On 11 June 2026, Anthropic CEO Dario Amodei published a long policy essay called “Policy on the AI Exponential.” His core argument: AI is improving so fast that the slow rhythm of legislation can no longer keep up, and voluntary transparency is no longer enough. He proposes binding rules for frontier AI models — mandatory third-party safety tests, government power to block dangerous releases, support for workers displaced by automation, faster drug approvals, civil-liberties guardrails, and a democratic alliance to control chips and standards. Alongside the essay, Anthropic published a draft bill and pledged $350 million for labor-market research and a fellowship program. The essay matters because it is the most detailed regulatory blueprint any frontier-lab CEO has put on the table, and it lands while Washington and Brussels are both rewriting their AI rulebooks.

Amodei opens not with a statistic but with a story. Two Hobbits try to rouse Treebeard, the wise but ponderous tree-creature of Middle-earth, to defend his forest. Treebeard takes “a full day simply to say hello to another tree.” That, Amodei writes, is what AI policy looks like from inside a frontier lab. “In the several years that it can take Congress to act, AI can go from an amusing toy to the full country of geniuses,” he warns, invoking his own earlier image of a “country of geniuses in a datacenter.” The pivot in the essay is unmistakable. For two years, Anthropic’s public posture was that binding rules were premature; the company instead lobbied for transparency laws — California’s SB 53, New York’s RAISE Act, Illinois’ SB 315. “However, now the risks are clearly here,” Amodei writes. “It is time to go beyond transparency to more serious and binding regulation of AI.” The proposal is built around five “perennial policy areas”: regulation and public safety, macroeconomics and tax, accelerating beneficial science, civil liberties and state power, and democratic leadership in the AI race. The flagship recommendation is structural: regulate frontier models the way the Federal Aviation Administration regulates airplanes. Models above a compute threshold would undergo mandatory third-party testing in four risk areas — cybersecurity, biological weapons, loss of control, and automated R&D that could accelerate the other three. A government body, or a network of accredited private evaluators under a “regulatory markets” model, would have the power to block or reverse a release that fails. Anthropic did not just publish a manifesto. It also released a draft legislative text on frontier-model testing and a separate Economic Policy Framework on job displacement, and pledged $350 million: $200 million for an Economic Futures Research Fund underwriting empirical studies of wage insurance, retraining grants and capital accounts, and $150 million for a “Claude Corps” fellowship paying 100 early-career Americans $85,000 a year to deploy Claude inside US communities. The timing is not accidental. Amodei’s essay landed one day after the White House signed an executive order on “Promoting Advanced AI Innovation and Security,” and roughly a week after Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) unveiled their 269-page bipartisan “Great American AI Act,” which would pre-empt state AI laws for three years. Anthropic, in effect, is trying to shape the federal bill in real time.

Three forces converged in the first half of 2026 to make this essay possible. The first is capability. Anthropic’s own red-teaming of “Claude Mythos Preview” — the internal codename for its latest frontier model — showed, according to the essay, that frontier models now pose “very real risks” to cybersecurity, including disruption of the financial sector and critical infrastructure. Amodei lists biological risk and “serious AI autonomy risks” as the next likely thresholds, citing Anthropic’s recursive self-improvement work. In a frontier lab’s own telling, the abstract risk scenarios of 2023 have become measured red-team findings in 2026. The second force is political. Senator Bernie Sanders introduced the American AI Sovereign Wealth Fund Act on 2 June, proposing a one-time 50% equity tax on the largest AI firms payable in shares. Days later, President Trump told reporters that “there are concepts where pieces could be given to the American public, where the American public essentially becomes a partner” — endorsing the principle of government equity in OpenAI, Anthropic and xAI. Anthropic has publicly declined those talks. Read against that backdrop, Amodei’s essay is a counter-offer: bind us with FAA-style safety rules and labor-market support, but leave the cap table alone. The third force is legislative. The Obernolte-Trahan draft would force frontier developers above $500 million in revenue to publish safety frameworks, report critical incidents, and submit to semi-annual third-party audits — essentially the architecture Amodei now endorses, plus a three-year pre-emption of state law that Brad Carson of Americans for Responsible Innovation has called “a generational mistake.” The historical analogue is the period between the 2001 collapse of Enron and the 2002 passage of Sarbanes-Oxley: a fast policy window opened by undeniable evidence of risk, in which an industry decides whether to write the rules or be written into them. Amodei has clearly chosen the former. For European leaders, the framing is sharper still. The EU AI Act entered into force in August 2024; the Council and Parliament reached political agreement on the “AI Omnibus” simplification package on 7 May 2026, postponing high-risk system deadlines. Amodei’s essay never mentions the EU AI Act by name, but it does propose an internationally coordinated regime built on compute thresholds, third-party audits and supply-chain controls. That is a different architecture from the Act’s risk-based, use-case-driven approach — closer to how the FAA regulates aircraft than to how Brussels classifies systems. If Washington adopts a version of Amodei’s framework, the question for Brussels becomes whether to converge or to insist on its own paradigm. Either path has costs for DAX40 companies trying to build one compliance stack.

