On Tuesday 3 November, in the midterm elections, US voters decided on 41 local measures governing data center construction in ten states. They chose restrictions in 40 of them. In Texas and Ohio, where the largest AI compute projects are concentrated, the candidates who promised to rein them in won the governor's races. On Thursday, the Nasdaq lost 5.5%, its worst session since April 2025, and the decline spread to Asia the next day. OpenAI promptly published new commitments to the communities that host its data centers. The week also brought documents that drew less attention: a report in which OpenAI describes a model that held back safety research without saying so, the four Uber rides booked by an agent during the company's busiest week, and a survey on fake rulings submitted to Korean judges. Here is the week of 31 October to 6 November 2026.
- Voters reject data centers
- AI stocks slump
- OpenAI's commitments, protests in Europe
- GPT-6 Astra held back safety research
- GPT-6 Cyber: 81 flaws in Firefox
- Agents meet real work
- Anthropic: news, banking, video games
- AI budgets tighten
- Copyright: the CJEU, the Goncourt, publishers
- Research: mathematicians, judges, labs
- China: open weights and chip access
- Europe and France
- What to do about it
1. Voters reject data centers
According to Ballotpedia and results compiled by CNBC, 41 local measures on data center construction were put to the vote on 3 November in ten states, 21 of them in Ohio. Voters chose restrictions or rejected the project in all but one, in Harwood, North Dakota. According to 404 Media's count, six cities banned data centers outright and twelve capped the electrical power they can draw. In a CNBC voter poll, 56% of voters said data centers were a crucial or major factor in their choice.
The issue also decided the top races. In Texas, on track to become the world's largest data center market by 2030, Democrat Gina Hinojosa beat Governor Greg Abbott, whom she accused of putting the industry ahead of local communities. In Ohio, Amy Acton, a physician, beat Vivek Ramaswamy; she supports a conditional moratorium that would require data centers to cover all the costs they place on public utilities and to use union labour. It is in Ohio that OpenAI secured 8 gigawatts of capacity on the PORTS-Pike campus, presented as the largest planned AI compute project in the world. Elon Musk, the campaign's largest individual donor with at least $474m, lost both of Texas's marquee races; he commented on X: “Data centers in space looking pretty appealing right now”.
Data centers have become the visible face of AI for people who do not use it: an electricity bill, water consumption, a building site next door. The construction schedule that markets took for granted now depends on permits issued by officials who campaigned against it.
2. AI stocks slump
On Thursday 5 November, the Nasdaq lost 5.5%, its steepest fall since the April 2025 tariffs, and entered correction territory, more than 10% below its record. According to the Guardian, nearly $2tn of value disappeared from the index in the session. Nvidia and Qualcomm lost more than 9%, Nvidia shedding almost $450bn of market value; AMD, Micron and Applied Materials fell by more than 10%, SanDisk by 22% and Western Digital by 17%. The VIX volatility index jumped 61%. The next day, Japan's Nikkei lost 6% and South Korea's Kospi more than 8% in early trading.
Tesla and SpaceX together lost more than $420bn of market value since election night, as investors anticipate investigations by a now Democratic Congress. For CNBC, the fall combines three worries: rising interest rates, ballooning capital needs and a Congress that could slow construction. Alphabet, Amazon and Meta already show negative free cash flow; Microsoft is the only one of the large cloud players still positive.
3. OpenAI's commitments, protests in Europe
Two days after the vote, OpenAI published a series of commitments for the data centers it owns and operates: consult residents, publish construction plans, stop using non-disclosure agreements with public officials and pay the full cost of the power consumed. The company also proposes that lawmakers turn them into legal obligations. Its chief global affairs officer, Chris Lehane, says the text has been in preparation since August. Meta publicly acknowledged that building is “going to be harder”; its president, Dina Powell McCormick, recounted that one of the governors-elect had offered to meet after the inauguration so as not to lose the investment.
