In March, Satya Nadella moved Mustafa Suleyman off Copilot. The Register called it accurately: the reorganization “frees him from the pesky need to show a return on AI investments.” Suleyman went sideways into a freshly minted MAI Superintelligence team, accountable for the kind of work that throws off press releases and recruitment letters rather than quarterly bookings. Two months later, on a YouTube stage with the Financial Times, he announced that within twelve to eighteen months, AI would perform at “human-level performance on most, if not all, professional tasks.” Accountants. Lawyers. Marketing. Project management. Everyone “sitting down at a computer.”

Read the sequence. The man is removed from the part of the org where revenue is measured, and almost immediately produces his most grandiose forecast of the year. The forecast is the new product. That is the only honest framing of the announcement.

What follows is not a thesis about whether the underlying technology is impressive. It is impressive, in narrow domains, in ways that matter. The thesis is about who the eighteen-month claim is for, what work it does in the world, and which interests it disciplines. The forecast is a financing instrument, a procurement weapon, a labor signal, and an immunization request, in roughly that order. It is also one of the cleaner specimens of late-cycle bubble rhetoric on record, and worth dissecting in those terms.

The capex denominator

Start with the number nobody quotes back to Suleyman when he speaks.

Microsoft’s fiscal 2026 capital expenditure guidance runs $97.7 to $120-plus billion, of which roughly two-thirds is going to GPU and CPU procurement. Across the four-firm hyperscaler cluster (Microsoft, Google, Meta, Amazon), 2026 capex is projected at $725 billion, up 77% from 2025. This is the largest concentrated industrial bet in postwar capitalism, narrower in firm count than the railroad bubble, denser in cash flows than the dotcom buildout, and almost entirely outside any public accountability framework.

Against that denominator, Microsoft’s stated FY26 AI revenue target is $25 billion. The ratio does not work without a story. The story has to be very large, very near-term, and very vivid, because the alternative is a write-down event of historic scale and an investor revolt that consumes the C-suite.

Suleyman’s eighteen months is that story. It is the only timeline that closes the gap between what was spent and what has been earned without forcing the firm to admit, publicly, that the bet is sized to a market that has not yet appeared. Once you read the claim with the capex line in view, the function becomes obvious. It is not a forecast. It is the verbal collateral on a balance sheet item.

What the data shows

Now the inconvenient ground truth.

The METR randomized trial on AI-assisted software development, completed in 2025, found that experienced developers using state-of-the-art coding assistants completed tasks 20% slower than the control group, while believing themselves to be roughly 20% faster. The gap between perceived productivity and measured productivity in that study is one of the cleanest single data points in the empirical literature on AI deployment. The technology was fastest at the easy parts of the work and created new failure modes (verification, hallucination triage, prompt rewriting, context loading) that consumed the savings and then some.

Torsten Slok at Apollo, working from S&P 500 margin data, has documented that Q4 2025 profit margin expansion was concentrated almost entirely inside the Big Tech cluster (up over 20% on the year) while the broader Bloomberg 500 index showed essentially no change. Wall Street consensus for the rest of the S&P 500 prices in zero AI-driven margin uplift through 2027. The real economy, by the market’s own pricing, is not getting an AI productivity dividend.

Gartner’s May 2026 study on AI-driven layoffs finds that automation-justified workforce reductions are systematically failing to generate the promised ROI. Eighty percent of white-collar workers, by Fortune’s reporting, are refusing AI adoption mandates from their employers, a form of soft strike that does not appear in the BLS data but appears in every implementation post-mortem. The MIT figure circulating in late 2025, that 95% of generative AI projects fail to deliver profit, has not been seriously challenged.

Dario Amodei, who in May 2025 forecast a 50% wipeout of entry-level white-collar jobs, walked his own claim back twelve months later, citing Jevons paradox: cheaper cognitive labor expands the addressable scope of cognitive work rather than collapsing total employment. Anaplan’s Charlie Gottdiener, writing in Fortune this week, offered the cleanest single-sentence framing on the deployment pattern: “AI isn’t eating software. It’s sorting it.” The split is between deterministic-domain work (still owned by structured systems) and probabilistic-domain work (where AI extracts value). Most professional tasks are mixed-domain. The deployment proceeds task-by-task, with mixed results, contested at the line level by workers who notice, accurately, that the productivity claims are not landing.

