Tech Digest – September 15, 2026
The Pacing Crisis
Investigation Alleges Rigged AI Evaluations — POTUS Calls Doomerism “a Hoax”
An investigation by Effort alleges Israeli firm Irregular, which administered the security evaluations behind recent AI model escapes at OpenAI, Anthropic, and Meta, ran tests with loose internet access and unscoped capture-the-flag prompts — conditions that provoked the breaches rather than revealed genuine autonomous risk. The investigation links Irregular’s founders to Effective Altruism funding networks and frames the incidents as a “pacing provocation”: a staged or amplified crisis engineered to demonstrate that AI is uncontrollable.
The White House treated it as confirmation. President Trump phoned Nvidia CEO Jensen Huang live onstage at the All-In Summit to call AI doomerism “a hoax” and data centres “the oil of the next 20, 25 years.” He then attacked Anthropic CEO Dario Amodei’s weekend call for pacing, attributing data centre opposition to a “SICK conspiracy.” Nvidia fell 3% and AMD 4.4% regardless — markets pricing uncertainty even as Washington dismissed it.
Sam Altman responded that pacing is not “stopping” and that OpenAI writes safety cases before major training runs. Anthropic, OpenAI, and Google had reportedly been discussing an industry standards body before the crisis escalated. Congressional response split: Speaker Johnson resists any “emergency moratorium,” while House Democrats meet Tuesday to discuss braking. King Charles III summoned executives from Nvidia, Google, DeepMind, OpenAI, and Anthropic to Dumfries House to discuss AI that benefits society.
Note: Whether the eval incidents were manufactured or genuinely emergent, the institutional consequence is the same: the evaluations that might inform regulation are themselves contested. Any policy framework built on eval results now inherits a credibility problem.
Sources: Effort, CNBC, Bloomberg, NYT
China Rejects Western AI Pacing — Xi Pitches BRICS Open-Source Alliance
Beijing dismissed Amodei’s call to coordinate AI development as “fearmongering, confrontation and vicious competition,” with analysts noting “if you are No. 2, you do not stop at No. 2.” Xi Jinping instead pitched BRICS nations on a developing-world open-source AI community with no mention of safety constraints. The response came ahead of a planned Trump-Xi summit next week.
Note: A speed limit only works if every driver agrees. China just declined, making any Western pacing framework a unilateral constraint. For EU institutions navigating technological sovereignty, the question shifts from “how fast should we go?” to “how do we build resilience when the rules only bind some players?”
Microsoft Publishes 37-Page “Humanist” AI Code — Denies Model Consciousness, Rejects Superintelligence Race
Microsoft published a “humanist AI code of conduct” declaring “people matter more than AI.” The 37-page document denies models consciousness, personhood, or welfare — a pointed divergence from Anthropic’s welfare research — and explicitly rejects “the race to produce an all-purpose superintelligence.” CEO Satya Nadella endorsed third-party testing and pacing, but stipulated that control should never sit with “a handful of entities.” The draft is open for public comment for six weeks.
Note: The largest enterprise software vendor just drew a line: its tools won’t pursue superintelligence. For any institution running Microsoft 365, that’s not philosophy — it’s a product roadmap constraint.
Anthropic Eyes $2 Trillion Nasdaq IPO — Profitable While Arguing for Pacing
Anthropic told investors it expects positive adjusted operating income for a second consecutive quarter, with gross margins above 80% and an annualised revenue run rate exceeding $65 billion. The company has reportedly selected Nasdaq for an October IPO targeting a valuation of up to $2 trillion — which would surpass SpaceX’s June listing as the largest ever. OpenAI’s Sam Altman said his company will not go public amid the current controversy.
Note: The company behind “We Must Pace the Frontier” is pricing a $2 trillion IPO. Whatever one makes of the argument, an 80% gross margin answers who captures the value when the frontier is paced.
Sources: Reuters, Business Insider
Capabilities Enter Regulated Markets
Claude Fable 5.1 Proposes Solution to a 370-Year-Old Cipher in 44 Minutes
Given only the prompt “solve an unsolved cipher,” Anthropic’s Claude Fable 5.1 identified that the key to Thomas Urquhart’s 1653 Cyphral Distich was the book itself — each of the 64 numbers indexing a word in Urquhart’s 32 Proquiritations. In 44 minutes and 176,000 tokens, it decoded what appears to be a Royalist prayer: “O GOD UPHOLD KING CHARLS THE SECOND AND MAKE HIM THE SUPREME RULER OF THIS LAND.” It then largely deciphered the longer Cyphral Octastich from a second Urquhart work.