The reception split along predictable lines, but the substance of the critique is worth taking seriously. Politico described the framework as “the most aggressive regulatory framework any major AI CEO has publicly backed.” Venture capitalist Steven Sinofsky and several open-source advocates called it regulatory capture in everything but name: a compute-threshold, accredited-evaluator, secured-weights regime that an incumbent lab can absorb without strain but that a Series A challenger cannot. Anthropic already red-teams its models, already operates an Economic Index, already produces the documentation a regulator would request — the smaller the lab, the higher the relative compliance cost. Cognitive scientist Gary Marcus, long an advocate of stricter AI rules, welcomed the shift away from “regulate us, but later” rhetoric while warning that the proposed four-risk taxonomy could harden into the same kind of brittle checklist Amodei himself critiques in a footnote about California’s SB 1047. Substack analyst Hybrid Horizons titled a long takedown “Dario Amodei’s Policy Essay Argues Against Itself,” noting that the essay invokes the Collingridge dilemma — the impossibility of regulating a technology before its impacts are clear — and then proposes a fixed risk list as the load-bearing mechanism. Amodei pre-empts the capture charge in the essay itself: “People are worried about AI because they correctly perceive that its risks are real, not because AI CEOs have been insufficiently Panglossian.” He also insists that AI companies need governance separation — citing Anthropic’s own Long-Term Benefit Trust — because power should not be concentrated in companies any more than in governments. Whether that self-criticism is sufficient is the open political question of the next six months.

For a decade, the binding constraint on AI was chips. In 2026 it is power, water and planning permission. Hyperscalers cannot find enough megawatts on the grid, cooling towers are clashing with municipal water tables, and substations in Virginia, Frankfurt and Dublin are quoting interconnection dates in 2031. The proposed escape route is unusual: lift the datacenter off the planet. In low Earth orbit, sunlight is nearly constant, cooling happens by radiating heat into the vacuum, and no city council can deny a permit. Three frontier projects now share this thesis. Google has Project Suncatcher, the startup Starcloud has a working H100 in orbit, and SpaceX has just shown the world its AI-1 satellite. The question senior buyers must answer is whether orbital compute is a real procurement option, an investor narrative for an IPO, or both at once.

On a hangar floor in Bastrop, Texas, on 8 June 2026, Elon Musk stood next to a flat aluminium panel the size of a freight pallet and called it the future of compute. Beside him was Ian Dahl, the Starlink engineer who has run the orbital-datacenter programme inside SpaceX for the past eighteen months. Behind them, projected on a black wall, was a render of the AI-1: a 70-metre wingspan craft, wider than a Boeing 747-8, with two enormous solar arrays canted toward the sun and a 110-square-metre liquid radiator pointed knife-edge into deep space. “It is much simpler than a Starlink satellite,” Musk said, almost dismissively. “Solar cells, a radiator, laser links. That is the whole machine.” The pitch, delivered days before SpaceX’s planned IPO at a reported $1.77 trillion valuation, was that each AI-1 carries roughly 150 kW peak and 120 kW sustained compute, comparable to a single NVIDIA GB300 rack on Earth, and that SpaceX intends to manufacture them at an 11-million-square-foot “Gigasat” factory rising on the same Bastrop site. The target is one gigawatt per year of orbital AI compute by late 2027, with aspirational ten-fold annual growth thereafter. The technical story matters less than the corporate one. In February, SpaceX absorbed xAI, folding Grok’s training workloads into the Starlink balance sheet and giving Musk a single vertically integrated stack — launch vehicle, satellite bus, solar panel line, model lab — of a kind nobody else can replicate. The AI-1 reveal was not a science demo. It was the prospectus exhibit. Investors who buy the IPO are being asked to underwrite a thesis that the next gigawatt of frontier compute will be assembled in Texas and bolted to a Falcon. The historical analogue Musk likes is the undersea cable buildout of the 1990s, when private capital laid more fibre than the internet needed for a decade, and the survivors compounded. The analogue his critics prefer is Iridium — the 1998 satellite-phone constellation that filed for bankruptcy nine months after launch, undone by the gravity of terrestrial alternatives. Which precedent applies will be decided not in Bastrop but in the cost-per-FLOP arithmetic of the next thirty-six months.

