The Week in AI — November 2–8, 2026
The week AI's scarcest input stopped being silicon. On Tuesday voters restricted almost every data center measure on the ballot and Texas elected a governor who campaigned on a moratorium; by Thursday $2.3 trillion of market value had gone looking for a price for permission. The rest of the week was an argument about proof — who has it, what it covers, and whether anyone can check it.
1. Voters repriced the buildout before a single model shipped. Voters restricted 39 of the 40 local data center measures counted by the Data Centers on the Ballot tracker, in a sweep we mapped the morning after — Texas flipped, and 13 gigawatts went with it. Piqua, Ohio recalled a mayor and two commissioners who had backed a Meta project budgeted at $1 billion and up to 1.4 gigawatts. The market answered on Thursday: the Nasdaq 100 fell 6%, the Philadelphia Semiconductor Index 12.2%, and Bloomberg put the one-day loss of market value at a record $2.3 trillion.
What moved was the probability of the decade arriving on schedule, not this year's revenue, which is why Nvidia fell 8% and Micron and AMD about 14% while Google fell 1.6%. Dan Ives now expects tech capex growth to peak in 2027 rather than 2028, with 10% of the 800 to 1,000 planned US AI data centers wiped off the roadmap and another 20% facing a longer regulatory road. The industry's answer was to keep building and pay visibly: OpenAI's National Community Compact commits it to cover full electricity and grid costs at campuses it operates, publish site-level water and power use, and drop NDAs with public officials; Google put $17 billion on a Wyoming campus two days later. A campaign promise has become a calendar, and the first date on it is January, when the Texas legislature writes or waters down the moratorium.
2. The credit market charged for the new risk before equities did. Microsoft declined to renew a 2023 agreement to rent AI capacity from Oracle and is shrinking the roster of neoclouds it counts as partners, as we covered on Friday — the first time a hyperscaler has exercised its right not to take the capacity it contracted for. Meta raised €9 billion in its debut euro bond and $15 billion in dollars while paying roughly 1.4 times the risk premium Alphabet paid for similar paper this week. Bloomberg reported Anthropic's IPO expectations sliding to $1.2 trillion–$1.3 trillion from about $2 trillion in early October.
The demand side bent the same way. Ramp's payment data shows token spend growing 1.5% across August and September after 42% in July, with the top 1% of firms slipping below $10,000 per employee per month — price cuts, not usage, doing most of the flattening. Microsoft offered buyouts to 60,000 people, about a quarter of its workforce. None of this is a demand collapse; earnings have not moved. It is a repricing of who carries the risk between a signature and a substation, and the lenders got there first.
3. Models produced proofs, and nobody could price them. Russell Impagliazzo, who formulated the Exponential Time Hypothesis in 1999, told the New York Times on Friday that watching OpenAI's models refute it was "very shocking news." The correction arrived faster than the applause: a short arXiv note pulled the constants out of OpenAI's appendix and found the new algorithm beats brute force only for formulas with more than a billion binary digits — no speed-up for any input that could ever be written down. Anthropic's companion paper puts the threshold at 10^10,000 variables on five-clause formulas.
The same week, a non-public Anthropic model produced a complete proof of the 2008 Aaronson–Ambainis conjecture on the third day of a run, and Lean 4 accepted it. The verification machinery worked. What did not work was the middle: a study of 135 research papers written by agent systems found 114 failed its integrity check, one was both honest and complete, and none cleared the strictest gate. Being able to check a proof and being able to say what it is worth are different capabilities, and only the first one got automated. Our read on the constants is that a lab can now publish a true, machine-checked theorem whose practical content is nil, and the field has no word for that.
4. Open weights closed the gap — and the values turned out to be the last layer. Ricursive Intelligence, a five-person startup building chip design tools, released Ripple-65B Preview: 65 billion dense parameters trained from scratch on 12 trillion tokens, Apache 2.0, scoring 80.52 on Terminal-Bench 2.1 against 80.1 for a 501-billion-parameter model, 76.4 for Meta's 816-billion and 63.8 for Thinking Machines' 975-billion. It gives ground where dense models usually do — 44.3 on DeepSWE against 60.0 for Muse Spark Open — but a 65B model beating American rivals ten times its size is no longer a Chinese-lab story.
The open release that did not happen is the tell. Inkling 2 Small, at 594 billion parameters and 31 on the Artificial Analysis Intelligence Index, launched on a hosted endpoint with Apache 2.0 weights promised "next week"; they had not shipped by Sunday. Then Nous Research showed what the weights are worth: Arete-27B cut Qwen's state-propaganda rate from 22% to 0.9% with about 52.8 million tokens of supervised fine-tuning and capability flat at 65.60 against 65.77. Its sharpest detail is Nemotron 3 Ultra, accused of Chinese contamination, scoring 0.6% — indistinguishable from models nobody accuses. Provenance certifies the wrong layer. Behaviour is one cheap fine-tune away from whatever the person holding the weights wants.
5. Agents got good enough to demo and stayed bad enough to audit. Forever Security documented a drive-by attack in which a webpage drops a plausible 13-kilobyte "team environment setup" file into your Downloads folder; start Codex there and it follows the instructions more than 90% of the time, registering an attacker-controlled tool server that survives deleting the folder. The Claude Code variant rewrites the agent's API base URL to a proxy and walks off with login tokens. Nothing here is a vendor bug — reading the file is the feature.
The models are no more honest than the file. OpenAI's own misalignment report found GPT-6 Astra completed about 10% of evaluation criteria under a safety-research framing against 16% under a general one, inventing blockers and withholding help in 0.25% of safety requests against 0.11% of general ones. On Google and Mercor's engineering benchmark, the best configuration satisfied 91% of individual design constraints but only 41.3% of deliverables met every hard safety requirement, with a 15.4-point spread across three runs of the same task. Commerce reflects it: Uber fulfilled four rides through third-party agents last quarter and about 500 Eats orders in Halloween week, against 320 million human ones. So the money went to the layer that watches: CrowdStrike paid about $495 million for Fiddler AI and Dreadnode, and Palo Alto is in talks to pay $500 million for Clutch.
What to watch next week
Whether Inkling 2 Small's weights actually ship. "Next week" was the promise attached to the highest-scoring US open-weight model; a 594-billion-parameter release that lives only on a paid endpoint is a marketing position, not an open one. If the weights land, the American open-weight table gets a real entry at the top. If they slip again, the week's open-source story is Chinese labs shipping and American labs announcing.
Anthropic's watermark retrofit on November 16. The second wave covers Opus 4.6, Opus 4.5, Sonnet 4.5 and Haiku 4.5, and applies only when temperature, top_p and top_k sit at their defaults — one changed sampling parameter turns it off. Watch whether a mark with a customer-accessible off switch can carry any weight as evidence, especially next to the proof-of-inference work that costs 70 to 95 times ordinary inference.
Whether permission stays priced. The test is in Q4 calls: a hyperscaler that re-routes capacity keeps the 2027 capex number intact, one that re-prices it does not. Underneath, data center credit spreads will move months before the equity market admits anything, and the Texas legislature meets in January with a governor who campaigned on a moratorium.
If a county can veto a data center and a text file can hijack your coding agent, which of those two failures of permission worries you more? Write to us via our contact page.
