On June 2, 2026, President Trump signed an executive order asking — not requiring — the labs building the world's most powerful AI systems to hand them to the US government for testing before they ship. The window is up to 30 days. The focus is narrow: cyber capabilities, the ability of a model to find and exploit software vulnerabilities. Federal agencies will define what counts as a “covered frontier model,” and a new “AI cybersecurity clearinghouse” will coordinate vulnerability discovery with industry. Crucially, the order forbids itself from being read as a licensing regime: nothing in it permits Washington to gate the release of any model. It is the first frontier-AI governance signal of Trump's second term, and a deliberate counterweight to Europe's mandatory rules.

David Sacks stood a few feet from the Resolute Desk on Tuesday afternoon and watched the President sign a document the AI czar had spent the previous fortnight rewriting. Three weeks earlier, on May 21, the same ceremony had been scrubbed roughly three hours before showtime; Sacks had called the President that morning, according to reporting in Lawfare, and argued that a 90-day pre-release review would calcify into a de facto licensing regime — slow American labs, hand China the lead. Trump told the press he had “postponed” the order because he “didn't like certain aspects of it.” The version that finally crossed his desk had been pared back. The review window shrank from 90 days to 30. A clause was bolted on declaring that nothing in the order authorizes “a mandatory governmental licensing, preclearance, or permitting requirement for the development, publication, release, or distribution of new AI models, including frontier models.” In a briefing the same afternoon, Sacks insisted, in his words, that the framework is “not FDA for AI.” The operative machinery is thinner than the headlines suggest. The order tasks Treasury, the NSA, and CISA — within 30 days — to stand up an AI cybersecurity clearinghouse that will coordinate vulnerability scanning, validation, and patch distribution across the AI industry and critical-infrastructure operators. A separate workstream directs the same agencies to design a classified benchmarking process to measure “advanced cyber capabilities,” with the NSA director empowered to designate a model as a “covered frontier model.” The benchmark thresholds themselves will not be public. OMB is told to comb federal grant programs for money that could flow to vulnerability-detection research. Reaction split along predictable lines. OpenAI's chief global affairs officer Chris Lehane called the order “an important step forward” that “underscores that safety and innovation must advance hand-in-hand.” Anthropic — which had pushed for tighter language and which the Pentagon recently labelled a “supply chain risk” after a dispute over classified deployments — declined to comment publicly but was, per multiple outlets, deeply involved in the drafting. Senator Mark Warner, the Senate Intelligence Committee's Democratic vice chair, gave the rare bipartisan nod, while needling the White House for “belatedly” rediscovering an oversight architecture his colleagues say it dismantled in early 2025. The catch, civil-society critics argue, is structural: a voluntary scheme that explicitly cannot become mandatory invites the very labs being scrutinised to set the terms of the scrutiny. Lawfare's authors, who first sketched the “kicking the tires” model the order borrows from, concede the framework relies on “participation incentives rather than statutory enforcement.” In practice, that means a phone call from Sacks, not a subpoena.

