Anthropic, the San Francisco AI company behind the Claude assistant, told its investors this month that it will earn its first ever operating profit in the April-to-June quarter: about $559 million on roughly $10.9 billion in revenue. That is more than its entire lifetime revenue to date, and it arrives two years ahead of the company's own previous guidance. A separate funding round in progress would value Anthropic at over $900 billion, more than OpenAI. The catch sits in a footnote: Anthropic has agreed to pay Elon Musk's SpaceX and xAI about $1.25 billion every month for data-centre capacity, and the discounted ramp-up period of that deal lines up almost exactly with the quarter being celebrated. For European CIOs, the numbers matter less than what they signal about lock-in.

Krishna Rao, Anthropic's chief financial officer, has spent much of the past year explaining very large numbers to very serious people. On 20 May, the Wall Street Journal reported the next one: a slide deck shared with investors projected $10.9 billion in second-quarter revenue, up 130% from $4.8 billion in the first quarter, and $559 million in operating profit. CNBC and Bloomberg confirmed the figures the same day. The operating-profit line includes training costs but excludes stock-based compensation — a definition that matters, as we will see. Three days later, the venture investor and technology essayist Azeem Azhar called the projection “an extraordinary achievement and one set to feature in the annals of business history.” For a company that as recently as last summer told the same investors not to expect any full-year profit before 2028, the reversal is striking. Dario Amodei, Anthropic's chief executive, has shifted the public framing from “safety-first research lab” to “most important commercial AI vendor on earth” in barely eighteen months. Roughly 80% of revenue now comes from enterprises, with more than 1,000 customers spending over $1 million a year on Claude — a number that has doubled since February. The partnership announcements arrived in a steady drumbeat through May. KPMG signed a global alliance that embeds Claude into the Digital Gateway platform used by all 276,000 of its employees and their clients. PwC expanded its existing alliance, committing to train 30,000 staff on Claude and standing up a joint Center of Excellence. Both firms told clients that delivery times on agentic builds had compressed by up to 70%. In Frankfurt and Munich, where DAX40 companies still buy most of their consulting hours from the Big Four, the practical effect is that the dominant productivity tool in the next audit, tax review or transformation programme is now a single American model from a single American vendor. Not by accident, the funding round being marketed off these numbers is the largest in private-market history. Bloomberg reported on 12 May that Anthropic was in talks to raise at least $30 billion at a pre-money valuation above $900 billion, surpassing OpenAI's $852 billion mark from March. By the week of 26 May, TechTimes reported the round could close imminently. Investors are being asked to underwrite the implicit promise that the second-quarter numbers represent a new floor, not a one-off. That is precisely the assumption Anthropic's loudest critics dispute.

On 20 May, the same day the WSJ broke the profit story, SpaceX filed for an initial public offering. Buried in the prospectus was a single line: Anthropic would pay SpaceX $1.25 billion per month, through May 2029, for capacity at the Colossus data centres that SpaceX co-owns with Musk's xAI. Axios put a round number on it: $15 billion a year, on a vendor whose own annual revenue is roughly $18 billion. The filing also confirmed what Anthropic had not previously disclosed publicly: the payments are reduced in May and June 2026 as the deal “ramps up.” May and June 2026 are two of the three months in the quarter Anthropic is projecting will be profitable. This is the pivot in the story. Ed Zitron, writing in his newsletter Where's Your Ed At under the title “Anthropic's ‘Profitability’ Swindle,” argued that the operating-profit figure is largely an artefact of the ramp-up discount, not a structural change in unit economics. The AI researcher and critic Gary Marcus reached the same conclusion independently, noting that the size of the discount may itself exceed the $559 million in projected operating profit. If Anthropic paid the full $1.25 billion run-rate for all three months of the quarter, the math reverts to what it has been for years: a business whose cost of goods rises roughly in step with its revenue. There is a second piece of disquieting context. In a sworn affidavit dated 9 March 2026, filed in Anthropic's lawsuit against the U.S. Department of Defense, Krishna Rao stated that the company's revenues “exceeded $5 billion to date.” “To date” in that filing reads as cumulative lifetime revenue. By contrast, the investor materials describe a $30 billion annualised run-rate at the end of Q1 and a single quarter, Q2, larger than that $5 billion cumulative figure. Both numbers can technically be reconciled, given how steeply the run-rate has climbed, but the gap between what gets sworn before a federal judge and what gets pitched to a $900 billion funding round is, at minimum, an awkward one. A historical comparison sharpens the picture. In 2003 Amazon recorded its first full-year operating profit, $271 million on $5.3 billion of revenue, after eight years of losses; the market treated it as the moment the business model was vindicated. Anthropic, three years out of stealth, is projecting roughly double Amazon's 2003 operating profit on roughly double its annual revenue — in a single quarter, and partly on the back of a counterparty discount that expires. The shape of the curve is not the same.

