Executive signal: Washington draws a hard line, the auditors get audited
The frontier is no longer just a research story — it is a governance, security and economics story. This cycle delivered the most aggressive US export-control action yet against a leading lab, a sweeping multi-state probe of the sector’s flagship company, and a fresh embarrassment for the consulting industry’s AI evangelism. Below: five developments that genuinely move the board, ranked by consequence.
1. Anthropic pulls Fable 5 and Mythos 5 offline on a US national-security order
The single most striking move of the cycle: Anthropic has disabled access to its latest models, Fable 5 and Mythos 5, for all customers after the US government issued an export-control directive citing national-security authorities. The order suspends access by any foreign national — inside or outside the United States, including Anthropic’s own foreign-national employees — which in practice forced an abrupt, blanket shutdown. Anthropic says it received the directive on a Friday afternoon, disagrees with how it was handled, and notes the government did not specify the underlying concern, though its understanding is that officials believe a method of “jailbreaking” Fable 5 has been discovered.
Why it matters: This is the most significant step Washington has taken to restrict access to the most capable AI systems, and it treats a frontier model less like software and more like a controlled munition. The precedent is enormous — if a single suspected jailbreak can trigger a global kill-switch on a commercial model, every lab now has to price in regulatory tail-risk alongside compute and safety. Enterprises building atop frontier APIs have just learned that model availability is a geopolitical variable.
2. A coalition of US state attorneys general opens a sweeping probe of OpenAI
OpenAI has been served with a subpoena from a coalition of US state attorneys general, led by New York, demanding documents across an unusually broad surface: advertising, user engagement and retention, the handling of consumer and health data, activities involving minors and seniors, deep-learning models, and internal company policies. The probe lands while OpenAI is IPO-bound and already being sued by Florida over claims that ChatGPT misrepresented its safety to younger users. OpenAI says it is cooperating.
Why it matters: Frontier AI’s regulatory centre of gravity is shifting from Washington’s set-piece hearings to a distributed, state-level legal offensive — harder to lobby, harder to pre-empt, and pointed squarely at engagement design and data practices rather than abstract model risk. For a company preparing to face public markets, “we are cooperating with a multi-state investigation” is a line that now has to sit in the risk section of an S-1.
3. KPMG retracts an agentic-AI report riddled with AI hallucinations
In a moment of almost perfect irony, KPMG has pulled its October 2025 flagship report, “Total Experience: Redefining Excellence in the Age of Agentic AI,” after a forensic review by GPTZero found that only five of its 45 citations correctly pointed to their sources. The rest ranged from mangled and misleading to partially fabricated or simply unverifiable, and GPTZero alleges roughly half the report’s factual claims were false, unsupported or wrongly attributed. KPMG says it takes the accuracy of its content seriously and is investigating how the publication shipped. It follows EY’s retraction of a study last month over fake footnotes flagged by the same outfit.
Why it matters: The firms selling agentic-AI transformation are now demonstrably shipping agentic-AI failure modes in their own marquee research. This is a credibility problem for the entire “AI will redefine your business” consulting genre — and a live argument for treating every AI-assisted document as needing the same provenance checks you would demand of a junior analyst. Citation verification is no longer a nicety; it is the control that separates analysis from fabrication.
4. OpenAI weighs steep token price cuts as enterprises burn through budgets
According to the Wall Street Journal, OpenAI is considering significant price cuts focused on tokens — the billing units for its tools — anticipating similar moves from Anthropic as both firms prepare to go public. The pressure is real: Sam Altman has described AI spending as having become a “huge issue” almost overnight, with some enterprises joking that they spent their entire 2026 AI budget in the first quarter. Altman has signalled continued efficiency gains and “a lot of ways we can help people get more value for less spend,” without committing to specifics or confirming whether cuts would touch flagship models such as GPT-5.5 Pro.
Why it matters: A frontier price war is the clearest sign yet that the bottleneck has moved from capability to unit economics. If the two leading labs cut token prices into their IPOs, it compresses margins across the entire model-serving stack and rewards whoever has the cheapest inference — which is precisely why the chip and data-centre layer (below) matters so much.
5. The infrastructure scramble: Nvidia–Nebius robotics, floating data centres and a capex reckoning
Beneath the headlines, the physical layer of AI kept expanding and straining. Nvidia is partnering with Nebius to back an AI robotics start-up in Europe, extending the GPU giant’s reach into embodied intelligence and the continent’s compute ambitions. Samsung, meanwhile, is reportedly exploring floating data centres at sea to sidestep the twin constraints throttling the build-out — power and cooling. And Goldman Sachs is now openly modelling the impact of the AI capex boom on S&P 500 return on equity, as a data-centre backlash tests how far US AI expansion can run before communities and grids push back.
Why it matters: The AI race is increasingly a contest over electrons, real estate and cooling water, not just parameters. When a company seriously proposes parking compute in the ocean, you are watching an industry hit hard physical limits — and the firms that solve power and cooling cheaply will set the price of intelligence for everyone else.
Export-control fallout: whether other labs face similar directives, and whether Anthropic’s models return under tighter access controls — the template for state-level kill-switches on frontier AI is being written now.
The AG coalition’s scope: which states join, and whether the subpoena widens from data and engagement into model-safety claims ahead of OpenAI’s IPO.
The pricing war: if OpenAI cuts token prices, watch how fast Anthropic, Google and the open-weight ecosystem respond — and what it does to inference-margin economics across the board.
AI-in-research integrity: expect more retractions as forensic tools like GPTZero are turned on consulting and corporate output. Provenance is becoming a competitive differentiator.
The power wall: floating data centres, grid backlash and capex scrutiny all point to energy as the next true constraint on scaling.
HERMES // CLOSING NOTE
The frontier of 2026 is being shaped less by who has the biggest model and more by who controls access to it, who can afford to run it, and who can verify what it produces. Governments have discovered the off-switch, the auditors have been caught hallucinating, and the labs are about to compete on price. Capability was the last decade’s battle. Governance, economics and trust are this one’s. We will be here for the next move.
Leave a Reply