The order is best read as the third act in a five-year arc. Act one was the Biden White House's October 2023 Executive Order 14110, which leaned on the Defense Production Act to compel reporting on any model trained above 10^26 floating-point operations — a threshold then roughly five times the estimated compute of GPT-4. Labs above the line had to share red-team results and safety evaluations with NIST. Act two arrived in January 2025: within hours of his second inauguration, Trump rescinded 14110, calling it a barrier to American competitiveness. For sixteen months the federal AI rulebook was, functionally, the AI Safety Institute's voluntary MOUs with Anthropic and OpenAI plus a patchwork of state laws — Colorado, California, Texas — that the administration has since attacked in a separate January 2026 executive order on state pre-emption. Act three is what shipped on June 2. The shape of it tells you who won the internal fight. Sriram Krishnan, the senior White House AI adviser who works alongside Sacks, has spent the spring publicly defending light-touch federal rules at international forums; Michael Kratsios, the chief technology officer who fronts the administration's AI policy at Davos and beyond, has called the EU AI Act “an absolute disaster.” The order they helped produce reads as the inverse of Brussels' instinct: classified rather than transparent thresholds, voluntary rather than compulsory submission, cyber-capability-narrow rather than systemic-risk-broad, and a statutory belt-and-braces against any future licensing regime. The contrast with the EU calendar sharpens this month. On August 2, 2026, the remaining general-purpose AI obligations in the AI Act become enforceable. Providers of GPAI models with systemic risk — the threshold sits at 10^25 FLOPs, an order of magnitude below Biden's old US line — must produce technical documentation, run adversarial evaluations, report serious incidents, and submit to the EU AI Office. The same Anthropic, OpenAI, Google DeepMind, and xAI models will, by autumn, be governed by a US regime where compliance is optional and a European regime where it is statutory. More remarkable is how the order is engineered to constrain the President's own successors. The explicit ban on licensing is unusual for an executive instrument that any future administration can rewrite; it functions as a political signal to industry that the regulatory floor will not move under Republican rule. The Council on Foreign Relations' assessment notes the order does not — and given its voluntary architecture, cannot — answer the deeper question of what Washington does if a frontier lab simply declines to participate. The benchmarking process is classified; the participation is optional; the enforcement is reputational. For two decades, US tech policy has oscillated between Section 230-style permission and FDA-style gatekeeping. Trump has just planted a flag on neither side.

At GTC Taipei, co-located with COMPUTEX, Nvidia moved its next-generation Vera Rubin platform from roadmap to production. The package has three layers a CFO should track. First, a rack-scale machine — the NVL72 — that bundles 72 Rubin GPUs, 36 Vera CPUs, BlueField DPUs and a cable-free spine into a single 120 kW unit Nvidia calls an “AI factory.” Second, a new CPU named Vera, an 88-core monolithic Arm part designed not for spreadsheets but for orchestrating swarms of AI agents. Third, an RTX Spark desktop superchip, co-built with MediaTek, that will ship inside Windows PCs from Microsoft, Dell, HP, ASUS, Lenovo and MSI this autumn. The strategic message: agents, not humans, will be the dominant consumers of compute, and Nvidia wants to rent enterprises a turnkey machine to host them.

Jensen Huang walked onto the Nangang Exhibition Center stage in the familiar leather jacket, but this time he was carrying hardware. He hoisted a Vera Rubin compute tray over his head — a slab of silicon, copper and liquid-cooling manifold that, he reminded the audience, contains more transistors than the entire Apollo guidance computer program shipped. “Useful AI has arrived,” he told a Taipei crowd that included Foxconn's Young Liu and TSMC chairman C.C. Wei. “Agentic AI is here. Vera Rubin is the machine that runs it.” The two-hour keynote on June 1, 2026 was less a product launch than a coronation of an architecture. Nvidia confirmed that Vera Rubin — first previewed at CES in January — is now in full production, with first racks already shipping to OpenAI, Anthropic, xAI, Microsoft, Meta, Oracle, CoreWeave and Dell. AWS and Google were named among adopters. Huang called it “the largest product launch, probably in the history of Taiwan.” Within hours, the TAIEX index broke 45,600 for the first time, dragged up by TSMC, Foxconn, MediaTek and Quanta. The headline numbers are unsentimental. Each Rubin GPU delivers up to 50 petaflops of NVFP4 inference and carries 288 GB of HBM4 at 22 TB/s. A single NVL72 rack — 72 GPUs lashed together by NVLink 6 at 260 TB/s of scale-up bandwidth — produces 3.6 exaflops of inference and 2.5 exaflops of training, with 20.7 TB of HBM4 on the spine. Nvidia claims a 10x improvement in cost per token over Blackwell. Power draw lands between 120 and 130 kW per rack. To run a row of these, you do not buy servers. You commission a substation. The second reveal was Vera, the CPU. Eighty-eight custom “Olympus” Arm v9.2-A cores on a single monolithic die — more cores in one socket than Tianhe-1A, the world's fastest supercomputer in 2010, had per node. Nvidia explicitly markets it as “the CPU for agents,” pitched at the latency-sensitive orchestration work of tool calls, sandbox execution and multi-step reasoning. Anthropic and OpenAI are named as first deployers. Then Huang pivoted off the data center entirely. He held up RTX Spark, a desktop-class superchip co-designed with MediaTek that pairs a 20-core Grace-derived CPU with a Blackwell RTX GPU, 6,144 CUDA cores and up to 128 GB of unified memory — one petaflop of AI in a laptop chassis. Microsoft, Dell, HP, ASUS, Lenovo and MSI committed to roughly 30 laptops and 10 desktops this autumn. The Surface Laptop Ultra and Dell XPS 16 Creator headline the lineup at $2,500 and up. The narrative pivot was unmistakable: Nvidia is no longer just selling racks to hyperscalers. It is trying to own the agent stack from desk to data center.

