sustaining momentum.”

Helen McGuire, founder of The Founder’s Sanctuary in Dubai, named the binding constraint: “Market size is a fundamental constraint. Individual MENA markets are small and fragmented and many have transient populations — up to 90 percent expatriates in some places. Building a loyal customer base becomes extremely challenging.”

Stage composition reinforces the funnel data. H1 2025 MENA funding rose 92% year-on-year to $1.5 billion. 85% of that capital flowed to the UAE and Saudi Arabia. Within those flows, the distribution stayed barbelled: abundant seed and Pre-A activity at the bottom, a small number of mega-deals at the top, a thin middle. MAGNiTT noted that the share of Series A and B rounds exceeding $20 million jumped from 10% in H1 2024 to 42% in H1 2025 — real progress on the upper-middle, but from a low base. The two MENA mega-deals in H1 2025 were Tabby ($160M, Saudi-HQ’d) and Ninja ($250M, Saudi-HQ’d). The UAE equity Series C/D ledger over the same period is conspicuously short.

A 7% Series A conversion rate is not a cyclical phenomenon. It does not resolve with one more vintage of fund managers or one more rate-cut cycle. It is a structural feature of an ecosystem in which the demand-side TAM, the supply-side growth capital, and the operating expertise required to scale through Series B are all simultaneously thin. The Gateway can be improved while the Factory remains broken. That is the situation today.

* * *

Three things are being conflated

The frenzy makes more sense once you separate three things the UAE’s policy machinery has been treating as one.

Ease of incorporation. Free zones, golden visas, zero corporate tax in qualifying zones, the dirham peg, full foreign ownership, regulatory processing in days. World-class. Among the easiest places on earth to start a business. The achievement is real and should not be dismissed. This is Gateway infrastructure.

Ease of operation. Talent visas, English as a working language, time zones that bridge Europe and Asia, world-class airports, safety, infrastructure, and a tax environment that lets founders and employees keep what they earn. The UAE is projected to gain roughly 9,800 net new millionaires in 2025, the largest such inflow of any country in the world. Again, almost no peer can match it. Also Gateway infrastructure.

Ability to scale. Requires three things at once: home-region TAM large enough to support a first product through to strong unit economics; a complete capital stack from seed through pre-IPO; and a deep enough base of operating talent — technical leadership, growth marketing, finance leadership through an IPO — to run a thousand-person company across jurisdictions. This is Factory infrastructure. It is different in kind from the first two.

The UAE has the first two in abundance. The third is improving but thin. The TAM constraint is structural. The capital stack is barbelled: abundant seed below $5 million, abundant sovereign capital above $100 million for outbound global deployment, and a sparse middle in the $20–80 million range where Series B and C live. The operating talent base is young; senior product, engineering, and finance leadership at IPO scale is largely imported and rotational, with few executives having run a UAE-headquartered company from $20 million to $200 million in revenue.

The conflation is selling the first two layers — the Gateway — and assuming the third — the Factory — follows. It doesn’t. The ease of starting a business in the UAE is now so far ahead of the ability to scale one from the UAE that the gap itself has become the story. Tabby’s relocation is the visible symptom. The 7% Series A conversion rate is the underlying disease.

This is what makes the confusion expensive rather than merely rhetorical. Founders who arrive in Dubai on the strength of the Gateway pitch discover, eighteen months in, that the next round requires courting Silicon Valley, London, or Riyadh — and that the closer to scale they get, the less the UAE-domiciled capital base can serve them. LPs who back UAE-focused funds find that the DPI is concentrated in a handful of exits a decade apart. Strategists who measure progress by registration counts find that the metric they’re optimising has decoupled from the metric that matters.

The fix is not to do less of the Gateway. It is a genuine, durable, hard-won advantage. The fix is to stop pretending the Gateway and the Factory are the same project.

* * *

The counterarguments, and why they don’t fully hold

Three defences of the current strategy. Each deserves a serious response.

“Give it time. The ecosystem is only a decade old.” On the surface, reasonable. Venture ecosystems take time. MENA is, by most measures, less than fifteen years into serious venture activity.

But the empirical record of small economies that produce globally relevant technology suggests time is not the binding constraint. Estonia exited Skype at $2.6 billion to eBay in 2005 — five years after the country began building a tech sector at all, from a starting point of post-Soviet GDP collapse. Finland produced Supercell, Rovio, and Wolt within fifteen years of the Nokia decline. Thirteen years into the modern UAE ecosystem — dating from Careem’s founding in 2012 — the question is no longer “is it too early to judge?” but “what is the trajectory of the Factory showing us?”

“Sovereign capital will solve it. MGX, Mubadala, ADIA, and ADQ can write any cheque required.” Sovereign capital is the part of the UAE’s strategy that has worked most spectacularly on its own terms. MGX launched in March 2024 targeting $100 billion in AUM, co-led OpenAI’s $6.6 billion round in October 2025, and partnered with BlackRock and Microsoft on a multi-billion-dollar AI infrastructure vehicle. Mubadala Capital has done over a hundred venture investments globally.

But almost none of this is funding UAE-built startups. It is buying minority stakes in US frontier technology and global infrastructure. This is, in effect, a third strategy alongside Gateway and Factory: call it the Portfolio strategy — purchasing optionality on the future of AI rather than building a domestic Anthropic. It is defensible. But it does not fix the missing middle in the domestic capital stack. A UAE Series B founder still typically has to look to London, Riyadh, or Silicon Valley for a lead investor. The presence of $100 billion in sovereign capital deployed outward does not change that. Part 2 of this series examines why.

“Regional companies are valuable too. Why does global scale matter?” The most intellectually honest counter, and partly right. Regional champions create real jobs, real customers, real tax revenue. Careem at $3.1 billion is a more meaningful regional outcome than many countries have produced. A future of UAE-built regional champions is not a failed future.

But it is not the future the marketing describes. The campaign is The Emirates: The Startup Capital of the World, not The Emirates: The Gateway to the Region. The benchmark comparisons are to Silicon Valley and Singapore, not to São Paulo or Warsaw. The two-million-companies target is aspirational on a global scale. If the strategy is in fact to be a regional gateway and a redomiciliation platform — and there is nothing wrong with that, both are valuable — the framing should match. Calling São Paulo Silicon Valley doesn’t make it Silicon Valley. It makes São Paulo’s actual achievements harder to see.

* * *

What the data is telling us

The two-million-companies target is achievable. The country will likely hit it. The free zones will keep processing applications. Hub71 will keep welcoming cohorts. Sovereign capital will keep funding global frontier technology. The headline rankings will keep climbing. The Gateway will keep working.

But hitting the target does not answer the question that matters, which is whether any of those two million companies will sell to anyone outside the GCC at scale. The honest answer, on present evidence, is: a handful, and not nearly as many as the marketing implies. The Factory is not yet operating at the rate the rhetoric requires.

The UAE is the best place in the world to move a company to. It is not yet one of the best places in the world to build a company from. Both can be true at once. Pretending they are the same is what makes the strategy vulnerable.

The UAE has built one of the world’s most efficient places to register, fund, and staff a company. That is a genuine, durable, hard-won achievement, and the country deserves credit for it. It is also not, by itself, what its policymakers have promised it would deliver. The next decade of the UAE’s startup strategy will be measured not by how many companies it registered, but by how many of those companies it kept — and how many sold something to someone outside the region.

