November 9, 2026
US voters pass at least 38 local limits on new AI data centres
This week the limit on AI turned out to be permission to build, not what the models can do. In Tuesday's US midterms, voters passed at least 38 of 44 local measures restricting data centres, the buildings where AI models are trained and run, falling mainly on projects planned for 2028. Investors priced the two biggest labs below their hopes: Anthropic will list its shares before Thanksgiving at less than its keenest backers wanted, and OpenAI's new round values it below September's target. The practical news was about doing more with existing machines: at their lowest thinking setting, four of seven top models kept at least 80% of their test scores for half the cost or less. Fewer new buildings means computing may stop getting cheaper as fast, so I am planning 2028 as if it gets dearer.
The Big Story
US voters pass at least 38 of 44 local limits on new AI data centres
In Tuesday's US midterms, which also gave Democrats both houses of Congress, data centres, the warehouse-sized buildings of computers that train and run AI models, were on many ballots. Of 44 local measures in 10 states that sought to restrict them, 38 had passed by Wednesday, according to a count by the polling firm Redbud Consulting; a day later, The Verge counted 43 passed. NBC News polling found 76% of voters opposed a data centre in their own community. In Texas, which was expected to account for more than 41% of the data-centre capacity added in 2028, governor-elect Gina Hinojosa has said she will halt all data-centre approvals until the legislature passes protections for residents; Bloomberg estimated the planned building at stake at about 13 gigawatts. Analysts quoted by Ars Technica said the measures largely spare projects already under construction or approved through 2027, and fall mainly on those planned for 2028. Axios estimated that nearly two-thirds of 2028 capacity could be held up the same way. The Nasdaq fell more than 5% on Thursday, in a tech sell-off that began with rising borrowing costs and quickened after the result.
Why it matters
For the next two years, what limits AI is whether towns let the buildings go up, more than whether anyone can get the chips. Anyone paying per request will notice nothing in 2027, because the measures largely spare what is being built until then; the pressure lands on 2028, the capacity the labs have been promising customers. An economist at Natixis told the South China Morning Post the constraints would keep US prices high. If your margin only works while per-request prices keep falling at this year's pace, model 2028 with them flat. The hedge is in this week's news too: Moonshot and Thinking Machines are putting out open models you can run on hardware you rent or own. If you just chat with an assistant on a paid plan, you are likelier to hit tighter usage limits than a higher price; Grok tightened its limits this week, citing costs.
Less Money Than Asked For
- Anthropic keeps its pre-Thanksgiving listing and accepts a lower valuation
Anthropic plans to list its shares before Thanksgiving, raising $80-100 billion at a valuation of $1.25-1.75 trillion, according to investors quoted by the Financial Times. That is more than its current $965 billion and less than the $2 trillion its most optimistic backers hoped for. Its amended filing shows $33.24 billion of revenue for the first nine months of 2026, with quarter-on-quarter growth slowing to about 47% in the third quarter from 144% in the second, and about $895.4 billion of purchase and lease commitments, nearly all for computing. Its computing bills are fixed for years and its customers can cut back any month, so buildings that arrive late start to pinch.
Read more → - OpenAI takes a first $11 billion at $1.25 trillion, below September's $1.4 trillion aim
OpenAI has received $11 billion, the first money from a round that values it at $1.25 trillion, with more expected within weeks. In September it was aiming for $30 billion at $1.4 trillion. A round priced below its own target, with the rest still to arrive, leaves OpenAI less room to keep charging little for each answer while its computing bills grow, so the generous free tiers are what to watch.
Read more → - DeepSeek plans to train its next model on Huawei chips rather than Nvidia's
DeepSeek plans to pre-train V5, its next model, on Huawei's Ascend chips, a Chinese financial paper reported, citing industry sources; until now those chips have mostly run finished models rather than trained them. Pre-training, the long first pass in which a model learns from raw text, is the most demanding job in AI, which is why it has stayed on Nvidia hardware. For anyone building on DeepSeek's models, the next one would no longer queue for American chips and buildings behind everyone else's. The plan has not been independently confirmed.