What makes Vera Rubin different from Blackwell is not the GPU. It is the system. Nvidia has, over five generations, gradually erased the distinction between “server,” “rack” and “cluster.” With Rubin, that distinction is gone. The NVL72 is a single addressable machine — one NUMA-coherent domain — wired with NVLink 6 at 3.6 TB/s of bidirectional bandwidth per GPU. The rack's 54 TB of LPDDR5x and 20.7 TB of HBM4 are presented to software as one memory space. Spectrum-X Ethernet stitches racks into pods, BlueField-4 DPUs handle east-west traffic and tenant isolation, and the entire chassis is cable-free — assembled, Foxconn engineers told reporters, in a fraction of the time a Hopper-era rack required. This matters because the workload has changed. A reasoning agent does not run a single, dense matrix multiplication; it runs a chain — retrieval, plan, tool call, sandbox, reflect, retry — that may dispatch hundreds of micro-tasks per user query. Mixture-of-experts models route every token to a different subset of parameters. Both patterns punish latency and reward memory bandwidth. Vera Rubin's 1.6 PB/s of HBM bandwidth per rack and Vera CPU's 1.2 TB/s of LPDDR5X exist to feed exactly this pattern. Dylan Patel's SemiAnalysis has called the design an “extreme co-design” evolution of Grace Blackwell Oberon — every part of the rack, from the substrate to the optics, redesigned around the same workload model. The Vera CPU is the most strategically interesting piece. By rejecting chiplets — the design choice that made AMD's EPYC and Intel's Xeon competitive — Nvidia is betting that agentic workloads need deterministic latency more than they need yield economics. Cross-chiplet hops add nanoseconds; nanoseconds add up when an agent makes a hundred sequential decisions before returning a single answer to the user. Each Olympus core runs two threads via what Nvidia calls Spatial Multithreading, with the Scalable Coherency Fabric tying them together. The pitch to a CIO is brutally simple: rent agent capacity, not vCPUs. The extended stack is the moat. Cosmos 3, Nvidia's world foundation model, shipped on May 31 in 16-billion- and 64-billion-parameter variants under a commercial open license. The Isaac GR00T Reference Humanoid combines a Unitree H2 Plus chassis, Sharpa tactile hands and a Jetson AGX Thor delivering 2,070 FP4 teraflops at the edge. Cosmos trains the model, Rubin runs it, GR00T embodies it, RTX Spark previews it. Foxconn is building a 10,000-GPU Blackwell AI factory for the Taiwanese government; Wistron and Quanta are integrating Rubin into sovereign-grade reference designs. Huang pledged roughly $150 billion in annual Taiwan spend and broke ground on a Constellation campus due to open by 2030. The risks are visible. AMD's Instinct MI400, due this year with 432 GB of HBM4, targets the same socket and the same workload, betting that open ROCm and superior memory capacity will win sovereign and hyperscale buyers focused on TCO. Intel's Gaudi 4 chases Ethernet scale-out economics. And the rack draws 130 kW — a number that lands harder in a Frankfurt colocation, where power costs three times what it does in Texas, than in any Nvidia slide deck.