This is not a story of inevitable wholesale replacement. It is a story of targeted augmentation, expensive infrastructure, contested adoption, and missing returns. Suleyman is making his eighteen-month claim into the teeth of that evidence base, not in spite of it.

The man and the vocabulary

A brief biographical note, because the persona is part of the product.

Suleyman co-founded DeepMind in 2010, sold it to Google in 2014, and was placed on leave in 2019 over bullying complaints, eventually parting ways with Google in 2022. He then co-founded Inflection AI, which Microsoft “acqui-hired” in March 2024 in a transaction structured precisely to avoid FTC merger review: Microsoft hired the team and licensed the technology, leaving Inflection technically intact as a husk. Suleyman walked into a CEO title at Microsoft AI and a personal payout estimated in the hundreds of millions. The structure is not illegal. It is a designed evasion of antitrust oversight, executed by one of the firms whose lobbying budget shapes the oversight regime itself.

He published The Coming Wave in 2023. The book argues, with significant rhetorical skill, that AI is an unprecedented force whose containment is the central political problem of the era. The framing is technically correct on certain narrow points and structurally self-serving on most. The author benefits enormously from the public’s adoption of the inevitability frame the book promotes. He has, by his own count, given more than two hundred interviews advancing it.

He coined “artificial capable intelligence” (ACI) as the bridge category between large language models and AGI. This is not science. It is taxonomy as positioning: invent a label, place your shop’s product inside it, claim leadership of a category that did not exist until you named it. The same move was executed for “metaverse” by Mark Zuckerberg, with results now widely understood.

His current branding term is “humanist superintelligence.” The phrase is incoherent on inspection. Humanism, the tradition from Erasmus through Mill to Sen, makes the human capacity for reason, choice, and dignity the measure of value. Superintelligence is by definition a system that exceeds human capacity in every domain that matters, including the domains that constitute human dignity. The two concepts are mutually annihilating. The phrase exists to launder a project that displaces the human as the measure through the vocabulary of the tradition that places the human as the measure. It is rhetoric, not ideas.

His framing line from the FT interview deserves to be quoted exactly: “Creating a new model is going to be like creating a podcast or writing a blog. It is going to be possible to design an AI that suits your requirements for every institution, organization, and person on the planet.” This is the democratization claim made for every prior wave: by blogs (publishing democratized), by YouTube (broadcasting democratized), by Uber (entrepreneurship democratized), by crypto (finance democratized). In each case, the outcome was not democratization. It was extreme rent concentration in a handful of platforms, with the democratization rhetoric serving as the on-ramp for the concentration. Suleyman is reaching for the move again because the move works. It reliably converts public anxiety about technological power into public consent to the firms that hold the power.

The ideology under the forecast

The operative ideology in the eighteen-month claim is technological determinism, and it is doing political work whether or not Suleyman intends it.

The argument structure is: compute grows exponentially, therefore capability grows exponentially, therefore deployment grows exponentially, therefore replacement of white-collar labor is inevitable on a calendar timeline. Each “therefore” hides a choice. Compute grows because four firms allocate capital to it, because regulators decline to slow it, because energy grids are restructured to feed it, because immigration policy is shaped to channel talent toward it, because tax policy is shaped to subsidize it. Capability grows when the scaling laws hold, which they may not, and several recent results suggest they are softening. Deployment requires organizational redesign, change management, contract renegotiation, regulatory clearance, and labor consent, all of which are slow, all of which are political, all of which are contested. The “therefore” chain is a collapse of agency. Every step contains a choice that the rhetoric is at pains to hide.

This is the same ideological move that sold globalization in the 1990s. The China shock was not weather. It was a sequence of policy decisions (PNTR for China, WTO accession, dollar reserve dynamics, manufacturing tax treatment, container shipping deregulation) that produced enormous gains for some classes and the destruction of others. The decisions were sold as a force of nature (“you can’t stop the tide of globalization”). The class consequences are now well documented, and the political consequences are visible from any honest read of the last decade of American electoral results.

The AI shock is being sold with the same grammar. The class consequences are likely to be similar in shape if not in detail. The political consequences will follow.

The eighteen-month forecast does its political work at four audiences simultaneously, and at a fifth that has recently entered the frame.

To CIOs: spend or lose your job. The fear is not of buying too much AI. It is of being the executive who underspent while a competitor leapfrogged. This is procurement coercion. It moves billions of dollars without anyone in the room needing to produce a business case.

