The reproduction of labor-power requires not only a reproduction of its skills, but also a reproduction of its submission to the rules of the established order.

Louis Althusser, Ideology and Ideological State Apparatuses (1970)

The Three Tweets

Between May 4 and May 13, 2026, three principals posted versions of the same announcement. The sequence is the operation.

May 4. Anthropic, Blackstone, Hellman & Friedman, Goldman Sachs, General Atlantic, Leonard Green, Apollo, GIC, and Sequoia jointly announce a new “AI-native enterprise services firm.” Marc Nachmann of Goldman Sachs supplies the giveaway line. The firm will “democratize access to forward-deployed engineers.” Goldman has been democratizing things for 156 years.

May 5. Brian Armstrong, Coinbase. Fourteen percent of the workforce. About 700 people. The reason, per the CEO’s memo: “Two forces are converging.” One is “current market conditions,” which is the English-language phrase for Bitcoin is down 23% from its October 2025 peak and retail volumes have collapsed. The other is AI. The memo emphasises the second. The SEC filing leads with the first.

May 8. Matthew Prince, Cloudflare. Twenty percent of the workforce. About 1,100 people. Quarterly revenue: $639.8 million, a 34% year-over-year increase, the largest single quarter in Cloudflare’s 16-year history. Prince’s metaphor on the earnings call: “like going from a manual to an electric screwdriver.” Prince’s footnote, on the same call: “I would guess that in 2027 we’ll have more employees than we did at any point in 2026.” He said both. He meant both.

Three principals. Three press releases. One operation. The work of the pamphlet is to read them together.

The Beautifully Written Tweet

Armstrong’s letter is “really beautifully written.” It was almost certainly produced by Claude.

The Coinbase memo performs a specific operation. It takes a cyclical downturn in a speculative asset class and reframes it as a structural transformation made necessary by a technology. The crypto downturn is the cause. The AI transformation is the alibi. The 700 workers are the payload.

This is not new. This blog catalogued the mechanism in its first post: the AI displacement narrative is a wage suppression instrument, regardless of whether the automation it claims has occurred. The narrative does the work. The technology does not need to.

Coinbase’s stock structure makes the play visible. Market cap roughly $52 billion. Revenue volatility brutal. The investor base wants the equity priced as a high-growth tech company, not as a cyclical exchange whose business correlates 0.9 with the Bitcoin price. “AI-native operating model” is the phrase you write in the memo when you cannot write “the bag is heavy and we need to be smaller for the next leg down.”

The wage suppression operation does not require any of the laid-off workers to be replaced by AI. It requires the remaining workers to believe the next layoff round is coming whether or not they perform. That is the labor discipline mechanism, and it is operating exactly as designed. Total compensation per remaining Coinbase employee is the metric that matters. The press release is the instrument.

There is one tell in the Armstrong memo that the analytics class has noticed and the labor class has been trained not to. The memo announces no headcount target. There is no “we will end 2026 with X employees.” There is only the cut. The reason is straightforward. The next hiring cycle, when it comes, will be at a lower wage scale. Mentioning it now would defeat the discipline.

The Electric Screwdriver

Cloudflare’s case is the case that admits the trick. Prince, on the same earnings call where he announced the 1,100 cuts: more employees in 2027 than at any point in 2026.

Reconcile those two sentences. Either AI made the support workers obsolete or it did not. If it did, why hire more next year. If it did not, what justified the cut.

The reconciliation is the Jevons paradox. William Stanley Jevons noticed in 1865 that improvements in the efficiency of coal use did not reduce coal consumption. They expanded it. When a resource becomes cheaper to use, the use cases multiply faster than the efficiency improves. Prince knows this. He just told you. The electric screwdriver does not put screwdriver factories out of business. It makes more things require screwdrivers.

So why the cut.

The cut is a one-time reset of the wage scale. The cohort that gets re-hired in 2027 will be younger, hungrier, more credentialed in agentic tooling, and considerably cheaper. The wage compression is the operation. The Jevons rebound is the alibi for the operation having ever made sense as productivity policy.

This is the textbook mechanism Marx described in Capital Volume I, Chapter 25. The reserve army of labor is not a metaphor. It is a wage discipline structure. Capital periodically expands and contracts the active workforce to keep the cost of labor compressed against the floor set by the unemployed. The “AI restructuring” of May 2026 is one such expansion of the reserve army. The fact that the reserve will be partly reabsorbed in 2027 does not undermine the operation. It is the operation.