Dry age-related macular degeneration (dAMD) is a slow-progressing eye disease that destroys central vision in tens of millions of older adults. It has no approved drug treatment. FutureHouse, a San Francisco non-profit founded by physicist Sam Rodriques and chemist Andrew White, has now published in Nature a multi-agent AI system called Robin that ran the full intellectual loop of biomedical discovery — reading the literature, forming a hypothesis, designing assays, analysing wet-lab data, and refining the hypothesis. Humans pipetted the cells. Robin nominated 30 drug candidates; five were tested; two looked promising. The standout is ripasudil, a Rho-kinase inhibitor sold in Japan for glaucoma — a drug no human researcher had ever connected to dAMD.

In a cell-culture room in San Francisco in early 2026, postdoc Ali Ghareeb watched a dish of retinal pigment epithelium cells engulf fluorescently labelled debris faster than they had any business doing. The molecule bathing them was ripasudil — a clinically approved eye drop for glaucoma, manufactured by Kowa, prescribed across Japan since 2014. Nobody had ever tested it for dry AMD. The reason Ghareeb tried it was that an AI system named Robin told him to. Robin is a workflow that orchestrates three of FutureHouse's specialised agents: Crow and Falcon for literature search and candidate evaluation, and Finch for data analysis. The paper that resulted, lead-authored by Ghareeb and his colleagues Michaela Hinks, Benjamin Chang, and Ludovico Mitchener, appeared in Nature on 19 May 2026 alongside a parallel paper from Google's Co-Scientist team. It is the first peer-reviewed demonstration of an AI system running the entire intellectual scaffold of a biological discovery — hypothesis, design, analysis, iteration — with humans relegated to the bench work. The loop ran like this. Robin began with a broad literature review and decided that enhancing retinal pigment epithelium (RPE) phagocytosis — the cellular housekeeping that clears photoreceptor debris and fails in dAMD — was a plausible therapeutic angle. Falcon ranked candidate molecules. Ten were tested in vitro. One ROCK inhibitor, the research-grade Y-27632, increased phagocytosis. Robin then proposed an RNA-sequencing follow-up, Finch analysed the reads, and the system flagged the lipid-efflux pump ABCA1 as the mechanistic story. With that signature in hand, Robin proposed a second, refined slate of candidates. This time the top hit was ripasudil — clinically approved, already used in the eye, with a known safety profile. “All hypotheses, experiment choices, data analyses, and main text figures in the manuscript were generated by Robin autonomously,” Sam Rodriques wrote in the company's research note. “Human researchers executed the physical experiments, but the intellectual framework was entirely AI-driven.” The entire arc — from conceiving Robin to submitting the paper — took two and a half months. That compression is the part that will discomfort pharma R&D leaders. Traditional small-molecule discovery for an eye disease typically runs ten to fifteen years from target identification to approval and costs upwards of two billion dollars per launched drug, according to industry estimates that have hardened over two decades. Drug repurposing — the pattern Robin was applied to — has always been faster, because it skips the worst of the safety pipeline. But repurposing has historically relied on serendipity, registry mining, or the kind of pattern-matching only senior clinicians with thirty-year careers manage to do well. Robin did it on a graduate-student timeline, in a lab the size of a small startup, by stitching together a glaucoma drug and a retinal degeneration on the strength of a phagocytosis pathway no human team had been chasing. Azeem Azhar, writing the day the paper landed, called it “AI finding the intellectual wormholes between specialised silos.”