Strip away the rendering and the engineering economics are unforgiving. A modern terrestrial AI campus delivers a watt of compute at roughly two to three dollars of capex amortised across a fifteen-year asset life. The most candid independent analyses of orbital compute — including a widely circulated teardown by Andrew McCalip and the IEEE Spectrum survey published last month — land at a three-to-four-times cost-per-watt premium for orbital, and that is before counting deorbit liabilities or the cost of redundancy against debris strikes. The drivers are mass and heat. Solar yield in a dawn-dusk sun-synchronous orbit is roughly eight times terrestrial output per square metre and runs near continuously, which is the genuine physics tailwind that brought Google’s research team to Project Suncatcher in November 2025. But every kilogram has to be lifted. At today’s Falcon 9 pricing of roughly $1,500/kg to LEO, and even at Starship’s projected $200/kg by the mid-2030s, the radiator mass dominates. The International Space Station rejects about 70 kW of heat using a 2,000-kilogram thermal loop; scaled linearly, a single AI-1 needs roughly three tonnes of plumbing just to keep the silicon below junction temperature. A gigawatt constellation, by independent estimate, would require something like 4,300 networked satellites massing thirty million kilograms in orbit, each unfurling a solar wing the length of half a football field. That is more mass than every object humanity has ever launched, combined. Bandwidth is the second wall. Optical inter-satellite links can move terabits between nearby craft, but the ground-to-orbit link still bottoms out at gigabit-class, which makes orbital compute suitable for training runs and batch inference but useless for the low-latency serving that pays the bills at OpenAI and Anthropic. Google’s Suncatcher paper concedes the point and positions orbital as a complement to terrestrial, not a replacement. Starcloud, the Nvidia-backed startup that put an H100 in orbit aboard Starcloud-1 in November 2025 and trained NanoGPT on the complete works of Shakespeare, raised a $170 million Series A in March on the same complement-not-replace pitch. Musk’s framing is more aggressive: he argues that terrestrial scaling hits a hard wall within a decade and that orbital is the only path to terawatt-class compute. The numbers say he may be directionally right on the wall and three to four times too optimistic on the price.

The strategic consequence for European buyers is that the orbital thesis has crystallised at exactly the wrong moment for Brussels. Six months ago, the EU’s sovereign-compute settlement looked complete: Deutsche Telekom and Nvidia opened the Industrial AI Cloud in Munich in February with 10,000 Blackwell GPUs and a billion-euro price tag, Mistral closed an $830 million debt facility in March for its Bruyères-le-Châtel campus, and the Commission was preparing to channel AI Act enforcement through these terrestrial, GDPR-bounded facilities. A frontier lab whose weights orbit outside any national jurisdiction breaks that frame. The Outer Space Treaty assigns state responsibility for objects to the launching state, which in the AI-1 case is the United States. ITAR covers the hardware export but, as JURIST and the European Journal of Law and Political Science both noted this spring, says nothing about model weights or inference performed in orbit. A DAX40 manufacturer that runs a sensitive process model on an AI-1 may be, depending on the analysis, simultaneously inside US export control, outside EU data jurisdiction, and unregulated by the AI Act for the orbital portion of the workload. The Commission’s emerging answer is the Orbital Triad — a coupling of the Space Act proposal, the AI Act and the Cyber Resilience Act — but it is years from binding force. In the meantime, procurement teams advising regulated clients face a real fork: contract with a sovereign terrestrial vendor at full price and full coverage, or with an orbital provider at a lower marginal cost but with unresolved legal exposure. The honest answer for the next two budget cycles is that orbital compute is not yet a regulated procurement category in the EU, and pricing it into any 2027 plan is premature.