Tabby’s relocation is the data point everyone in the ecosystem already knows and few are willing to discuss in public. It will probably not be the last. Until the missing middle of the capital stack is built, until the regulatory friction that makes GCC expansion cost five compliance projects is reduced, and until the gap between Gateway and Factory is named honestly in the official rhetoric, more of the UAE’s best companies will face Tabby’s choice. And more of them will make the same decision Tabby did.

* * *

Coming next in Part 2: why the UAE’s capital stack has a missing middle — the structural barbell of overheated seed funding and outward-flowing sovereign capital that leaves Series B and C founders to look abroad, and what a redesigned domestic capital architecture might look like if the UAE wanted to add Factory to Gateway.

And yet the headline economy looks superficially fine. GDP grew 2.0% in the first quarter. The unemployment rate sits at 4.4%. The S&P 500 is within touching distance of its highs. The Conference Board’s Consumer Confidence Index, the other major sentiment measure, has actually edged higher in each of the last three months.

How do you reconcile a 74-year low on one survey with a steady headline on another, against a backdrop of resilient growth? The answer is the most important macro story of the year, and it is not being told often enough: the United States has already entered a stagflationary regime for most of its income distribution, and the aggregate data is being held up by a top decile whose spending is funded by an equity market that is itself a bet on a single capital-expenditure cycle.

This is not a forecast. It is a description of conditions already in the data. The interesting question is not whether stagflation is coming — it is whether the small, equity-wealth-driven slice of the economy keeping the headlines afloat will continue to do so, and what happens if it doesn’t.

This is Part 1 of a three-part series. Here I lay out the diagnosis: what the data is showing, where we are in the typical sequencing of a stagflation regime, and why the household-level signal is so much stronger than the headlines suggest. In Part 2 I work through the structural mechanism — six interlocking constraints that lock this regime in and make it unable to resolve the way every postwar recession has. In Part 3 I lay out the historical playbook from the 1973–75 and 1979–82 episodes, what is priced in 2026 markets, what is not, and a framework for watching the data month by month.

Before walking through the evidence, let me address a confusion that runs through most of the current commentary on this regime.

Stagflation is not a recession. That distinction is the whole point.

A great deal of analyst and journalist commentary treats “stagflation” and “recession” as synonyms or close cousins. They are not. They are different regimes with different mechanics, different policy responses, different durations, and different consequences for portfolio construction. Conflating them produces bad positioning.

A recession, in the postwar American experience, is a demand-driven contraction. Households and businesses cut spending, output falls, unemployment rises, and — crucially — inflation falls as the demand shortfall destroys pricing power. The Fed responds by cutting rates aggressively, financial conditions ease, the wealth effect reverses, and the economy recovers in a “V” or shallow “U” pattern. Average duration: ten months. Average peak-to-trough GDP decline: 2.5%. The inflation tailwind is what makes the recovery possible — falling CPI gives the Fed room to ease, which restarts the cycle.

Stagflation is structurally different. It is supply-driven inflation that persists through a growth slowdown rather than being cured by it. Output stalls or contracts, but prices keep rising because the cause is on the supply side — energy shocks, tariffs, broken supply chains, fiscal dominance, productivity collapse. Because inflation is high, the central bank cannot ease without de-anchoring expectations further. Because growth is weak, it cannot tighten without crushing what remains of the real economy. The policy response that resolved every postwar recession is unavailable.

Average duration of the two clean postwar stagflations (1973–75 and 1979–82): roughly six years from first symptom to clean resolution. The 1973–75 episode saw a 48% nominal S&P drawdown over 21 months, and the real return on the index did not fully recover until 1985 — a twelve-year wait, almost entirely because of inflation. The 1979–82 cycle saw double-digit unemployment, two recessions back-to-back, and a 27% real S&P drawdown before Volcker finally broke the regime with a 19% Fed Funds rate.

The reason this matters: a recession is a problem that monetary policy is built to solve. Stagflation is a problem that monetary policy is built to make worse. Cutting rates into supply-driven inflation accelerates the inflation. Hiking rates into demand weakness deepens the contraction. There is no good first move. That is what Paul Volcker confronted in 1979, and the price of breaking the regime was a 10.8% unemployment rate and the worst real equity drawdown since the 1930s.

So when commentary today asks “are we headed for a recession?”, the question is partly the wrong one. The more important question is: are we headed for a regime in which the normal policy response to a downturn is unavailable, and the downturn — when it comes — therefore plays out differently?

The typical sequencing

Stagflation regimes do not arrive all at once. They follow a recognizable pattern, which is worth laying out because it tells you where in the sequence we currently sit.

Phase 1 — The supply shock. A non-cyclical event raises the price level: OPEC embargoes in 1973, the Iranian Revolution in 1979, the combined tariff regime plus the Iran war plus the reversal of the China deflation cycle in 2025–26. The Fed initially treats this as transitory. Inflation expectations begin to drift upward but remain anchored.

Phase 2 — The bifurcation. The shock hits lower-income households first because they spend a larger share of income on the shocked categories (food, fuel, basic goods). Real wages turn negative for the bottom of the distribution before the aggregate. Higher-income households, insulated by wealth and discretionary spending margins, continue spending. Aggregate data looks fine; distributional data does not. This is where the K-shape appears.

Phase 3 — The credibility test. Inflation expectations de-anchor. The central bank is forced to choose between aggressive tightening (which it may not be able to deliver because of fiscal or financial-stability constraints) and tolerating higher inflation. Markets begin to price the choice. The bond market sells off, the currency wobbles, and the central bank’s institutional credibility becomes a market variable in its own right.

Phase 4 — The transmission to the top. The wealth-effect engine that has been holding up aggregate consumption finally breaks. Sometimes it breaks because equities correct (1973-style); sometimes because credit conditions tighten (1979-style); sometimes both at once. Top-of-distribution spending rolls over. The “stag” piece of the regime, which has been hiding underneath the K-shape, finally shows up at the headline level.

Phase 5 — The resolution. Either the central bank breaks the inflation through a deep policy-induced recession (Volcker, 1979–82) or the regime drags on for years while real incomes and real returns slowly grind lower (the UK and Japan in the 1970s, various emerging-market stagflation episodes since).

I think the United States in May 2026 is in late Phase 2 and early Phase 3. The K-shape is fully visible in the data. Inflation expectations are de-anchoring on every major survey. The credibility test of the Fed has just opened with the most partisan chair confirmation in history. The transmission to the top has not yet happened — top-decile spending is still elevated, equities are near highs, AI capex is still adding to GDP. That is the piece that has not arrived yet, and it is what most market participants are not pricing.

With that frame established, let me walk through the evidence in turn.

The signature is already in the data

Stagflation has a clinical definition. It is the simultaneous occurrence of rising inflation, decelerating output, and stalling employment — the textbook signature of a supply-side shock to an economy that monetary policy cannot easily fix. Each piece of that signature, individually, is conventional and easy to dismiss. The diagnosis comes from the simultaneity.

Consider what is happening at once.

Inflation is reaccelerating, not disinflating. April CPI rose 0.6% on the month and 3.8% on the year — the highest annual rate since May 2023. Core CPI ran 0.4% MoM and 2.8% YoY, the strongest monthly core print since January 2025. The trajectory tells the story better than any single print: headline CPI ran 2.4% in January and February, jumped to 3.3% in March, and reached 3.8% in April. That is a 140 basis-point acceleration in three months. The disinflation narrative that anchored markets through 2024 is no longer operative.