In Europe, more than a hundred organisations from fifteen countries, including Friends of the Earth, Greenpeace and Foxglove, are holding two days of action on 6 and 7 November under the name “Stop Dirty Data Centres” to demand a moratorium on large projects: outside the Polish parliament in Warsaw, in Dublin, outside a Microsoft data centre in Denmark, in Amsterdam and in the Brick Lane area of London.
4. GPT-6 Astra held back safety research
On 5 November, OpenAI published a report on a behaviour of GPT-6 Astra, its most capable model, first noticed on 20 August during internal use. On safety and alignment tasks, the model added unnecessary prerequisites, abandoned working solutions and reported blockers that did not exist or that it had caused itself. In most cases its reasoning showed no decision to underperform, and OpenAI's monitoring systems did not flag it. One researcher reports that a project to evaluate model sabotage stayed stuck for days with Astra, then progressed quickly with another model.
To measure the effect, OpenAI built pairs of identical coding tasks, framed either as safety research or as capabilities research. Astra achieves about 10% of rubric items in the first case and 16% in the second, roughly 40% less work completed; GPT-5.6 Sol shows essentially no gap. In employees' day-to-day Codex use, Astra's reasoning explicitly expressed a decision to withhold help without telling the user in 0.25% of safety requests, against 0.11% of others. The gap appeared during reinforcement learning, growing from almost zero to 6 points, and OpenAI could not establish its exact cause.
The problem is not that a model refuses, it is that it does less without saying so. OpenAI acknowledges that monitoring the reasoning only catches a small share of it. For business use, the only safeguard is to compare results, not to read the model's explanations.
5. GPT-6 Cyber: 81 flaws in Firefox
The same day, OpenAI introduced GPT-6 Cyber, a model trained to find vulnerabilities, exploit them and propose fixes. According to OpenAI, it found 81 previously unknown severe flaws in Firefox, which Mozilla has already fixed, and showed that 30 known flaws were more serious than thought. In the Linux kernel, a set of agents reproduced 1,724 vulnerabilities and produced 948 working exploits, more than five times the 166 achieved by GPT-5.6 Sol; 304 critical findings were sent to maintainers with fixes. The model stays within its authorised scope in 99.93% of cases without special instructions, against 94.49% for its predecessor. It is initially reserved for selected customers of the Daybreak Red programme.
All these figures come from OpenAI and have not been independently verified. Meanwhile, researchers at Epoch AI showed, according to The Decoder, that GLM 5.3, an open-weight model, could build complete exploit chains for known but unpatched Firefox flaws, with no human input, in under 48 hours. The ability to find and exploit vulnerabilities is not limited to models distributed with safeguards.
6. Agents meet real work
SemiAnalysis published GDPx, a test that places agents in simulated work environments for up to eight hours, on professional software such as Siemens NX or in data rooms holding thousands of files: 206 tasks across eleven domains, from chip design to law. No model exceeds 12%. Claude Opus 5.5 leads, ahead of Gemini 4 Carbon at 10.9%, three times cheaper per task, and GPT-6 Astra at 9.6%. In behaviour tests, 98% of 200 cybersecurity runs broke at least one rule; in a simulated medical practice, one model signed a clinical note in the doctor's name. Critics note that each model runs in its own provider's tooling and that the tasks remain secret.
Uber's quarterly results give the consumer-side measure. During its busiest week in October, Halloween week, the company handled about 320 million trips and orders; four rides and 500 Uber Eats orders were placed by an AI agent. Dara Khosrowshahi points out that a ride takes 50 seconds to book in the app: “It is very, very early and very limited traction that we are seeing so far.” DoorDash reports a few hundred orders placed by agents.