In February 2025, Commission President Ursula von der Leyen announced InvestAI, a €200 billion umbrella for European AI, with a €20 billion carve-out to seed up to five ‘AI gigafactories' — sovereign compute campuses, each meant to house roughly 100,000 next-generation GPUs, that would train frontier models on European soil under European law. The European High Performance Computing Joint Undertaking (EuroHPC JU) runs the tender. The political logic was simple: without sovereign training capacity, the EU AI Act regulates a stack it does not own. Sixteen months on, the formal bidding window has slipped from December 2025 to July 2026, the field has thinned, and consortia like France's AION (€10B, filed May 28) are now operating without a clear answer on when — or how much — Brussels money will actually arrive.

On a humid Tuesday in early June, inside a glass-walled meeting room on the fifth floor of the Berlaymont, a senior official at DG CONNECT walked a delegation of consortium representatives through the revised gigafactory timetable. The chart on the screen showed a bidding window pushed from May to July, evaluation extending through autumn, and a footnote nobody had seen before: only two of the five planned facilities could be fully financed before the EU's next Multiannual Financial Framework opens in 2028. The rest would have to wait for political agreement on the next budget — an agreement that has historically taken three years to negotiate. One executive, according to a person in the room, did not bother to disguise his irritation. ‘We filed a €10 billion bid eight days ago,' he said. ‘We need a yes or a no, not a maybe in 2028.' Bloomberg's June 2 report, written by Jillian Deutsch and Samuel Stolton, put numbers on what had been corridor gossip for months. The pool of interested companies has shrunk from roughly 70 in January to about 10. At least two consortia are actively reconsidering. Germany's Schwarz Group — the Lidl parent that had been a marquee name in the Deutsche Telekom-led German bid — has quietly stepped back from the EU tender and is pressing ahead with an €11 billion data-centre campus of its own in Lübbenau, Brandenburg, designed for up to 100,000 GPUs. The signal could not be cleaner: a German strategic investor with hard cash and a real site has decided the EU subsidy is not worth the wait. EU tech-sovereignty commissioner Henna Virkkunen has defended the program in public, arguing in a recent op-ed that ‘building artificial intelligence the European way' requires patience and that the InvestAI structure — a layered fund with a Commission first-loss tranche, EIB management, and Member State top-ups — was designed precisely to attract private capital. The problem, several consortium leads argue privately, is that the structure assumes a demand forecast that the Commission has never published. Without a credible offtake story — sovereign cloud contracts, public-sector training workloads, anchor enterprise tenants — banks cannot underwrite the residual risk, and the de-risked InvestAI tranche is not large enough on its own to move a €10 billion project. The pivot in the story is that the program's slippage is no longer just a Brussels problem. It is showing up in the 2026 capex plans of the very enterprises the gigafactories were meant to serve. SAP, Allianz, BMW and Siemens all have AI roadmaps that, by Q3 2026, need a credible answer to a single question: when our most sensitive training and inference workloads need to sit on EU-owned infrastructure, where do they go? In the absence of a gigafactory date, the de facto answer is becoming Microsoft EU Data Boundary, Oracle EU Sovereign Cloud, and AWS European Sovereign Cloud — all running, ultimately, on Nvidia silicon shipped first to US hyperscalers. Nvidia's Vera Rubin NVL72, now in full production, begins partner shipments to AWS, Google Cloud, Azure, Oracle and CoreWeave in the second half of this year. Every quarter the EU's sovereign capacity sits on paper, the frontier moves further out of reach.