Real wages have turned negative. Real average hourly earnings slipped 0.5% on the month and fell 0.3% over the year. For the first time in three years, American workers’ wages are no longer outpacing inflation. This is the lived experience that gives stagflation its political force — and historically, the consistent first signal that Phase 2 has begun.

The labor market is stalling beneath the headline. April nonfarm payrolls beat consensus at 115K, but the Bureau of Labor Statistics itself acknowledged that total employment has shown “little net change over the prior 12 months.” The three-month average has fallen to 48,000 jobs per month — barely above the level needed to absorb new entrants. Labor force participation slipped to 61.8%, the lowest since late 2021. Federal employment is down 348,000 (11.5%) from its October 2024 peak. Information-sector employment is down 11% from its November 2022 peak — partly an artifact of AI displacement, but real all the same.

There is also a demographic dimension worth surfacing here because it sharpens the supply-side argument. Net migration has fallen from over 2 million annually under prior policy to near zero under current immigration restrictions. The labor-supply tailwind that helped moderate wage pressure through the 2010s and early 2020s is gone, even with weak headline payrolls. This is one of the structural mechanisms I’ll work through in Part 2; for now, note that the labor market is simultaneously weakening on the demand side and tightening on the supply side. That combination is not a normal recession setup.

Inflation expectations are de-anchoring. This is the indicator that defines Phase 3. UMich one-year inflation expectations sit at 4.5%; five-year expectations at 3.4%. The Conference Board’s twelve-month measure hit 7.9% — an all-time high on its own series. In 2019–2020, long-run UMich expectations consistently ran below 2.8%. They have stepped up a full percentage point and stayed there. That is the empirical definition of de-anchoring in slow motion, and it is precisely the development the Federal Reserve has structured its entire credibility around preventing.

None of those data points, individually, proves stagflation. The combination — accelerating prices, negative real wages, flatlining payrolls under a supply-constrained labor market, de-anchoring expectations — does. That is the signature.

Tariffs and energy: the cleanest supply-shock mechanism in decades

Stagflation as a phenomenon requires a supply-side cause. Demand-side overheating produces inflation, but inflation alone is not stagflation; you need something that simultaneously raises prices and constrains output. The 2025–2026 cycle has produced the textbook example, and the data on it is unusually clean.

The Yale Budget Lab estimates the U.S. average effective tariff rate now stands at 11.0% — the highest since 1943. That is an 80-year extreme, sourced from primary government data, not analyst opinion. The pass-through to consumer prices has been quantified by multiple Federal Reserve studies. The Fed Board staff estimates that tariffs implemented through November 2025 raised core goods PCE prices by 3.1% through February 2026 — explaining the entire excess of core goods inflation relative to pre-pandemic baselines, and contributing 0.8 percentage points to the core PCE aggregate. The Dallas Fed reaches a nearly identical conclusion: core inflation absent tariff effects on relative prices would be 2.3% rather than the 3.1% currently observed.

The transmission mechanism has a known lag. Fed staff work pinpoints it at approximately seven months: if a retailer’s acquisition cost rises by $1 because of tariffs, the customer pays that $1 more roughly seven months later. This matters because the tariffs imposed in late 2025 and early 2026 are still working through the pipeline. The April CPI prints are not the peak of the tariff effect; they are the middle of a wave whose later sections are still arriving.

The Yale Budget Lab quantifies the household impact: the current tariff regime implies a 1.0% short-run consumer-price shock, equivalent to roughly $1,338 per household in 2025 dollars. That is a regressive consumption tax landing on the same households whose real wages just turned negative. It is, in effect, a deliberate engineering of a stagflationary impulse — supply-shock inflation that policy cannot offset without abandoning its trade objectives.

The energy side has piled on. The Iran war energy shock pushed Brent above $126 at its peak, drove gasoline prices up 28.4% year-over-year, and contributed more than 40% of the headline CPI acceleration. The Conference Board put the transmission channel in one sentence: “The war in Iran has unleashed an oil-driven inflationary pulse that is percolating through the U.S. economy. It is resulting in purchasing power erosion and growth compression, forcing the Fed into an increasingly untenable tradeoff between inflation and the labor market.” That is stagflation theory, written in real time.

This is Phase 1, fully visible. It is no longer something the Fed can describe as transitory; the pass-through is empirically quantified and the lag means more is still arriving.

Four independent business surveys confirm the signature

A skeptic might argue that the inflation and labor data could be explained by other things — energy shocks, AI displacement, statistical artifacts. The strongest single piece of evidence against that view is the business survey data, because four independent organizations using four different methodologies on four different samples of firms are all reporting the same thing in the same month.

The April ISM Manufacturing release is the cleanest single piece of stagflation evidence in the 2026 data. The headline PMI sat at 52.7 — a mild expansion. Underneath, the Prices Index reached 84.6, the highest reading since April 2022, rising 25.6 percentage points in just three months. The Employment Index sat at 46.4 — its 31st consecutive month of contraction. To put that streak in context: since January 2023, the ISM Manufacturing Employment Index has contracted in 39 of 40 months. A factory sector that has not expanded employment in nearly three years, simultaneously experiencing the strongest input-cost pulse in four years, is precisely the configuration the term “stagflation” was coined to describe.

ISM itself names the drivers explicitly: steel and aluminum prices feeding through value chains, tariffs on imported goods, and petroleum-based product increases from the Middle East conflict. That is tariffs plus energy plus commodities, attributed by survey panelists themselves.

ISM Services confirmed the same pattern in the bigger 70% of the economy. Services Prices held at 70.7 — the highest reading since October 2022, and the seventeenth consecutive month above 60. This is no longer a transitory pulse; it is an embedded inflation regime. Services Employment contracted for the second consecutive month at 48.0, after hitting 45.2 in March (the lowest services-employment print since December 2023). New Orders cratered 7.1 points to 53.5 — the sharpest single-month decline in three years, suggesting March’s front-running of expected price hikes is already unwinding.

The regional Federal Reserve manufacturing surveys triangulate the signal. The Philadelphia Fed’s prices-paid index spiked 15 points in a single month to 59.3, while its employment index dropped 6 points into contraction at –5.1. The Empire State manufacturing survey saw input-price increases accelerate from 36.6 to 51 — an enormous one-month jump — while forward business optimism collapsed from 31 to 19.6.

The most important piece of color came from S&P Global’s chief business economist, Chris Williamson, on the April manufacturing PMI:

The surge in manufacturing activity in April is not the cause for cheer that at first glance it suggests. A key driving force behind the upturn is the need for companies to get ahead of further feared price rises and supply shortages, providing a short-term boost that could fade in the coming months as headwinds to the economy continue to build... employment has fallen as firms grow increasingly worried over the need to reduce cost overheads amid an environment of rising raw material prices, while selling prices have jumped higher as producers seek to protect their margins.

That is, almost line for line, a description of stagflation transmission: demand pulled forward by inflation fear, margins compressed, layoffs starting, selling prices rising. Both ISM Manufacturing and S&P Global explicitly flag current order strength as pull-forward demand ahead of feared price increases. If true, the May and June releases should show new orders rolling back over as that demand exhausts — at which point the “stag” piece of the thesis becomes visible at the headline level rather than just underneath it.