Recursive self-improvement, or RSI, is the moment an AI system can meaningfully design and train its own successor without a human in the loop. It is the technical milestone OpenAI’s roadmap treats as the qualitative threshold to something it calls AGI-relevant capability. On 10 June 2026, Sam Altman told OpenAI staff in an internal Slack message that the company expects to go public within the next year, but also that an RSI takeoff would be a reason to stay private. The remark, first reported by The Information and confirmed by Reuters, makes a capability claim load-bearing for a public-markets timeline. For investors holding indirect OpenAI exposure through SoftBank, Microsoft equity, or late-stage venture funds, the IPO date is now contingent on what GPT-5.6 and its successors can do to themselves.

The message landed in OpenAI’s all-hands Slack channel just before midday Pacific on Wednesday, 10 June. Altman, writing in the conversational register he reserves for staff, framed the next twelve months as a planning exercise rather than a commitment. OpenAI had filed confidentially with the SEC on 22 May; Goldman Sachs and Morgan Stanley were lined up to lead; a target valuation between $852 billion and $1 trillion was being modelled. Yet the sentence that travelled was the caveat. According to The Information, Altman wrote that the faster the potential RSI takeoff looked, the more advantageous it could be to delay an IPO. Rapid capability gains, he argued, could create circumstances where staying private was preferable. Hours later, chief scientist Jakub Pachocki briefed engineers on GPT-5.6, codenamed internally as 5.6, a model he described as a meaningful improvement over GPT-5.5 and scheduled for release this month. Pachocki also walked staff through internal milestones: an automated intern-level researcher by September 2026, a full automated AI researcher by March 2028. To complete the package, Altman told employees the company was preparing a tender offer at $687.69 per share, the same implied valuation used in the S-1 modelling. The choreography is the point. In a single internal communication, OpenAI bundled three things normally kept apart: a public-markets timetable, a capability roadmap, and a liquidity event for employees and early backers. Each leg reinforces the others. The tender gives staff and angels a cash-out at IPO-adjacent prices without waiting for a roadshow. The model release primes the narrative going into the S-1 publication. The RSI caveat, meanwhile, gives Altman optionality: if the next training run delivers something genuinely recursive, he can withdraw the filing and frame the delay as responsible stewardship rather than as a stumble. The historical comparison is less Tesla’s Master Plan than Apple’s iPhone reveal cadence. Steve Jobs trained a public to expect a January announcement, a June ship, and a financial quarter calibrated around it. Altman is attempting something more ambitious: a capability disclosure cadence on which a $1 trillion listing is timed. It is also, importantly, the first time an AGI claim has been spoken in the same sentence as an SEC filing. The lawyers at Sullivan & Cromwell who are drafting OpenAI’s risk factors will have noticed.

OpenAI’s path to a public listing has been compressed and rewritten more often than any private company in living memory. The October 2025 reorganisation crystallised Microsoft’s economic interest at 26.79 percent of a new Public Benefit Corporation, OpenAI Group PBC, valued at roughly $500 billion. By March 2026 the company had closed a $122 billion primary round at an $852 billion post-money, led by SoftBank with participation from Amazon, Nvidia and a long tail of crossover funds: Altimeter, Coatue, D1, Dragoneer, Fidelity, T. Rowe Price. A $6.6 billion employee tender in May allowed roughly seventy-five staff to take $30 million each off the table at the $400 billion floor. The S-1 followed on 22 May. Q4 2026 became the working target, with September floated as the optimistic case. Against this run-up, Altman’s RSI caveat lands as a deliberate brake. The signal to the buyside is that OpenAI will not be pushed onto the tape by infrastructure debt or by SoftBank’s liquidity needs, which CNBC reported on 4 June as a growing concern around Masayoshi Son’s ability to service the $64.6 billion already committed to OpenAI by October. The signal to staff is different again: the tender is real, the price is set, but the IPO behind it is conditional on a capability event no one can yet timetable. The contrast with Anthropic could not be sharper. On the same day Altman wrote his Slack message, Dario Amodei published Policy on the AI Exponential, a five-pillar essay arguing for FAA-style pre-deployment testing, government authority to block frontier models that fail third-party assessment, and a $350 million Anthropic-funded package for labour-market dislocation. Anthropic also filed its own confidential S-1 the same week. Two of the three frontier labs are now within striking distance of a listing; their public framings are almost mirror opposites. OpenAI is offering capability as the trigger event. Anthropic is offering governance as the precondition. For Goldman’s syndicate desk, that is a genuine pricing question: does the marginal long-only fund pay a premium for an AGI option, or for a regulated-compounder narrative? The two playbooks recall Google’s 2004 dual-class moonshot prospectus and the 2010 SPAC-era trigger structures, but neither analogy quite fits. Google did not promise the search engine would rewrite itself. SPAC triggers were revenue gates, not capability gates. What Altman has introduced is a public-markets timing function indexed to a phenomenon that critics argue is not measurable and proponents argue is not yet here.