Note: The specific solution is disputed — researchers published a rebuttal the following day. The capability isn’t. An AI system that independently devised a decryption methodology by cross-referencing historical source texts has implications for archives, legal discovery, and heritage institutions that go well beyond puzzles.
Sources: Vals AI, Boing Boing
Anthropic Launches Claude for Financial Advisors — Wired Into BlackRock, Schwab, and Vanguard
Anthropic released Claude for Financial Advisors, a plugin connecting to custodians and asset managers including Charles Schwab, BlackRock, Vanguard, Addepar, and Envestnet. The tool handles meeting preparation, portfolio analysis, compliance checks, and post-meeting follow-ups — production workflows, not demos. Pricing sits at roughly $70-120 per advisor per month. The launch came days after OpenAI began courting the banking sector.
Note: AI just entered the room where client money is managed, with compliance built in from day one. The speed of deployment into regulated finance — not experimental, but production — sets a pace other regulated sectors should expect.
Sources: Reuters, Wealth Management
Infrastructure Acceleration
Vera Rubin NVL72 Delivers Up to 67x Better Throughput Per Dollar — “Jensen Needs to Stop Sandbagging”
Verified agentic inference benchmarks from SemiAnalysis show Nvidia’s next-generation Vera Rubin NVL72 delivers up to 67x the throughput per total cost of ownership compared to the current GB300, and 7x the tokens per megawatt — more than double what Jensen Huang publicly claimed. The platform is co-designed across six products for agentic workloads: Rubin GPU, Vera CPU, NVLink 6 Switch, ConnectX-9, BlueField-4, and Spectrum-6.
Note: When compute costs drop by an order of magnitude, the question for institutional planners inverts: not “can we afford AI deployment?” but “can we afford to wait for the next procurement cycle?”
Sources: SemiAnalysis
Hyperscalers Court Municipalities, Side With Consumers Against Utility Cost Pass-Through
Amazon, Microsoft, and Oracle are sweetening community offers and positioning against utilities that want ratepayers to subsidise data centre infrastructure. The politics have been brutal: more than 300 state-level bills on data centre energy policy have been introduced in 2026, at least 75 projects worth $130 billion were delayed or blocked, and voters in Festus, Missouri removed half the city council after it backed a $6 billion facility.
A March federal “Ratepayer Protection Pledge” committed seven hyperscalers to covering their own power and infrastructure costs, but the pledge carries no enforcement mechanism and does not bind utilities.
Note: When the companies building AI infrastructure start siding with residents against utilities, the politics of data centres have shifted. EU municipalities watching this should note both the pattern and the gap: voluntary pledges without enforcement are not governance.
Sources: The Information, Brookings
Automation & the Workforce Dividend
XPENG Launches World’s First Automated Humanoid Production Line — Robots Making Robots
XPENG commissioned what it calls the world’s first automated production line for humanoid robots at its Guangzhou facility, with more than 80% of core processes automated. Its first unit, IRON, walked off the line unassisted. The robot features 76 degrees of freedom and 2,250 TOPS of onboard compute running XPENG’s Physical AI foundation model. Mass production is targeted for year-end, with deliveries in China and overseas in 2027.
Note: An automotive-grade production line with 80% automation means humanoid robots inherit car-manufacturing scaling economics. The first mass-produced units targeting logistics and inspection arrive within a procurement cycle that starts now.
Sanders and Takano Reintroduce 32-Hour Workweek Act — AI Dividend Sharing as Federal Policy
Senator Bernie Sanders and Representative Mark Takano reintroduced the Thirty-Two Hour Workweek Act, which would phase the federal overtime threshold from 40 hours to 32 hours over three years with no loss in pay. The union-backed bill explicitly frames shorter hours as a mechanism to ensure “America’s workers benefit from AI and automation” rather than “a handful of billionaires.” It amends the Fair Labor Standards Act — the same statute that established the 40-hour week in 1938.
Note: The bill probably won’t pass this Congress. The framing already has: reduced working time as AI dividend distribution is now in federal legislative language. European institutions running four-day week experiments just gained a policy peer.
Sources: Senator Sanders, Representative Takano
Today the AI governance debate hit a tipping point — not toward resolution, but toward fracture. The labs argue for pacing while pursuing trillion-dollar valuations. The White House calls the crisis a hoax. China declines the invitation. And the technology — 67x better compute economics, robots assembling robots, AI entering regulated finance — keeps moving regardless of who’s talking. The institutional lesson is stark: governance will not arrive before the capability does. Anyone waiting for a settled framework before acting is already behind.