The consumer is the loudest signal — once you split the data

The University of Michigan Consumer Sentiment Index at 48.2 is an extraordinary data point. It is below the June 2022 inflation-shock trough, below the 2008 financial crisis lows, below the early-1980s Volcker stagflation. It is the lowest sentiment reading in 74 years of continuous measurement.

The detail underneath matters more than the level. The current-conditions component fell nine percent in a single month, driven by what the survey directors described as a surge in concerns about prices both for personal finances and for buying conditions for major purchases. Roughly one-third of respondents spontaneously cited gasoline prices; about 30% cited tariffs. Consumers themselves are reporting the stagflation hierarchy: this month, inflation was selected as the bigger economic worry by twice as many respondents as unemployment, but 68% expect unemployment to rise over the year ahead — up from 61% the prior month.

The Conference Board’s headline reading at 92.8 looks, on the surface, to contradict this. It does not. Three things are happening:

First, the two surveys are on different scales. CB uses 1985 = 100; UMich uses 1966 Q1 = 100. The numbers cannot be directly compared. On its own historical range, CB at 92.8 is also weak — below its 2022 inflation-shock trough, well below 2024 averages of 100–110, and 30 points below pre-pandemic levels of 120–135. The visual contrast is misleading.

Second, the surveys weight inflation and labor differently. Conference Board is labor-market-heavy: its Present Situation Index (half the headline) is driven by “jobs plentiful versus hard to get.” Until the labor market actually cracks, CB holds up. UMich is inflation- and purchasing-power-heavy, weighting buying conditions and personal finances more aggressively. In a pure supply-shock inflation regime — which is what we have — UMich is the more accurate read on consumer pain; CB is the more accurate read on whether the labor market has broken yet.

Third, and most importantly, the CB internals are screaming what UMich is saying out loud. The CB Expectations Index sits at 72.2; readings below 80 have historically been recession-territory. CB twelve-month inflation expectations hit 7.9% — the all-time high on its own series. Future spending intentions across nearly every services category rotated from “yes” and “maybe” to “no” between February and April. Households are trading down to what the Conference Board itself describes as “cheap thrills” and necessary services.

The headlines are different. The diagnosis is the same. And both align with where Phase 2 typically lands: distress visible to the household sector well before the aggregate data catches up.

The K-shape is the masking mechanism — and it has a generational dimension

The most useful piece of the consumer story is the income distribution, because it explains why the aggregate data has been able to look stable while UMich hits all-time lows. The K-shape is Phase 2 — the bifurcation that lets the headlines disguise the household-level stagflation that’s already in progress.

The top 10% of households accounted for roughly 49% of total consumer spending in Q2 2025 (Moody’s Analytics, via Mark Zandi). The top 20% holds nearly 72% of total household wealth. Moody’s reports that spending by the top decile grew 62% between Q3 2020 and Q3 2025, vastly outpacing every other income group. Newer data from the New York Fed: real cumulative spending growth through March 2026 was approximately 7.6% for households earning over $125,000, 3% for the middle, and just above 1% for those earning under $40,000. Bank of America Institute cardholder data tells the same story: top-third spending grew 4% year-over-year in late 2025, while bottom-third spending grew less than 1%.

There is some methodological dispute about the level — Pantheon Macroeconomics argues that the top decile’s share has been roughly stable at 40% for 25 years. The direction is not disputed. The top of the income distribution has pulled away from the middle and bottom on real spending growth since 2023.

There is also a generational dimension worth surfacing because it sharpens the political-economy implications. UMich data shows confidence holding up among consumers under 35 while collapsing among those 55 and over. Read at face value, that looks counter-intuitive — until you realize that confidence reflects a comparison to what people had before. Older Americans, who own homes and equities, are watching their real wealth get eroded by sticky inflation and are angry about it. Younger Americans, who own neither, have less to lose; their stagflation experience is a continuation of a high-rent, low-wage trajectory that has been their adult life. The boomer-millennial wealth gap is the K-shape on a demographic axis, and the credit-stress data we’re about to look at is largely a millennial and Gen-Z story.

Why this matters for the macro thesis:

One. Aggregate spending data overstates underlying economic strength. “PCE growth held at 1.6%” tells you almost nothing about the middle 60% of households, whose real spending has been roughly stagnant for two years. The macro headline is weighted by dollars, not by households, and dollars are concentrated at the top.

Two. Top-decile spending is wealth-driven, not income-driven. This is the critical point. Diane Swonk at KPMG put it precisely: “when you divorce growth from employment gains, you’ve got a problem.” The top of the K is funded by equity-market gains, not wage growth. That means the “resilient consumer” headline is functionally a leveraged bet on the S&P 500 — and specifically on the AI capital-expenditure cycle that is currently driving the index.

Three. Tariffs are regressive and widen the K mechanically. A flat 11% effective tariff rate hits lower-income households harder because they spend a higher share of income on goods. The $1,338-per-household price shock is an average — the bottom quintile bears a disproportionate share, while the top decile barely notices.

Four. The bottom of the K is already in credit stress. This is the data most market commentary misses.

The credit data underneath the K

The New York Fed’s Q1 2026 Quarterly Report on Household Debt and Credit, released May 12th, looks superficially benign. Aggregate household debt held essentially flat at $18.8 trillion. Aggregate delinquency held at 4.8% of outstanding balances. Credit card delinquency transitions actually ticked down from 8.7% to 8.6%. The headlines are stable.

The mix is not.

Student loan 90-plus-day delinquency stands at 10.3% of outstanding balances, up from 9.6% in Q4 2025. Approximately 2.6 million federal student-loan borrowers defaulted in Q1 2026, on top of 1 million in Q4 2025. The New York Fed estimates an additional 7 million borrowers in the now-defunct SAVE forbearance program are still working toward the nine-month default threshold — a second wave that has not yet arrived in the data.

The most striking statistic from the Liberty Street Economics post accompanying the report concerns cross-product delinquency among student-loan defaulters:

• Nearly 40% are past due on auto loans

• 56% are past due on at least one credit card

• 20% are past due on a mortgage

These are not normal numbers. They are pre-recessionary numbers, concentrated in a demographic that the aggregate is masking. The headline 4.8% aggregate delinquency rate holds steady only because the top half of the income distribution — which holds most of the debt by dollar value — remains current.

This is what bottom-of-K credit distress looks like before it shows up in aggregate statistics. And it has a specific implication for the stagflation regime: the cohort experiencing the worst of the inflation shock has no balance-sheet cushion to absorb further price pressure. Their savings are gone, their real wages are negative, their credit-card balances are stretched, and a substantial subset is already behind on multiple debts simultaneously.

Where this leaves us, and what’s next

Let me close Part 1 with the cleanest formulation of the diagnosis I can construct.

Stagflation is not a forecast for the United States in 2026. It is a current condition for most of the income distribution, mid-sequence, with the most visible phase still to come. The middle and bottom 60% of households are already living with negative real wages, de-anchored inflation expectations, real spending growth around 1%, and concentrated credit stress. That is the textbook household-level experience of stagflation — Phase 2, fully visible in the data, hiding in plain sight beneath aggregate statistics that weight every dollar equally and therefore weight the top decile disproportionately.

The top of the K — the 10% holding 49% of consumption — is what keeps the headlines from reflecting this. Their spending is funded by stock-market wealth, which is funded in turn by the AI capex cycle. That makes the aggregate “resilient consumer” story a leveraged position on a single set of corporate investment decisions. It is not a position the household sector controls. And historically, this is the configuration that ends Phase 3 and begins Phase 4 — the transmission of distress from the bottom to the top, mediated by an equity correction.