Not everyone reads the Slack message as candour. Gary Marcus, who spent the spring picking apart Anthropic’s own RSI claims as a bait-and-switch in which faster coding is rebranded as autonomous self-design, would put the odds of human-free recursive self-improvement before 2028 below ten percent. Ed Zitron, writing on Where’s Your Ed At, has spent the year describing the OpenAI-Anthropic capital cycle as the most annoying bubble in history and would treat the RSI caveat as a narrative escape hatch: a way to defer the only thing that matters, which is whether the unit economics of frontier models can survive a public-markets cost-of-capital. Both critics make the same structural point from different angles. RSI is unfalsifiable on any timeline an IPO syndicate cares about. If GPT-5.6 ships and accelerates internal R&D by twenty percent, is that RSI or just better tooling? The S-1 will have to draw that line, and the line will be argued by lawyers, not researchers. There is a second, quieter reading available to enterprise buyers. The RSI caveat may be less about delaying an IPO than about pre-emptively explaining a missed quarter after one. If OpenAI lists in Q4 and then misses revenue or burns through its compute budget faster than modelled, Altman has already planted the disclosure: capability moves faster than capital markets, and the company reserves the right to act accordingly. That is a useful piece of conditioning to have in the record before the bell rings.

For two years, enterprise engineering leaders treated AI tooling spend the way 2010-era CTOs treated AWS: a line item that grew faster than budget but was easier to defend than to question. The Pragmatic Engineer’s 11 June Pulse documents the inflection. Across Uber, DoorDash, an unnamed retirement-savings provider, and Gergely Orosz’s own circle, engineering departments are imposing token caps, throttling premium models, and challenging the productivity claims that justified last year’s budgets. The signal matters because the cuts are coming from inside engineering, not from CFO governance committees. When the people who fought for the budget start questioning the return, the FinOps conversation shifts from accounting to strategy — and from procurement risk to architectural risk.

On the Rapid Response podcast in late May, Uber president and chief operating officer Andrew Macdonald spoke in the careful cadence of an executive auditing his own enthusiasm. He recalled a conversation with senior engineering leaders elsewhere in the industry who told him, with evident pride, that 25 percent of code commits over the past quarter had been driven by AI. Macdonald’s response was the kind of question that, six months earlier, would have read as heretical: how many projects, he asked, actually moved from below the line to above the line because of those commits? “That link,” he said, “is not there yet.” Coming from the operating chief of a company that had already burned through its entire 2026 AI coding budget by the end of March — and whose CTO Praveen Neppalli Naga went viral telling The Information he was “back to the drawing board, because the budget I thought I would need is blown away already” — the question landed with the weight of an internal memo leaked in public. Gergely Orosz pulled the threads together in his Sunday newsletter. The Pulse of 11 June reads less like commentary than like a field report. DoorDash, Orosz writes, has pushed accountability for token spend down to individual developers: high monthly allowances, but justification required for overruns and a plan for next month’s efficiency. A large US retirement-savings firm has gone the other direction, imposing strict monthly Copilot limits and forcing engineers onto cheaper “0x” models once the cap is hit. Smaller startups are routing developers off raw API access onto subsidised Claude Code Max and Codex Max seats, where the unit economics are at least predictable. Orosz coined the term tokenmaxxing earlier in the spring for the leaderboard-driven adoption tactics that produced these bills; the June Pulse is the moment he documents the hangover. A week before Orosz’s piece, Sam Altman validated the shift from the supply side. AI budgeting, the OpenAI chief executive told an audience, has become “a huge issue” for customer firms — a complaint, he noted, that “never came up earlier this year” but had arrived “all of a sudden.” Bloomberg subsequently reported that OpenAI is preparing price cuts. The historical analogue is not 2022’s SaaS rationalisation but something older: the 2001 server-spend retreat, when CIOs who had ordered Sun racks by the cage suddenly demanded utilisation reports. Then, as now, the cuts started inside the technical organisation, not in finance.