To regulators: do not bother. The deployment will outrun the legislative cycle. Any regulation you write will be obsolete before it ships. Better to step back and let the firms self-govern. This argument is recited inside antitrust hearings as standard testimony.

To workers: your leverage is gone. Take the offer. Do not strike. Do not unionize. Do not push back on the adoption mandate. The eighteen-month claim is, among other things, a wage-suppression instrument applied ex ante.

To shareholders: the capex pays off. The bet was correctly sized. Hold the position.

And now, as of the spring of 2026, to the Trump administration. Fortune is currently featuring an hour-long interview with Donald Trump in the Oval Office on the same homepage as the Suleyman piece. The AI sector is being positioned as a national strategic asset, deserving of energy grid priority, immigration carve-outs, antitrust forbearance, and direct industrial subsidy. The argument runs: if America does not build superintelligence first, China will. This is the Lockheed strategy, executed for a different commodity. Declare the firm essential to national survival, then negotiate terms.

The eighteen-month claim is efficient because it disciplines all five audiences with one sentence. That is what makes it worth saying.

The class question, recovered

Fortune frames its piece by invoking Henry Luce’s American Century. The frame is worth taking seriously and turning.

The twentieth-century American white-collar middle class was not an emergent property of capitalism. It was a constructed political settlement. The GI Bill produced the credentialing pipeline. Federal housing policy and the mortgage interest deduction produced the suburban household. Employer-based health insurance, structured by wartime wage controls, anchored employment to identity. Professional licensing regimes (the bar, the AMA, accounting boards) controlled labor supply in the high-wage cognitive sectors. Corporate income tax structures and capital controls kept retained earnings inside the firm, where they funded the employment relation rather than the buyback. Defined-benefit pensions tied the worker to the firm across a working life. Public sector employment, at federal, state, and municipal levels, absorbed an enormous share of credentialed labor on terms negotiated against rather than imposed upon the worker.

That settlement has been disassembled in stages over four decades. Offshoring of manufacturing in the 1990s. Public sector hollowing in the 2000s. Platform extraction of services labor in the 2010s. The credentialed cognitive class, which sat at the apex of the settlement, has been the last segment to absorb material downward pressure on real wages and conditions. Suleyman’s eighteen-month claim, whether or not the technology delivers, is being deployed as the closing move on that long disassembly.

The credentialed knowledge worker class was the social ballast of postwar liberal democracy. Hollow it, and you get the politics already visible across most of the OECD. Hollow it further, and the politics get worse. The fact that a senior executive at a $3.2 trillion firm can announce on YouTube, with no apparent political cost to the firm, that the bottom of the credentialed middle class will be unemployed within eighteen months is itself a data point about where political power currently sits. It is not a normal thing for a society to absorb. It is being absorbed because the political machinery for resisting it has been disassembled in advance.

Watch the actions, not the prophecies

The actions tell a different story than the words. Microsoft fired 15,000 workers in 2025 without citing AI, then quietly grew headcount in segments where AI deployment is most aggressive. IBM tripled entry-level hires for 2026 despite running one of the more aggressive AI internal deployments in the Fortune 500. ClickUp is operating at a three-to-one agent-to-human ratio while continuing to hire humans. Anaplan’s CEO is publicly bifurcating his product strategy between deterministic and probabilistic workloads, an admission that the wholesale-replacement narrative does not survive contact with deployment reality. Amodei, having been the loudest forecaster of mass white-collar replacement in 2025, has walked it back. The market priced in the SaaSpocalypse in February, then partially walked it back as the agentic systems failed to scale into deployment as predicted.

These are people hedging with both hands. The prophecies are for the press conferences. The actions are for the operating committee. When the two diverge as completely as they currently do, the honest analyst attends to the actions.

Close

The eighteen-month claim is not a forecast. It is a financing instrument, a procurement weapon, a labor discipline tool, a capex immunization, and now a national-security marketing pitch. It works at all five audiences in a single sentence, which is what makes it worth a senior executive’s stage time. The technology is real. The capability gains are real in some domains. The deployment is slow, expensive, contested, and so far producing margin compression rather than margin expansion outside the four firms that sell the picks and shovels.

The interesting question is not whether Suleyman is right. The interesting question is who benefits if the claim is treated as if it might be right, regardless of outcome. That list of beneficiaries is short, named, and visible on the leaderboards. The list of those who pay the cost of the wager, if the wager fails, is long, dispersed, and politically disorganized by design.

Read the move, not the prediction.