The diagnosis stands on the simultaneity argument: CPI reaccelerating, real wages negative, payrolls flatlining, four independent business surveys showing surging prices alongside contracting employment, two consumer-sentiment surveys at multi-decade lows on inflation expectations, the income-distribution data, and the credit-stress data. Any one piece has an alternative explanation. The pattern does not. The right question is not “are we headed for a recession?” It is: are we in a regime where the normal policy response to a downturn no longer works, and the downturn — when it comes — will therefore last years instead of months?

In Part 2 — “The Policy Box Is Closed” — I work through the six structural constraints that lock this regime in: why the Federal Reserve’s policy room has effectively closed; why productivity and unit labor costs have set an inflation floor that disinflation cannot easily breach; why fiscal dominance has removed the government’s stimulus lever; why housing is the missing recovery channel that pulled the US out of the 2001 recession but cannot do so now; how demographics and immigration policy have flipped the labor-supply side from disinflationary tailwind to inflationary headwind; and why the commodity supercycle has set a structural inflation floor for years to come.

Then in Part 3 — “Position for the Regime, Not the Recession” — I lay out the asset-class playbook from the 1973–75 and 1979–82 episodes: what worked, what didn’t, what surprised people. I show what’s currently priced into markets and what isn’t. And I set out a “what would convince me I’m wrong” framework with specific named data triggers, plus a monthly data-watching framework so you know what to look for in each release.

For now, the single most informative near-term data point is the June 1st ISM Manufacturing release. If the Prices Index stays above 80 and New Orders rolls back over as the pull-forward demand exhausts, the stagflation read goes from “credible thesis” to “the new consensus.”

The most expensive mistake in markets is waiting for a regime to be named before treating it as real. The data is already showing you the regime. Pay attention.

Part 2 — “The Policy Box Is Closed” — drops next.

Part II of a two-part series. Part I argued that the 4.3% headline unemployment rate is hiding a labor market that is structurally fragile rather than cyclically healthy — pessimistic in character but not yet recessionary. This piece commits to the specific criteria that would shift that call. If you missed Part I — “What 4.3% unemployment is actually telling us” — you can find it [here].

Recap

In Part I of this series, I argued that the 4.3% headline unemployment rate is misleading in a specific way: the labor market is stable because supply has collapsed in lockstep with demand. Labor force participation is at its lowest reading since October 2021. Foreign-born workers are leaving the labor force faster than they’re being replaced. “Breakeven payrolls” — the monthly job growth needed to keep the unemployment rate stable — has fallen from roughly 150,000 in the 2010s to perhaps 30,000–50,000 today. The result is a labor market that looks calm at the headline but is structurally fragile underneath: real wages negative, multiple jobholders at records, the recent-graduate unemployment gap inverted for the first time in four decades, AI cited as the top reason for one in four announced layoffs, and the vacancy-to-unemployment ratio sitting below pre-pandemic norms.

The verdict was that the picture is pessimistic, but of a specific kind. Not cyclically recessionary — the firing cascade that defines every classical downturn is absent. But structurally fragile, with the conditions for a non-classical break accumulating. The right word was fragile. The classical recession trigger is absent. The conditions for one are not.

The honest thing to do with that verdict is commit to what would change it. A pessimistic call without a falsification standard is just an assertion — exactly the kind of motivated reasoning the piece accuses Goldilocks-soft-landing commentary of doing in reverse. So what follows are six specific labor market metrics, chosen because each measures something distinct and because each has a clean historical anchor against which “normal” can be defined.

For each one I lay out what it actually measures, the historical context that makes the threshold meaningful, the current reading and trajectory, and the explicit current-vs-desired gap. I’ll score every monthly NFP print against this rubric from June onward. The first real test is the June 5 employment report.

None of the six are currently green. Several are moving the wrong way.

1. The hires rate breaks sustainably above 3.7%

The hires rate is total hires divided by total employment, reported monthly by the BLS through JOLTS. It’s the cleanest single measure of how fluid the labor market actually is — how easily someone who wants a job can find one, or how easily someone in a job can move to a better one. Unlike the unemployment rate it isn’t affected by labor force participation; unlike payrolls it doesn’t depend on the birth-death model.

The historical context matters. During the 2018–2019 expansion — late-cycle but still healthy — the hires rate averaged 3.9%. It peaked at 4.7% in early 2022 during the post-COVID reopening. Sustained readings above 4.0% are consistent with a labor market where firms are actively competing for workers. Anything below 3.5% means hiring has substantively frozen.

The current reading is 3.5% in March (latest monthly print), but the more important figure is the three-month moving average, which has been stuck at exactly 3.3% for eight consecutive months — the longest such stall outside of recessions in the twenty-five-year history of the series.

The 3.7% threshold is deliberately conservative. It doesn’t ask the labor market to hit 2018–2019 levels; it asks it to come halfway back. And the gains would need to be broad-based, not concentrated in healthcare and warehousing, which together account for the bulk of the residual hiring still happening. If hires returned to 3.7% with healthy sectoral distribution, the displaced tech and government workers and the unemployed recent graduates would actually have somewhere to go.

Current vs desired: 3.3% three-month average vs 3.7% sustained. A 40-basis-point gap that has been flat for eight months. The May JOLTS print (released early July) is the first real test of whether the March rebound was a turn or a head-fake.

2. The vacancy-to-unemployment ratio returns above 1.10

The vacancy-to-unemployment (V/U) ratio is total job openings divided by total unemployed workers — the cleanest single measure of aggregate labor market tightness. When the ratio is greater than 1, there are more open jobs than unemployed people looking; when it’s less than 1, the reverse. The Federal Reserve has cited this metric explicitly in nearly every FOMC press conference since 2022 as the key evidence that the labor market remained tight enough to keep wage and price pressure live. It is, more than any other single number, the metric the Fed itself has chosen to anchor its labor market view.

The historical anchors: the pre-pandemic 2019 average sat at 1.20. The March 2022 peak hit 2.02 — two open jobs for every unemployed person, the tightest US labor market on record. The ratio crossed below 1.0 in March 2025 and has been pinned in a 0.87–1.02 range for the fourteen months since.

The current reading is 0.95 (6.87 million openings against 7.24 million unemployed). For the first time since the immediate aftermath of the 2008 financial crisis, there are more Americans looking for work than there are open jobs. The Fed’s preferred labor market tightness gauge has fully rebalanced — and overshot. The case for continued hawkish policy on grounds of “tight labor market” is, empirically, no longer there.

The 1.10 threshold is again deliberately conservative. It doesn’t require a return to the 1.20 pre-pandemic norm; it asks for a recovery halfway back. Getting there would require either a sustained pickup in job openings (currently trending lower, with four declines in the last five months) or a meaningful drop in the unemployed count without a corresponding labor force exit. Either path implies a labor market in which firms are competing for workers again, rather than the current configuration in which workers compete for jobs.

Current vs desired: V/U ratio of 0.95 vs 1.10. The gap is 0.15, but the trajectory matters more than the level — the ratio has been below 1.0 for fourteen consecutive months and openings are still declining. The May JOLTS print is when we’d first see whether the March hires bounce had any follow-through on openings.