The arithmetic Orosz lays out is unforgiving. Uber’s engineering organisation of roughly 5,000 developers saw Claude Code adoption climb from 32 percent in late 2025 to 84 percent by April 2026, propelled by an internal leaderboard that ranked teams on total AI tool usage — a gamification mechanic Orosz had warned about in earlier columns. The result: the entire fiscal-year 2026 AI tooling budget consumed in four months, and a corrective hard cap of $1,500 per engineer per month across Claude Code, Cursor, and the GitHub Copilot CLI. At fully loaded engineering cost in San Francisco, $1,500 per month is roughly four percent of a senior developer’s compensation. As a sanity threshold for tool spend, that is high; as evidence of where the unconstrained curve was heading, it is higher still. The macro picture aligns. OpenAI has disclosed that its single largest token customer now consumes 100 billion tokens monthly, a millionfold increase over six years, driven almost entirely by the shift from prompt-and-wait chat to autonomous agents that consume context continuously. Altman’s “huge issue” comment was the supply-side acknowledgement of a demand-side problem visible in every enterprise AP ledger. The structural explanation is simple: agentic coding workflows do not bill per request; they bill per turn, per tool call, per planning step. A single autonomous task can consume in hours what a developer used to consume in a week. The demand-side response is now organising into a doctrine. Ramp closed a $750 million round on 4 June at a $44 billion valuation specifically on the strength of its AI token spend-management product, which pulls usage data directly from model providers into corporate spend platforms. CEO Eric Glyman framed the thesis bluntly: for five centuries, business spend ran on people and vendors; in twenty-four months, a third pillar — intelligence, billed by the token and invisible to legacy procurement — has appeared. The FinOps Foundation used its June FinOps X keynote to declare token governance the discipline’s new mandate, complete with a proposed Tokenomics framework focused on cost per outcome rather than cost per call. Together with Orosz’s Pulse, these data points sketch a coherent picture: the first half of 2026 will be remembered as the moment the FinOps curve bent.

For DAX40 CFOs entering Q3 reforecast season, the Pragmatic Engineer signal arrives at a structurally awkward moment. Most German large-cap rolling plans submit revised opex envelopes in late June and early July; AI tooling has been the fastest-growing variable line item in nearly every IT budget filed since 2024. The temptation will be to translate the Uber story directly — impose a per-seat cap, force routing to cheaper models, declare victory. That approach manages cash but mis-reads what Macdonald actually said. His critique was not that AI tools are expensive; it was that engineering productivity metrics — commits driven by AI, percent of developers active — have been mistaken for business outcomes. A 25 percent AI-commit rate that does not move a single project from below to above the line is not a procurement problem. It is an alignment problem between engineering leadership and product economics. The more disciplined response runs in two layers. At the cost layer, treat token spend as compute, not as software: route by task complexity, cap by developer cohort, audit by team, and tie premium model access to documented use cases the way GPU quotas were rationed in 2024. At the value layer — the layer Macdonald was actually pointing at — require engineering organisations to publish, alongside the commit and adoption metrics they already report, a quarterly attribution of which shipped initiatives the tools materially accelerated. The bear case Ed Zitron has been making — that AI simply has no ROI — is the easy story to dismiss, because individual developers feel faster. The harder middle-ground question Orosz, Macdonald, and Altman have collectively put on the table is the one consultancies should be helping clients answer this quarter: AI ships features, but does it ship the right ones?