The narrative pivot is worth pausing on. Robin is not, on inspection, a single monolithic AI. It is a small, interpretable orchestration of agents — three of them — each tuned for one job. That design choice matters because it is the opposite of where most of the frontier-lab discourse has been pointing. The fashionable assumption through 2025 was that scientific reasoning would emerge from ever-larger general-purpose models, given enough context. Robin is built from the opposite hypothesis: that durable scientific work needs structure, separation of concerns, and a human-readable trace. Crow is the deep-literature agent, drawing on FutureHouse's earlier PaperQA architecture to read and synthesise primary research at scale. Falcon ranks and evaluates candidate molecules against a hypothesis. Finch is the data-analysis agent, ingesting raw experimental outputs — cell counts, RNA-seq reads, dose-response curves — and producing the figures and statistical summaries that ended up in the Nature paper. A supervisor loop wires them together: literature → hypothesis → candidate set → human wet-lab → Finch analysis → refined hypothesis → next candidate set. Every step writes a trajectory log, which FutureHouse has open-sourced on GitHub alongside the Robin code. That openness is part of the story. Robin is released under permissive terms; the trajectories are inspectable; the agents themselves are available on FutureHouse's platform. For a field that has watched closed proprietary “AI scientist” demos from frontier labs accumulate without independent reproduction, having a working pipeline drop into the public domain on the day of publication is a different kind of event. Anyone with cell-culture capacity and a GPT-class API budget can, in principle, run the loop on a different disease tomorrow. The sceptical reading matters too. Resultsense, summarising critiques from The Conversation, noted that Robin's analytical agent “did poorly on statistics and bioinformatics questions and relied heavily on human-supplied prompts.” Two of five tested candidates being promising is real but small; ripasudil has not been validated in an animal model of dAMD, let alone in a human eye. Many drug candidates that clear cell-culture assays fail at the next stringency level. Karandeep Singh, who leads AI strategy for UC San Diego Health, told Nature: “You don't know how it works in reality until it's been made available to a broad set of people.” Critics also point out that natural-language reasoning, even when orchestrated, cannot model the quantitative complexity of biological systems the way structural tools like AlphaFold can. The honest reading is that Robin is a credible research instrument, not an autonomous scientist. Still, the context cuts the other way. Robin is the second major automated-science result in two weeks: Google Research's ERA paper landed days before, and Google's own Co-Scientist published in the same Nature issue, identifying three promising leukaemia repurposing candidates out of five tested. Two independent labs, two parallel Nature papers, one converging signal. The regime in which a small team can run an end-to-end discovery loop in a quarter, rather than a decade, has arrived in peer-reviewed form.

For two years the AI debate in boardrooms was a productivity argument. This week it became a legitimacy argument. Inside seven days, three very different audiences — graduating students in Arizona, a congresswoman in Washington, and a town council outside Frankfurt — pushed back against AI infrastructure in ways that traditional corporate communications cannot easily neutralise. None of these signals is, on its own, decisive. Taken together they describe a pattern: the public face of AI is becoming harder to defend in front of ordinary people. For European enterprises that have spent 18 months telling staff, customers and regulators that generative AI is a controlled, beneficial roll-out, the question is no longer whether sentiment is shifting, but how quickly that shift turns into permitting delays, employee resistance and political risk on home soil.

Eric Schmidt, the former Google chief executive, walked onto the University of Arizona stage on 16 May and within minutes lost the room. He compared AI to the personal computer. The audience booed. He tried again, telling graduates that the machines were coming but that they had agency, then delivered a line that had worked for a generation of Silicon Valley speakers: “When someone offers you a seat on the rocketship, you do not ask which seat, you just get on.” The booing got louder. NBC News and Fox Business both filmed it. Four days earlier, students at the University of Central Florida had booed Gloria Caulfield off the stage after she called AI “the next industrial revolution.” One UCF graduate, journalism student Houda Eletr, told reporters she had bought a cassette player to push back against algorithmic life: “It feels like humans are left on the back burner.” The line landed because it was the opposite of corporate scripting. A week later, on Capitol Hill, Representative Alexandria Ocasio-Cortez lifted two glass jars onto the witness desk of a House Energy and Commerce subcommittee. The water inside was the colour of weak tea. She told EPA Assistant Administrator Jessica Kramer that the jars were drawn from kitchen taps in Morgan County, Georgia, where Meta is operating a hyperscale data centre. Residents told her appliances had stopped working, water bills were set to rise by roughly a third, and the data centre was consuming about a tenth of the community's daily water. “Neither one of these things are drinkable, right?” Ocasio-Cortez asked. Kramer agreed and promised an EPA review. Bloomberg Law, HuffPost and Tom's Hardware all carried the image of the jars within hours. It was the kind of single-frame political moment that polling cannot manufacture. The third front opened in Germany. In early February the council of Groß-Gerau, twenty minutes from Frankfurt's data-centre cluster, voted 18 to 14 to refuse planning permission for Vantage Data Centers' €2.5 billion, 174-megawatt FRA5 campus. The coalition against the project crossed the political spectrum — SPD, Greens, FDP, Free Voters and Die Linke all voted no. The city's own urban planning office flagged terrorism risk, heat stress, doubtful business tax revenue and a paltry number of direct jobs. By May, AlgorithmWatch had published “How to Resist Data Centers: A Guide For Local Communities in Europe,” a thirty-page playbook lifting tactics from Friends of the Earth Ireland and the Catalan Guifi.net cooperative. Earlier in the year, the militant Vulkangruppe had sabotaged a Berlin substation, knocking out power to 45,000 households, and named data centres in its communiqué as a driver of an “insatiable” hunger for energy. Azeem Azhar, writing in Exponential View on 23 May under the headline “The AI backlash is the only thing growing faster than AI revenues,” put it bluntly: AI leaders, he argued, are poor ambassadors for their own technology.