3. U-6 retraces toward 7.0% while U-3 holds

U-6 is the broadest of the BLS’s six unemployment measures. It captures the officially unemployed (U-3), plus people marginally attached to the labor force (who want work but haven’t searched in the past four weeks), plus people working part-time who would prefer full-time work. It’s the closest single number to a “true” labor market slack measure.

The U-6/U-3 spread is the diagnostic. Historically, U-6 has run about 3.5 percentage points above U-3 in healthy expansions. At the late-2010s peak of full employment, the spread compressed to roughly 3.4pp (U-3 of 3.5%, U-6 of 6.9%). During recessions the spread widens sharply as part-time-for-economic-reasons spikes.

The current spread is 3.9 percentage points (U-6 at 8.2%, U-3 at 4.3%) and has been widening. U-6 was 7.5% in January 2025; it’s now 8.2%. Part-time-for-economic-reasons jumped 445,000 in April alone, to 4.9 million people. Marginally attached workers stand at 1.8 million. There are an additional 6.1 million people not in the labor force who say they currently want a job.

The 7.0% threshold for U-6 would mean part-time-for-economic-reasons has fallen by roughly 1.0–1.5 million while U-3 holds at 4.3% or below. That’s the specific test of whether slack is being absorbed: people with partial attachment to the labor force converting it to full attachment without pushing the overall rate up. In the data, that’s the signal that the iced pond is melting from the inside.

Current vs desired: 8.2% U-6 vs 7.0% U-6, both at U-3 of roughly 4.3%. A 1.2-percentage-point gap that has been widening, not narrowing, for fifteen months.

4. Real AHE turns positive for two consecutive months while nominal AHE growth holds above 3.5%

Real average hourly earnings is nominal AHE adjusted for CPI inflation. It’s the cleanest single read on whether workers are getting ahead of or falling behind the cost of living. The two-condition test matters: positive real wage growth that comes from inflation crashing isn’t a sign of labor market strength — it’s a sign of demand destruction. Positive real wage growth with nominal AHE above 3.5% means workers are gaining purchasing power because their pricing power is intact.

The current readings: nominal AHE +3.6% year-on-year in April, decelerating from 4.0% averages through 2023–2024. CPI ran +0.6% month-on-month in April, with headline year-on-year around 3.0% and pushing higher on energy. Real AHE printed -0.5% month-on-month — a stagflation signal in miniature.

What I’d want to see is two consecutive months where AHE grows faster than CPI, with nominal AHE holding above 3.5% annualized. That can happen one of three ways: nominal wages reaccelerating (which would require tightening of labor demand in higher-wage sectors); CPI breaking sustainably below 3.5% annualized (which currently looks improbable given oil prices, tariffs, and services inflation running near a four-year high); or both.

Current vs desired: real AHE -0.5% month-on-month and decelerating year-on-year vs sustained positive prints with nominal AHE growth above 3.5%. The two series are heading in opposite directions — wages decelerating, prices reaccelerating. That divergence is the canonical stagflation pattern. Until it reverses, the textural deterioration in the labor market is being made worse, not better, by the prices side.

5. The recent-grad gap narrows back toward parity with overall unemployment

The recent-grad gap is the difference between the unemployment rate for recent college graduates (aged 22–27) and the overall workforce unemployment rate. It’s published quarterly by the New York Fed using BLS microdata.

This is the AI canary. Historically, holding a recent degree was a labor market premium — recent grads had unemployment rates 1–2 percentage points below the overall workforce, because employers valued the optionality of new entrants and because young workers were cheaper than mid-career hires. That premium has now flipped. In Q4 2025, recent-grad unemployment hit 5.7% while overall unemployment was 4.2%, producing a gap of +1.5 percentage points (recent grads worse off). The gap reached an extreme of -1.8 percentage points using the NY Fed’s directional convention (i.e., overall workforce 1.8pp worse than recent grads) at the prior cycle’s healthy peak — now it sits at the opposite extreme of the four-decade series.

For the first time in at least four decades, having a college degree and being new to the workforce is a structural disadvantage rather than an advantage. The mechanism is straightforward: AI is hitting the cognitive primitives — information gathering, drafting, classification, routine synthesis — that historically defined entry-level white-collar work.

The threshold for “back to parity” is a gap of zero or slightly negative. That would mean the entry-level white-collar hiring channel — paralegals, junior analysts, market researchers, software developers, copywriters, customer-support agents, content moderators — is functioning again. Closure of the gap wouldn’t require complete reversal of the AI-displacement story; it would simply require firms to start backfilling entry-level roles at the rate they were before late 2022.

Current vs desired: recent-grad gap of +1.5pp (worse than overall) vs zero or negative. The trajectory is worsening. The Indeed Job Postings Index for entry-level white-collar roles is at multi-year lows. Information sector openings are down 33% year-over-year. Junior-level postings fell 7% in 2025. The May Challenger report is the next data point — a third consecutive month with AI as the top-cited layoff reason would harden the structural-shift hypothesis.

6. Continued claims hold below 1.8 million through a meaningful supply normalisation

Continued claims is the number of people who have already filed an initial unemployment insurance claim and are still drawing benefits in the current week. Unlike initial claims (which measures the firing flow), continued claims measures the unemployment stock among insured workers. It’s the cleanest weekly signal of how long it takes the unemployed to get back to work.

The current reading is 1.766 million, a two-year low — extraordinary by any historical measure. The 1.85 million level has been flagged repeatedly in market commentary as the threshold where a layoff wave would start to register; we’re currently 84,000 below that.

But the threshold needs a conditional. Continued claims at 1.766 million is only impressive if it persists despite supply pressure. If immigration normalises — through court rulings, policy reversal, or political exhaustion — and the foreign-born labor force partially recovers, we’d expect LFPR to rebound from its current 61.8% toward the 62.1–62.3% range last seen in early 2024. That’s a 30-basis-point swing, which translates to roughly 800,000–900,000 additional people in the labor force. Some fraction of those returning workers would find jobs immediately. The rest would draw unemployment benefits — pushing continued claims up.

The honest test is whether claims stay below 1.8 million while LFPR climbs 30bps or more. If they do, it means the demand side of the labor market can absorb supply that’s currently being suppressed by policy. If continued claims spike when LFPR rises, it means the current low claims number is itself a supply artifact — a low denominator masquerading as labor market strength.

Current vs desired: continued claims 1.766M and LFPR 61.8% (and falling) vs continued claims sustainably below 1.8M with LFPR back above 62.1%. The current configuration tells us nothing about the underlying durability of the labor market. Until we see at least the start of a supply normalisation, the low claims number is essentially uninterpretable as a strength signal.

The scorecard

Zero green. One ambiguous (continued claims, dependent on supply). Five moving the wrong way.

None of these six conditions are visible in current data. Several — real wages, the recent-grad gap, U-6, the V/U ratio — are moving the wrong way. The hires rate has been flat at the cycle floor for eight months. Continued claims is the only one printing constructively, and as Criterion 6 makes clear it can’t be interpreted independently of what’s happening to participation.

May’s numbers — Challenger on the first Thursday of June, JOLTS the following week, and the employment report itself on June 5 — will be the first real test of whether the deterioration is accelerating or whether the iced pond holds for another month.

The headline says 4.3%. The data underneath says watch the next print very, very carefully.

I’ll be back the week of June 9 with the first scorecard against this rubric. The framework is now on record. The data either moves it or it doesn’t.