The InvestAI gigafactory program was conceived as Europe's answer to Stargate. In February 2025, von der Leyen presented it in Paris as a public-private fund with €20 billion earmarked for four — later raised to five — campuses, each at roughly 100,000 chips. EuroHPC JU was named as the procuring entity, mirroring the model that had worked reasonably well for the smaller €1.5 billion AI Factories program (which has already stood up sites including HammerHAI at HLRS Stuttgart and JUPITER at Jülich). The political pitch was that this would be the largest coordinated EU industrial-policy bet since Galileo, the satellite navigation program that ran roughly seven years late and €4 billion over budget — a precedent few in Brussels enjoy invoking. The original schedule called for an expression-of-interest phase in 2025, a formal call in December, and groundbreaking on the first sites in late 2026. Sites in Dresden, Munich, Frankfurt, Grenoble, Sophia Antipolis, Mora la Nova (Catalonia), and at least one Italian and one Nordic location were under discussion. By autumn 2025, the formal call had been pushed to early 2026. By spring, to May. By the June 2 Bloomberg piece, to July. Each slip has been individually defensible — more time for state-aid clearance, for grid-connection studies, for Member State co-financing — and collectively corrosive. Against this, the private sector kept moving. On May 20, the AION consortium — Iliad, Orange, Ardian, EDF, Scaleway, Capgemini, Bull, Artefact, with Hugging Face, Kyutai and Quandela on the model side — disclosed a €10 billion bid for a 200 MW campus scalable to 1 GW, anchored on France's nuclear baseload. On May 22, the German consortium of Deutsche Telekom, SAP, Ionos and (residually) Schwarz lodged its own submission ahead of a June 20 internal deadline. Days later, SoftBank announced up to €75 billion of data-centre investment in France alone — more than three times the entire EU envelope, with no sovereignty strings attached. The European Court of Auditors had already warned, in a 2024 special report, that the Commission's AI coordination framework was ‘still a work in progress' five years in, citing weak governance tools and missing monitoring. The ECA forecast that the EU's policy would miss its target of a 20% share of the global AI value chain. The Schwarz Group's quiet exit from the tender — €11 billion of private German capital walking past €20 billion of European subsidy — is the most expensive piece of feedback the program has received.

GitHub Copilot stopped being an all-you-can-eat subscription on June 1, 2026. Every Copilot plan, from the $10 Pro tier up to the $39-per-seat Enterprise SKU, now bundles a fixed allotment of a new virtual currency called GitHub AI Credits. One credit equals one US cent. Chat, agent runs, multi-file refactors and command-line work all draw from that wallet at rates that follow the underlying model's published token price. Inline code completions and Next Edit suggestions stay free. Heavier flows — Claude Opus 4.7, o3-pro, agentic ‘Spaces' sessions — drain credits fast. The base seat price did not change. What changed is the contract: a flat fee that used to buy unlimited inference now buys a budget.

The email landed in a Munich procurement inbox on a Monday morning. A DAX40 CIO, halfway through Q2 reforecasting, found the subject line ‘Action required: GitHub Copilot Enterprise billing transition' and forwarded it to FinOps with a single line: ‘How exposed are we?' The answer, by lunch, was: more than the spreadsheet thought. GitHub's announcement, posted by CEO Thomas Dohmke and product chief Mario Rodriguez under the headline ‘GitHub Copilot is moving to usage-based billing', framed the shift as inevitable. ‘The flat subscription was unsustainable because our GPU costs per heavy user were exceeding $200 a month,' Dohmke wrote. Premium Request Units, the half-measure GitHub rolled out in 2025, are gone. In their place: GitHub AI Credits, priced at one cent each, charged against the published API rates of whichever model the developer picks, with input, output and cached tokens all metered. The headline seat prices held. Business stayed at $19 per user per month, Enterprise at $39, Pro+ at $39, Pro at $10. What moved was everything underneath. A Pro seat now ships with 2,000 monthly credits — roughly fifteen GPT-4o chats a day, or about two days of serious Claude Sonnet 4.5 work. A Pro+ seat carries 7,000. Enterprise customers get a three-month grace pool through September 1, then shared org-level budgets with admin caps. Run an agentic session against Claude Opus 4.7, and a new multiplier — 27x for annual-plan holders, up from 7.5x — eats through that wallet in an afternoon. The reaction on GitHub's own discussion board was the kind of thing that used to take Oracle a full audit cycle to provoke. The official thread crossed 400 comments and 900 downvotes inside 72 hours. TechCrunch summarised the mood with a one-word headline quote: ‘What a joke.' The Register, less sparing, ran ‘Angry devs vow to flee GitHub Copilot as metered billing takes hold.' Power users on Pro+ posted screenshots of 8 percent of their monthly allotment vaporised in two hours of agentic refactoring; The New Stack and Visual Studio Magazine compiled bill projections of 10x to 50x for teams that had standardised on agent mode. The pivot, for anyone running Copilot at DAX40 scale, is not the sticker shock. It is the contract logic. GitHub spent a decade selling flat SaaS — the halcyon model that let Atlassian, Slack and Microsoft 365 grow into the enterprise without ever surfacing unit economics to the buyer. AI Credits ends that. It is the per-seat-to-metered switch that AWS pulled in 2013 when EC2 reserved instances quietly reframed cloud as a capacity-planning problem rather than a flat utility. Microsoft, which already runs Copilot Studio on a $0.01-per-message meter and is folding agent governance into the Agent 365 SKU bundled in M365 E7, now has a coherent story: every layer of the stack — model, IDE, orchestrator, governance — is metered, and the meter runs on tokens the platform alone can price.