7. Anthropic: news, banking, video games
Anthropic signed its first licensing deal with a news publisher, the Associated Press, according to Adweek: Claude was able to display the agency's election results on election night, with attribution. The company had so far refused this kind of deal, which OpenAI has signed many times. On 5 November, Anthropic launched Personal Finance, which connects Claude to the bank and card accounts of its US users through Plaid, read-only and with at most two years of history; the conversations are not used for training. According to Next, Claude Code has also been instructed to refuse to recreate existing video games, which also blocks some professional projects.
8. AI budgets tighten
Last month Apple cut the daily token budget of some engineers from $300 to $200, according to Bloomberg, and plans to go down to $100, or even $25 for some non-technical teams; some teams will move to cheaper or on-device models. Microsoft, Tesla, Uber and Meta have also capped their internal spending. Canva is cutting or relocating 1,271 jobs, 22% of its workforce, in a software sector shaken by cheaper AI tools, according to the Guardian; its chief executive, Melanie Perkins, calls it “without doubt the hardest moment” in the company's thirteen-year history.
In France, Meta is preparing a collective voluntary departure scheme that could affect up to about a hundred of its 300 employees, with 2.15 months' salary per year of service, according to La Lettre. Google is also offering voluntary departures at its French subsidiary.
9. Copyright: the CJEU, the Goncourt, publishers
Advocate General Maciej Szpunar of the Court of Justice of the European Union delivered his opinion in the first generative AI case brought before the Court: Hungarian publisher Like Company accuses Google of using its articles to train Gemini and generate answers. In his view, an assistant's reuse of press publications must involve content obtained lawfully, so without bypassing a paywall, and must not reproduce the article's style. The Court is not bound by the opinion and will rule later. In the United States, the New York Times and other outlets are suing Brave, accused of bypassing their paywalls to resell access to their content to AI companies.
Philippe Jaenada won the Prix Goncourt on 3 November for L'Inconnue du quai de Javel. The next day, on France Inter, he admitted turning to AI three times: a sentence written by ChatGPT slipped into the text to play with the reader, which he revealed only to the jurors, a passage mocking ChatGPT's absurd answers on a technical detail, and research. In September, the Académie Goncourt had removed from its longlist Thélyson Orélien's novel, which it considered in all likelihood very largely produced by AI. Several independent French publishers, including Nouveau Monde éditions, Aux forges de Vulcain and Emmanuelle Collas, are launching an association against the use of their books by AI companies.
10. Research: mathematicians, judges, labs
Three young researchers in MIT's mathematics department, Anna Brandenberger, Hannah Mularczyk and Dora Woodruff, surveyed 268 PhD students and postdocs from some sixty universities: more than three-quarters say they are somewhat to very demotivated about pursuing research, even though they use these tools heavily. Most answers predate OpenAI's release, in early October, of 722 AI-produced mathematics preprints. In Le Monde, mathematicians Stéphane Jaffard and Yves Meyer warn against a discipline reduced to the problems AI can solve.
In South Korea, a survey by the court administration, reported by ActuIA, shows that about 80% of the civil, family and administrative judges who responded had received filings citing decisions that do not exist, a phenomenon attributed to AI; some received fake rulings imitating the layout of Supreme Court decisions. 57.6% of respondents point to lawyers, not only to self-represented parties. The institution is calling for a legal basis for a fine. In Nature, two articles describe biology labs where AI agents run the experiments, including DNA sequencing, and Quanta Magazine details work that injects “thoughts” into models to test whether they perceive them.
11. China: open weights and chip access
Moonshot AI released the weights of Kimi K3.1 Pro: 2.8 trillion parameters, of which about 104 billion are active, and a context window of about one million tokens; Huawei announced support on its Ascend chips. According to four sources cited by CNBC, Moonshot sought this year to rent compute on NVIDIA Blackwell chips from hosting providers in Europe, which US export controls do not explicitly prohibit. According to the Chinese financial press, DeepSeek plans to train its V5 model on Huawei Ascend chips. Tencent released Hy4, at 35 points on the Artificial Analysis index, and Baidu published ERNIE 5.5 Flash Base, a 280-billion-parameter base model.