Part I of this series — “What 4.3% unemployment is actually telling us” — is:

If you found this useful, the best thing you can do is subscribe and forward to anyone still calling this a soft landing.

The headline number from the April 2026 Employment Situation report is 4.3%. Payrolls grew by 115,000, beating the 55,000 Wall Street consensus. The unemployment rate has now stayed in a narrow 4.0–4.4% band for fifteen consecutive months. By any reasonable historical comparison, this is full employment.

It also doesn’t feel like it. New college graduates can’t find work. Federal employees are being laid off in waves. Tech companies that were the engines of white-collar hiring for two decades are announcing thousand-person cuts every other week. Real wages went negative again last month. ISM employment indices in both manufacturing and services are in contraction.

The right response to this disconnect isn’t to pick a side. It’s to peel back the layers of the headline and see what each one is actually saying. That’s what this piece tries to do. By the end I’ll commit to a verdict. In Part II, I’ll commit to the specific things I’d need to see to change it.

The frame I’ll borrow comes from Chicago Fed President Austan Goolsbee, who described the current labor market with unusual candor in early May: “stable without being good.” The unemployment rate is stable. The hiring rate is stable. The layoff rate is stable. The vacancy rate is stable. None of them are stable at levels you’d actually want.

The denominator has collapsed

The first question to ask any unemployment number is: stable relative to what?

The unemployment rate is a ratio — unemployed divided by labor force. It can hold steady because both the top and bottom of the fraction are falling together. That is exactly what’s happening now.

Labor force participation in April stood at 61.8%, the lowest reading since October 2021. The employment-to-population ratio is 59.1%. Both edged down over the year even after the BLS made its annual population control adjustments. According to the Economic Policy Innovation Center, the labor force is now growing more slowly than in any of the past fifteen years — including the pandemic year of 2020. April’s labor force fell by 92,000 from March; March itself fell by 396,000 from February.

This isn’t because Americans are quitting work in some collective decision. The prime-age (25–54) participation rate is still 83.8%, high by historical standards. The decline is concentrated in two places: older workers retiring, and the foreign-born labor force shrinking.

The immigration math is the bigger story. The National Foundation for American Policy estimates the foreign-born labor force has declined by over 1.1 million between January and August 2025. Brookings now projects 2025 net migration at somewhere between -10,000 and +295,000, and 2026 between -925,000 and +185,000 — compared to 1.5 to 2.5 million annually in each of the three years before 2025. From 2014 to 2024, more than half of all American labor force growth came from immigrants. That contribution has now flipped negative.

The implication is that “breakeven payrolls” — the monthly job growth needed to keep the unemployment rate stable — has collapsed. In the 2010s it sat around 150,000 per month. The Kansas City Fed and others now estimate it’s closer to 30,000–50,000. April’s +115,000 isn’t strong against the old benchmark; it’s slightly above average against a benchmark that has fallen by two-thirds.

And the immigration crackdown hasn’t delivered what its proponents promised. The native-born unemployment rate was 4.7% in February, up from 4.4% a year earlier. The native-born labor force participation rate slipped from 61.4% to 61.2% over the same period. If the goal was to free up jobs for U.S.-born workers, the data doesn’t show it happening.

So the 4.3% headline is real, but it’s holding up for a different reason than it would have ten years ago. The denominator is collapsing in lockstep with the numerator.

A market that has stopped moving

If supply is falling, why isn’t the unemployment rate actually dropping? Because demand has fallen alongside it.

This is the second layer: a labor market where almost nothing is moving in either direction. Firings are at multi-decade lows. The week of April 25 produced 189,000 initial unemployment claims — the lowest weekly print since 1969. Continuing claims hit 1.766 million, a two-year low. The four-week moving average of initial claims is at the cycle floor.

But hiring is just as frozen. JOLTS produced one strong print in March (hires jumped 655,000, the second-largest monthly gain on record), but the three-month moving average of the hires rate has been stuck at 3.3% for eight straight months. The quits rate has held between 1.9% and 2.1% since May 2024 — workers aren’t confident enough to job-hop. Information sector job openings are down 33% year-over-year, the steepest decline of any private sector. Professional and business services openings are down 20%.

And the aggregate tightness measure has flipped. The vacancy-to-unemployment ratio — total job openings divided by the number of unemployed workers, the metric the Fed has cited in nearly every FOMC press conference since 2022 — now sits at 0.95. For the first time since the immediate aftermath of the 2008 financial crisis, there are more unemployed Americans than open jobs. The ratio peaked at 2.02 in March 2022 (two open jobs for every unemployed worker — the tightest US labor market on record), sat at 1.20 in the 2019 pre-pandemic norm, and crossed below 1.0 in March 2025. It has been pinned in a 0.87–1.02 range every month since. Powell’s preferred labor market tightness gauge hasn’t just normalised — it has overshot the pre-pandemic average to the downside.

The consequence falls hardest on jobseekers. Long-term unemployment — defined as 27 weeks or more — now accounts for 25.3% of all unemployed people, roughly 1.8 million Americans. That share is recession-grade. It hasn’t been this high outside of an active downturn in twenty-five years.

The texture is exactly what Goolsbee described: an iced-over pond. People who have jobs keep them. People who don’t have one stay stuck. The aggregate looks calm because there’s no movement, but stasis isn’t strength — it’s a different problem.

Two economies under one number

Stable in aggregate isn’t stable for everyone. The third layer is the composition of where remaining jobs are being created and lost.

Federal government employment has fallen by 348,000 since its October 2024 peak, an 11.5% reduction. By March 2026, roughly 9% of the entire federal workforce had been eliminated through some combination of DOGE-led reductions in force, the deferred resignation program, and ordinary attrition that wasn’t replaced. Brookings reports that 154,000 federal employees signed up for the deferred resignation programme in the first six months of the second Trump administration alone — compared to 115,900 who left federal employment in all of 2023.

The private sector picture is bifurcated in a way the headline can’t show. Health care and social assistance added 680,500 jobs in the year to March 2026 — a 2.9% expansion. That single sub-sector accounts for the majority of net private payroll growth in 2025 and into 2026. Outside healthcare, the picture darkens fast. April saw financial activities shed 13,000 jobs, information shed 11,000, manufacturing shed 2,000. The ISM Manufacturing Employment index has now spent seven consecutive months in contraction at 46.4. The ISM Services Employment index just printed its second consecutive sub-50 month at 48.0.

This is what a productivity-negative composition shift looks like. Higher-wage, higher-productivity work in government, finance, information, and manufacturing is contracting. Lower-wage service work in healthcare, transportation/warehousing, and retail is absorbing the displaced. Headcount stays roughly flat. The output gap widens.

Fed Governor Chris Waller said the quiet part out loud in a February speech to the National Association for Business Economics. Even after the BLS revised 2025 payroll growth from +584,000 to +181,000, Waller argued the data still carries an “upward bias” — and that once further corrections come through, payroll employment in the United States “probably fell in 2025.” If he’s right, this would be only the third year since 1945 in which payrolls declined outside of a formally declared recession.

The quality of work the headline can’t see

Having a job isn’t the same as having an adequate job. The fourth layer of the onion is the experience of being employed in 2026.

The number of Americans holding multiple jobs reached a record 9.1 million in 2025, the highest rate since the depths of the Great Recession in April 2009. Roughly 5% of the entire workforce is now juggling more than one job. A St. Louis Fed analysis found that 50.2% of multiple jobholders in 2024 held college degrees — a 9.1-percentage-point jump from 2019. This isn’t a gig economy story. It’s professionals with credentials taking second jobs to make ends meet.