The class composition matters. Per Goldman Sachs economist Elsie Peng’s April 6, 2026 U.S. Daily note (cited in this blog’s $725 Billion Enclosure), net U.S. job loss to AI substitution is running roughly 16,000/month. Workers aged 22 to 25 in AI-exposed roles are down 16% in under three years. The cohort being severed from the credentialing ladder is the cohort that has not yet built the leverage to bargain. The Cloudflare CFO disclosed that the cuts will spare “salespeople who carry revenue quotas.” The wage discipline is calibrated precisely. It exempts the workers closest to the revenue and absorbs the workers furthest from it.

Cloudflare did not lay off Matthew Prince. Cloudflare laid off the 1,100 workers least able to refuse. This is not a flaw of the model. It is the model.

The Forward Deployed Engineer

Now we arrive at the contradiction the press releases cannot absorb.

On the same days Cloudflare and Coinbase are firing workers because they have been rendered obsolete by AI, Anthropic and a consortium representing the largest concentrations of private capital on earth are standing up a new firm whose entire purpose is to hire humans, lots of them, to sit inside customer companies and make AI work.

The job title is Forward Deployed Engineer. Palantir invented the role, called it “Delta,” and ran the company that way for over a decade. Up to 2016, Palantir had more FDEs than software engineers. The Pragmatic Engineer’s August 2025 deep dive on the role is candid about the function: the FDE alternates between embedded customer work and core product engineering, and Palantir could not have integrated its software into enterprises and government agencies without them. That is a more polite way of saying the product does not work without a human standing next to it.

OpenAI built an FDE team in early 2025. The Head of FDE at OpenAI has gone, on the record, to John Deere customer sites in Iowa to debug agentic deployments alongside farmers. Anthropic published an FDE job spec in early 2026 and on May 4, with the Blackstone consortium, blew the function up to industrial scale. The press release uses the verb “democratize.” The Goldman quote uses “expanding the number of highly skilled implementation partners.” The Blackstone quote uses “break down one of the most significant bottlenecks to enterprise AI adoption.” Translate. The model does not deploy itself. The frontier AI lab is hiring a consultancy. The consultancy will be wholly-owned by private equity.

The most expensive way ever invented to install enterprise software is to hire a $300,000 engineer to fly to a customer site, sit on the factory floor (Palantir’s Anjor Kanekar describes literally working on the Airbus final assembly line for years), debug the integration, escalate the unsolved problems to a research team, then come back the next month and do it again. This is not what the marketing claimed. The marketing claimed the agents would do this. The agents cannot. They cannot because the technical capability is real and partial, and the partial gap is exactly the gap a billable consultant fits into.

The actual business model: the hyperscaler captures the rent on compute, the model lab captures the rent on tokens, and the FDE captures the rent on the gap between what the demo showed and what the deployment requires. The consortium that just funded the new Anthropic-Blackstone services entity (General Atlantic, Leonard Green, Apollo, GIC, Sequoia, alongside the named principals) is not betting on AGI. It is betting that the gap between AGI marketing and enterprise reality is wide enough to drive a Palantir-sized truck through. Palantir’s market cap is over $400 billion. The truck is large.

This is the contradiction Coinbase and Cloudflare cannot acknowledge in their layoff press releases. If AI is so productive that 14% of a crypto exchange and 20% of an internet infrastructure company can be made redundant, what is Anthropic’s new firm selling to the rest of the Fortune 500. The answer is the same answer the labs sell to themselves. The model is impressive. The integration is hard. The integration is hard because the model still requires constant supervision, custom evals, prompt engineering, infrastructure tuning, and a human in the loop to recover from the failure modes. The Forward Deployed Engineer is the loop.

The Circuit

Three principals. One operation. Reading them together, dialectically:

The lie (Armstrong, May 5): AI is so productive your job is at risk.

The mechanism (Prince, May 8): The layoffs are temporary. The wage reset is permanent.

The contradiction (Anthropic / Blackstone, May 4): The AI does not work unless we send humans to make it work.

The narrative cannot hold all three at the same time. It does not have to. The audience for each tweet is different. The Coinbase memo is addressed to investors who want the multiple expanded and to the remaining workforce who want to keep their jobs. The Cloudflare call is addressed to the same investors and to a different segment of labor: the engineers being told they are now “two, ten, even 100 times more productive.” The Anthropic-Blackstone announcement is addressed to enterprise procurement officers who need to be told the technology is ready, and to the limited partners of every major PE fund who need to be told the consulting business is coming.

The three audiences cannot easily compare notes. That is the architecture.

The base-superstructure read, which is the read this blog has been working since post #001, holds cleanly. The base is the reallocation of capital from labor to compute infrastructure: $725 billion in the four hyperscalers’ 2026 capex alone. Cloudflare’s 34% revenue growth on a 20% workforce cut is not productivity. It is the rising organic composition of capital that Marx described in 1867, scaled to the largest fiscal-year capital reallocation in American corporate history.