The cleanest way to read these episodes is as a sequence, not a coincidence. The US polling layer is the oldest signal. A 21 May briefing already covered the headline number: roughly seven in ten Americans now say they reject generative AI in some part of their daily lives. That polling told boards sentiment was sour. It did not tell them what shape the resistance would take. The past fortnight has answered that question. Resistance is becoming local, physical and theatrical — a jar of brown water, a booed speaker, a no-vote in a Hessian council chamber. The DACH layer is the part that should worry German enterprises most. Groß-Gerau is not an outlier. Data Center Dynamics has logged a string of municipal vetoes across the Frankfurt belt over the past eighteen months, and the AlgorithmWatch guide is explicitly designed to industrialise that pattern. The legal architecture matters. Under German Baurecht, a §34 BauGB judgement about how a development “fits” into the character of its surroundings is at the discretion of the municipality. Frankfurt is now the second-largest data-centre hub in Europe after London, and the easy sites along the autobahn are already taken. The new sites are in places like Groß-Gerau, where councils have something the old sites did not: a precedent that says no is a viable answer. Historical comparison helps here. The closest analogue is the anti-nuclear movement of the late 1970s in West Germany — Wyhl, Brokdorf, Gorleben — where small towns proved able to delay, redesign or kill industrial projects backed by federal policy and global capital. That movement was also dismissed early as fringe. It ended up reshaping German energy policy for forty years. NIMBY movements against airports in the same period — Frankfurt's own Startbahn West, Munich II — followed a similar arc: local, visceral, eventually national. AI data centres are now travelling that road faster than any of those did, partly because the infrastructure footprint is bigger, partly because the social contract is thinner. Nuclear plants at least promised baseload electricity for everyone. A hyperscale campus exports its product to a hyperscaler's customers, not to the town it sits in. But Germany is different in one important way. The opposition is not principally green. The Groß-Gerau vote crossed party lines including the centre-right CDU's coalition partners. The objections were about jobs that did not materialise, business tax revenue that looked unreliable, district heating costs that the operator wanted the city to swallow, and security concerns. That is a far harder coalition to defuse than a single-issue environmental protest, because it includes the people normally counted on to wave projects through. Ben Thompson, writing in Stratechery on 18 May under the title “Data Center Discontent,” reached a related conclusion from a different angle. Communities, he argued, suddenly have a veto over AI that they did not have over globalisation, because AI has to land somewhere physical to exist at all. His proposed solution — pay communities off, generously and visibly — is logically sound and politically uncomfortable, and it does not address the Schmidt or Eletr signals at all. Those are not about money. They are about whether the people building the systems are people the public is willing to be ruled by.

On Wednesday May 20, 2026, Meta cut about 8,000 jobs — roughly ten percent of its workforce — and Intuit told 3,000 employees, seventeen percent of its global staff, that their roles were gone. Both companies named the same cause: AI is now productive enough that fewer humans can run the work. Meta is moving 7,000 surviving staff into four new AI organisations and has cancelled 6,000 open requisitions; Intuit is taking a $300–340 million restructuring charge for roughly $500 million in annualised savings by H2 2026. The day pushed the 2026 tech-layoff tracker past 113,000. The question for enterprise boards in Frankfurt, Walldorf and Munich is no longer whether AI changes headcount, but whether their next earnings call will sound like Mark Zuckerberg's or like Mercedes' quieter n8n rollout.