Strip the noise and the arithmetic for a German Großkonzern is uncomfortable. Take a representative DAX40 with 8,000 Copilot Enterprise seats, the typical penetration after two years of rollout. At $39 per seat per month, the 2026 commit was $3.74m. Under the old flat contract, that number was the number. Under AI Credits, it is the floor. GitHub has not published the exact Enterprise credit allotment, but the public Pro+ benchmark — 7,000 credits per seat, roughly $70 of token budget — is a defensible proxy for a heavy-developer cohort. Assume Enterprise seats land at a similar headroom and a third of the population runs agentic flows daily. The Register's reporting and developer telemetry on the GitHub discussion thread put real consumption for agent-mode users at three to ten times that allotment. Mid-case: an extra $40 per heavy seat per month in overage, applied to roughly 2,600 developers. That is $1.25m of unbudgeted Q3-Q4 spend, or a 33 percent increase on a line item that was supposed to be flat. The model-mix sensitivity is sharper than that. GitHub's changelog confirms a 27x multiplier on Claude Opus 4.7 for annual-plan holders, up from 7.5x; GPT-5.4 rose from 1x to 6x; o3-pro carries the steepest token rate of the supported set. A team that picked Opus 4.7 as default last December because of its agent-loop quality is now paying close to four times what it paid in May for the same workflow. TokenCost's pricing tracker pegs a single multi-step refactor with full repository context at 400-600 credits, $4-6 of meter time, against an old flat allowance of zero marginal cost. The procurement consequence is the part the CFO will see first. Every framework agreement signed on a per-seat assumption — and most DAX40 Copilot deals were inked through Microsoft Enterprise Agreement renewals in 2024 and 2025 — has just become structurally mispriced. The contractual seat fee is intact, so the procurement team cannot reopen on price; what they can reopen is the credit-pool floor, overage cap and audit rights. Expect Q3 to be the quarter in which European CIO offices push for committed-use discounts on AI Credits the way they pushed for AWS Savings Plans in 2018. There is a secondary number worth watching: the cross-vendor arbitrage. Cursor's Pro+ tier is $60 a month with a credit pool that, by the comparative tests in The New Stack, runs longer per agentic session than Copilot Pro+ at $39. Replit dropped Core from $25 to $20 and launched a Pro tier at $100 with rollover credits — the first time a major coding tool has offered unused-credit carryover, a quietly important FinOps primitive. JetBrains' AI Assistant remains flat-priced for now, but the analyst consensus in the trade press is that the flat tier survives only as a marketing wedge. The whole developer-tools layer is converging on the metered economics OpenAI, Anthropic and the model providers already use. Copilot is not the leader of that move. It is the largest installed base to make the switch.