The superstructure is the narrative apparatus that legitimates the reallocation in the eyes of the dispossessed. There are three layers, each visible in the three tweets:

1. The productivity narrative (Armstrong, Prince). AI is doing the work. This is the lie that disciplines the remaining workforce.
2. The temporary narrative (Prince’s 2027 admission). The pain is short-term. This is the lie that prevents collective response, because organising requires time and the targets believe they will be rehired.
3. The integration narrative (Anthropic-Blackstone). The capability is real, you just need our consultants. This is the lie that captures the rent on the gap between marketing and reality.

Each lie is, in isolation, partially true. The productivity gains are real. The job market is cyclical. The integration work is genuinely hard. The dialectical operation is that each truth, deployed selectively, conceals a structural fact about the others. The productivity gain is real and the wage suppression is also real and the second does not follow from the first. The cyclical recovery is real and the wage reset is also real and the wages will not recover with the cycle. The integration work is real and the model marketing systematically conceals this from the public. The three lies cohere because they tell three different audiences three different parts of one truth, none of which add up to the actual structure of the operation, which is class redistribution at a scale not seen since the breaking of the post-war labor compact in the 1970s.

This is what Antonio Gramsci called transformismo: the incorporation of potentially oppositional positions into the hegemonic project. The laid-off Coinbase support worker who buys the productivity narrative is not resisting. He is enrolling. The Cloudflare engineer who upskills for the 2027 rehire is not resisting. He is competing for a lower-paid slot in his own former job. The mid-career professional reading the Anthropic-Blackstone press release and considering an FDE role is not resisting. He is being recruited into the apparatus that produces the next round of his peers’ redundancy memos. The recruitment is, structurally, the function. Capital does not need to crush its critics. It needs them on the implementation team.

The Lie Is The Product

The structural insight lands cleanly. The Armstrong tweet was beautifully written. The Prince screwdriver metaphor was crisp. The Anthropic press release deployed the word “democratize” in a context that should make democracy file a defamation suit. These are not minor literary observations. The lie is not a defect of the product. The lie is the product.

What capital is selling, in May 2026, is not artificial intelligence. Capital is selling a narrative apparatus under which the cost of labor can be compressed, the gains from compute investment can be captured at the top of the stack, and the inevitable failure of the marketed capability to match the actual deployment can be monetised as a consulting service. Each of these revenue streams is large. Together they constitute roughly the size of the global GDP shift currently underway.

The model itself is real. The capability is partial. The press release is fluent. The Forward Deployed Engineer is hired. The workers are gone. The investors are paid. The consortium will return capital at well over its cost. The narrative will be sustained until the next cyclical reset, at which point the next cohort of CEO memos will be written by a model trained on the previous cohort’s memos, and the recursive consolidation of the wage discipline apparatus will continue until something physical (energy, a strike, a procurement scandal, the next sovereign debt crisis, a war that breaks a logistics chain a hyperscaler depends on) breaks the loop from outside.

This blog has, over its first 26 entries, named the principals, the mechanism, and the political economy. Nothing in this pamphlet is novel. The escalation in May 2026 is that the three operations (the lie, the layoff, the Forward Deployed Engineer) are now visible in the same nine-day window, in plain text, in language a procurement officer could parse. The architecture is not concealed. The architecture is described, on the record, with named principals quoting one another in approving press releases.

The fence is not going up. The fence went up. We are now learning what the rent will be.

Three things follow.

The Forward Deployed Engineer is not, structurally, an engineer. He is a Pinkerton with a laptop. His function is to absorb, on the customer site, the failure modes of a product that the marketing claimed did not have failure modes, so that the marketing can continue to claim the product does not have failure modes. The function is structurally similar to the role the Pinkerton agencies performed for the rail and mining cartels of the 1880s. The form is different. The labor relation is the same.

The “AI layoffs” of May 2026 are not, structurally, AI layoffs. They are coordinated wage resets executed under cover of a technological narrative that the CEOs themselves do not believe and that their own actions (the FDE hiring, the 2027 rehire admission, the Blackstone-fronted consulting buildout) systematically refute. The discipline is the product. The technology is the alibi.

The press releases are written, at some non-trivial rate, by the same model that the press releases describe. This is funny in a way that does not survive sustained attention.

Beautifully written tweets. Electric screwdrivers. Forward-deployed engineers. Three principals reading from the same prompt deck, telling three different audiences three partial truths that, together, conceal the operation that is class redistribution under cover of technological inevitability.

The pamphlet refuses the median. The principals are named. The mechanism is described. The lie is the product.

Read the next memo accordingly.