A Meta engineer in Singapore opened her laptop at 4:07 a.m. local time on Wednesday and found the screen frozen on a login error. The badge-disable mail had already landed in her personal Gmail: sixteen weeks of severance, two additional weeks per year of tenure, an HR portal link, and a short paragraph from Mark Zuckerberg about “running the company more efficiently to fund the investments we are making in personal superintelligence.” Within six hours the same email pattern fanned out across Dublin, London, Menlo Park and Austin. By the time the U.S. east coast logged on, roughly 8,000 Meta employees — about ten percent of the company — had been cut. Meta cancelled 6,000 open requisitions the same day and announced that 7,000 of the surviving employees would be reorganised into four new AI-focused groups covering products, agents, infrastructure and what an internal memo called “reality labs research after Reality Labs.” In total Meta restructured roughly 21,000 positions in a single week. Three time zones away in Mountain View, Intuit CEO Sasan Goodarzi was on a six-fifteen a.m. Zoom with his executive team rehearsing the language he would use at 9 a.m. “There is no good way to do this,” he told the all-hands later that morning, according to two people on the call. “But the work we do for customers is changing faster than our org chart, and we owe people honesty about that.” Intuit announced 3,000 cuts — seventeen percent of its 18,200-person headcount — in a Form 8-K filed alongside Q3 FY26 earnings. Affected U.S. employees receive sixteen weeks of base pay plus two weeks per year of tenure, with a final employment date of July 31, 2026. The cuts span seven countries and touch TurboTax, QuickBooks, Credit Karma and Mailchimp. Goodarzi told staff Intuit had signed multi-year deals to wire Anthropic and OpenAI models directly into the product surface; the company would rebuild around “smaller, AI-native teams shipping more.” The juxtaposition is the point. Meta reported record quarterly revenue of $56.31 billion the week before and raised its 2026 AI infrastructure spend to as much as $145 billion. Intuit is not in distress either — it raised guidance in the same 8-K that announced the layoffs. Neither company is cutting because demand collapsed. They are cutting because, in Zuckerberg's words to staff, “we are seeing more and more examples where one or two people are building something in a week that would have previously taken dozens of people months.” That sentence is the difference between this round and the 2023 wave. Three years ago, Meta, Google, Microsoft and Salesforce shed roughly 260,000 roles citing over-hiring during Covid and rising interest rates. AI was a rumour in the press release. In May 2026 it is the press release. The closest historical analogue is not 2023; it is the manufacturing automation milestones of 1985–1995, when robotic body-in-white lines at Toyota, VW and GM permanently reset the headcount per vehicle produced. The white-collar equivalent has arrived at Fortune 100 scale on a Wednesday.

Strip out the rhetoric and the financials tell their own story. Intuit's 8-K guides to $300–340 million in restructuring charges, largely in Q4 FY26 ending July 31, against an expected $500 million annualised run-rate saving by the second half of fiscal 2026. At blended fully-loaded cost per employee of roughly $170,000 across the affected geographies, the headline 3,000-person cut implies about $510 million in gross compensation removed — almost exactly the $500 million savings number, suggesting Intuit is reinvesting essentially nothing of the headcount budget into new hires elsewhere. That is a tell. In a normal restructuring, companies cut and rehire into priority skills; Intuit is cutting and not rehiring. The cash payback on the restructuring charge is therefore inside eight months, an unusually fast payoff that only works if AI tooling genuinely closes the productivity gap left by the departing workers. Meta's math is larger and harder to verify. The company has not disclosed a restructuring charge for the 8,000-person cut, but at Meta's blended cost-per-employee of roughly $400,000 (including stock-based comp), the gross compensation removed is on the order of $3.2 billion. Cancelling 6,000 open requisitions removes another ~$2.4 billion of avoided spend. Against that, Zuckerberg lifted 2026 capex guidance toward $145 billion, with the marginal increase funding GPU clusters, custom MTIA silicon and the four new AI organisations. The implicit trade is roughly five-to-one — every dollar saved in headcount is matched by ~five dollars in additional AI infrastructure. Meta is not cutting to save money. It is cutting to redirect compensation budget into compute and to demonstrate to public-market investors that the company can run leaner per dollar of revenue while it absorbs that capex bill. The productivity claim itself deserves harder scrutiny. Layoffs.fyi recorded 113,000+ tech roles cut across 179 companies year-to-date by May 18, with about 48 percent of those cuts publicly attributed to AI — up from under 8 percent in 2025 and 20 percent through Q1 2026. A December 2025 hiring-manager survey found 59 percent of respondents admit they emphasise AI in layoff communications because, as one respondent put it, “it plays better with stakeholders”; 17 percent explicitly blame AI when the real driver is financial. Ed Zitron has been hammering this point for months: CEOs rarely say AI is replacing workers, they say AI is making them “more efficient” — language that softens both the equity story and the legal exposure. The honest read of May 20 is that Meta and Intuit are doing both things at once. Real AI tooling is real, code-completion and customer-support automation are removing actual hours, and the AI framing also conveniently rewrites a routine cost-cut as a forward-looking strategy. The thing to watch is the next earnings cycle: if Meta's revenue-per-employee climbs sharply by Q3 and Intuit's gross margins expand without a step-down in product velocity, the AI productivity story holds. If revenue-per-employee is flat in twelve months, the cover-story critique will land harder than Zitron's most caustic post.