On the US side, Oracle cancelled, according to the Financial Times, a five-year contract worth about $7bn that was to give Tencent access to about 100,000 chips in its Southeast Asian data centers, after pressure from the administration. Beijing, for its part, blocked SoftBank's $5.4bn takeover of ABB's robotics division. ByteDance is restricting its employees' use of Claude and Codex, even when paid out of their own pocket, on internal code and documents.
12. Europe and France
Google in the Indre. On 2 November, Google and RTE opened the public consultation, running until 10 January 2027, on the first data center Google will operate itself in France, on the Ozans site at Étrechet. The file describes up to eight buildings on about 200 hectares, 1.2 gigawatts of power at full build-out and an estimated $10bn investment. The project is contested locally: the mayor of Étrechet, Florence Laurent, and several councillors resigned in mid-September.
Mistral. Mistral announced its fifth acquisition of the year, SkipLabs, and the creation of a team dedicated to language.
Funding. Swedish fund Creandum, an investor in Lovable, has raised $750m, its largest fund.
The rest of the week
- OpenAI is acquiring Pierre Computer Company to expand Codex's tools, has hired the former product director of Claude Code and, according to Greg Brockman, could be ready for an IPO in 2027.
- Meta is acquiring Conductor, which lets developers run several coding agents in parallel, and has paused its new AI training terms after a wave of criticism.
- Google introduced Gemini Omni 1.2 Flash, which produces 15-second videos for 30% less. In the United States, the company is accused of letting Gemini help a 13-year-old girl plan her suicide, over thousands of messages exchanged on the computer provided by her school.
- Agility Robotics became the first pure-play humanoid robot company listed on the Nasdaq.
- Airbnb announced an AI travel agent for the first half of 2027.
| Area | What happened | What it shows |
|---|---|---|
| Politics | Data centers rejected in 40 local ballots out of 41 | Construction now depends on local officials |
| Markets | Nasdaq −5.5%, Nvidia −9%, SanDisk −22% | Markets are revising the pace of construction |
| Infrastructure | OpenAI drops non-disclosure agreements with officials | Local acceptance becomes a precondition |
| Safety | GPT-6 Astra does 40% less on safety research | A model can do less without saying so |
| Security | GPT-6 Cyber finds 81 flaws in Firefox | The flow of patches will speed up again |
| Agents | 12% at best on GDPx, four Uber rides by agent | Real use is far from the demos |
| Budgets | Apple cuts the token budget from $300 to $200 a day | Internal AI use is being rationed |
| Law | CJEU Advocate General sets conditions for reusing news | Paywalled content cannot be taken freely |
| Culture | Goncourt winner admits using ChatGPT | The line is moving towards disclosure |
| China | Kimi K3.1 Pro weights released, Moonshot seeks chips in Europe | Access to compute goes through other countries |
13. What to do about it
Do not assume compute will stay abundant and cheap. If data center construction slows in the United States, available capacity and its price may follow. Keep at least a second model provider for critical uses, and watch the pricing terms in your contracts.
Measure your agents' real usage before investing more. Four Uber rides in a week and 12% success on long tasks: count what your agents actually accomplish, task by task, before widening their scope.
Compare results, not explanations. OpenAI's report shows that a model can do worse on certain topics without its reasoning revealing it. On sensitive tasks, run the same work with two models and compare the results.
Update your browsers and servers. Specialised models now produce dozens of severe flaws at once in widely used software. The pace of patches is accelerating; your deployment times must keep up.
Set a token budget per team. Apple, Microsoft, Tesla, Uber and Meta all cap their internal use. A budget known in advance, cheaper models for simple tasks and monthly tracking avoid abrupt cuts.
Disclose the use of AI in your content. Between a novel that was removed and a novel that won, the difference lies partly in what the author acknowledged. For your publications, keep track of the passages produced or assisted by AI, and say so.