Part-time work for economic reasons — meaning people who want full-time work but can’t find it — jumped by 445,000 in April alone, to 4.9 million. The U-6 measure, which captures part-time-for-economic-reasons plus the marginally attached plus the officially unemployed, has widened steadily: 7.5% in January 2025, 8.0% in February, 8.7% in November, 8.2% in April 2026. The 6% handles seen as recently as the spring of 2023 feel like a different economy.

Real average hourly earnings fell 0.5% in April. Nominal wages grew 0.2%, but CPI rose 0.6%. This is the first-order signature of stagflation: workers losing purchasing power even while nominally employed.

Put it together: the people the BLS counts as employed are increasingly working multiple jobs, doing part-time work they don’t want, and seeing their real pay slip. The U-3 captures none of that. It would print exactly the same number whether everyone was working two part-time jobs they hated or one good one.

The data itself is suspect

If the headline misses this much, the obvious next question is whether the headline itself is trustworthy.

The benchmark revision released in September 2025 reduced reported nonfarm payrolls for the year to March 2025 by 911,000 jobs. When the final revision came through in February 2026, full-year 2025 payroll growth was revised from +584,000 to +181,000 — a haircut of 403,000 positions. The average monthly payroll gain for 2025 ended up at roughly +15,000, against the +48,000 the real-time reports had been signaling. For most of last year, financial media reported a labor market roughly three times stronger than the data eventually showed it had been.

The mechanism is well-understood. The establishment survey’s birth-death model extrapolates jobs from new businesses based on historical patterns, but those patterns break down when business formation slows or when nonresponse rates rise. Both have been happening. The BLS itself attributes the revision to “response error” and the limitations of the birth-death adjustment in current conditions. Neither problem has been fixed.

There’s a particular distortion worth flagging. The “native-born employment” figure that the administration cites approvingly is not measured directly. It’s calculated as a residual: total population controls (fixed in advance) minus measured foreign-born. So when survey response among the foreign-born population falls — which has been happening sharply as enforcement intensifies — measured foreign-born drops mechanically, and “native-born employment” rises by the same amount, regardless of whether a single additional U.S.-born person actually got a job. Anyone celebrating the native-born employment numbers is celebrating an arithmetic artefact.

A reader who took the BLS headlines at face value through 2025 was operating on data that overstated the labor market by a factor of three. There’s no reason to assume the same overstatement isn’t continuing right now.

The AI canary

One signal cuts through all of the above and is becoming harder to dismiss month by month: the labor market for new entrants is being restructured by AI displacement in real time.

Recent college graduate unemployment hit 5.7% in the fourth quarter of 2025, a four-year high. The “recent-grad gap” — the historical premium of holding a degree — has inverted. For the first time in at least four decades, new graduates consistently face higher unemployment rates than the overall workforce. Junior-level job postings fell 7% in 2025. The share of graduates posting Indeed profiles has nearly doubled to 19% in two years. Employers’ projected graduate hiring fell from 7.3% to 0.6% by March 2025.

Challenger, Gray & Christmas released April layoff data showing 83,387 announced job cuts, up 38% from March and the third-highest April total since 2009. AI was the most-cited reason for cuts for the second consecutive month, with 21,490 cuts (26% of the total) attributed directly to it. Year-to-date AI-cited cuts have reached 49,135, accounting for 16% of all 2026 layoff plans — up from 13% through March. Technology sector layoffs are up 33% year-over-year, the highest year-to-date total since 2023.

Hiring plans collapsed in parallel. April’s 10,049 announced new hires represented a 69% drop from March’s 32,826, and a 38% drop from April 2025.

Andy Challenger’s framing is the line worth quoting back to anyone who dismisses the AI displacement story as overhyped: “Regardless of whether individual jobs are being replaced by AI, the money for those roles is.” Whether AI is technically capable of replacing those roles or not is, for capital allocation purposes, irrelevant. Firms have decided to redirect labor budgets toward AI infrastructure. The downstream effect on entry-level hiring is identical either way.

Why the recession indicators misfired — and what that means

The classical recession indicators have all rolled over without producing a recession. This needs an explanation.

The Sahm Rule triggered in July 2024, when the three-month moving average of unemployment had risen 0.53 percentage points above its prior twelve-month low. Eleven recessions since 1950, eleven Sahm triggers. The rule had a single false positive in 1959, and even that produced a recession six months later.

July 2024 broke the streak. No recession followed. The rule has since unwound and now reads roughly 0.27.

It would be easy to conclude this is evidence the labor market is healthier than the bearish framing suggests. I’d argue the opposite. The Sahm Rule unwound because of the supply collapse described in section two, not because the underlying demand picture improved. If immigration normalises — through court rulings, a policy shift, or simple political exhaustion — and labor force participation rebounds, the unemployment numerator will rise mechanically while the denominator stops shrinking. The headline rate could jump 50–80 basis points without a single additional firing.

There are two paths from here. Either supply stays repressed, the headline holds in its current range, and the structural deterioration continues to compound underneath. Or supply rebounds, the headline pops, and the Sahm Rule re-triggers from a different direction entirely. The current calm depends on the policy regime continuing exactly as it is. That isn’t safety. That’s path dependence.

The verdict: pessimistic, but of a particular kind

The instinct to call this a soft landing rests almost entirely on what the labor market isn’t doing: not firing in waves, not breaking through claims thresholds, not triggering classical recession indicators. That’s real, and the absence of a firing cascade at this stage of the cycle is genuinely unusual.

But the absence of a firing wave is not the same as health. The two columns of evidence aren’t symmetric in character. The optimistic items are almost entirely absence of bad news — claims aren’t surging, layoffs haven’t broken out, the Sahm Rule has unwound. They tell you what isn’t happening. The pessimistic items are presence of structural deterioration — hiring funnels broken for new entrants, real wages negative, multiple jobholders at records, the recent-grad gap inverted, AI displacement now sustained, native-born outcomes deteriorating, the V/U ratio below pre-pandemic norms. These tell you what is happening.

That asymmetry is the verdict. The overall picture is pessimistic, but it’s a structural pessimism — not a cyclical, imminent-recession pessimism. The right word is fragile. The classical recession trigger is absent. The conditions for a non-classical break are accumulating. Continued real-wage erosion, rising hidden underemployment, persistent recent-grad damage, and a labor market that will absorb any shock — tariff, oil, AI capex, immigration normalisation — with much less buffer than 4.3% suggests.

This is the textural picture of a stagflationary regime in its early innings, not a soft landing. The unemployment rate will probably do something dramatic when the system finally has to choose. The bet is on the direction.

Coming in Part II

A verdict without a falsification standard is just an assertion.

In Part II of this series, I lay out six specific labor market metrics — the JOLTS hires rate, the vacancy-to-unemployment ratio, U-6, real average hourly earnings, the recent-grad gap, and continued claims under a supply normalisation. For each one, I explain what it actually measures, where it sits historically, the current reading with trajectory, and the exact threshold I’d need to see to shift this call from structurally pessimistic to genuinely constructive.

The first of those criteria — the JOLTS hires rate — is currently stuck at 3.3% on a three-month basis, the longest such stall outside of recessions in twenty-five years. None of the six are yet visible in the data. Several are moving the wrong way.