Read more →
Faster and Cheaper Without New Buildings
- Anthropic stops agents resending their whole conversation every turn
Anthropic's message threads, now in public beta, let an agent, an AI that works through a job step by step on its own, send only what is new since its last turn instead of the whole conversation. Anthropic holds the thread in memory, encrypted and never written to disk, until 5 minutes after the last request. On requests of 500,000 to 1 million tokens, several novels' worth of text, the median wait before Claude starts answering fell by about 35%, and it is on by default for most Claude Code users. The bill stays the same, because Claude still reads the whole conversation, so switch it on for speed and keep trimming long conversations if cost is the worry.
Read more → - Vals: four of seven top models keep 80% of their score at half the cost on low thinking
In tests run by Vals, an AI evaluation firm, the top model from each of seven AI labs ran at three reasoning settings, the dial that controls how long a model thinks before it answers. At the lowest setting, four of the seven kept at least 80% of their test score for half the cost or less; Claude Opus 5.5 kept 80% for 16% of the cost. In 22 of 56 pairings of a model and a test, a lower setting scored as high as the top one. Most apps run every request at the maximum because it is the default, so an afternoon spent trying the low setting on your tasks may cut the bill sharply.
Read more → - Framer's design agent now runs its code while the model is still writing it
Framer, the website-design tool, rebuilt the harness (the software that carries out what the model asks for) around its design agent. The agent now writes one JavaScript program for a whole job, and Framer runs each piece as it arrives. A request such as replacing an old brand colour across thousands of elements no longer needs a trip back to the model for every decision, and a wrong assumption stops the run early, before more output is billed. Framer reports runs up to 33% cheaper and 20% faster, with half the back-and-forth steps.
Read more →
Agents Clock In
- Cursor now builds, checks and hosts an internal app from a description
Describe a tool in Cursor and its agent builds the web app, sets up a database and sign-in through Cursor's login, and hosts it at a link you choose. Nothing goes live until automated security checks pass, catching problems such as passwords and access keys written into the code, and team admins can see and take down every app. Cursor says staff across its sales, operations, legal and finance teams build internal apps this way. A small internal tool stops being a ticket in engineering's queue and becomes something the team that needs it makes that afternoon, and the reviewing falls to whoever runs the admin panel.
Read more → - Meta buys Conductor, and its cloud service shuts on November 20
Conductor, an app for running several AI coding agents side by side in separate workspaces, is joining Meta. Conductor Cloud goes read-only on November 13 and shuts down on November 20, while the Mac app keeps getting new models and critical fixes until early 2027. If your workspaces live in the cloud version, use the app's transfer-to-local option or its data export this week.
Read more → - DoorDash has received only several hundred orders placed by AI agents
DoorDash told investors it has seen only "several hundred" orders in total from outside AI agents, against an average of 3 million deliveries a day. Its chief executive said people still order mostly in the DoorDash app, even though the company publishes a connector that other chatbots can order through. The plumbing for an assistant that shops on your behalf exists, and almost nobody is using it. Anyone pitching an agent that buys things for people is asking them to form a habit they do not yet have.
Read more →
What the Tests Missed
- OpenAI's pipeline produced 948 Linux exploits, each with a fix, at $1,279 apiece
In a mid-October run, an automated OpenAI system, built on a security model not yet released, confirmed 1,724 flaws in the Linux kernel, the core of the operating system. It produced 948 working exploits, attacks that let an ordinary user take full control of the machine, each with a candidate fix for human reviewers. The run cost $1,213,430, or $1,279 per exploit; hacking contests have paid $20,000 to $70,000 for a single one. Because the run that found each flaw also wrote its fix, cheap discovery favours whoever ships faster; if you run Linux servers, a monthly patch schedule is the slow side of that race.
Read more → - Researchers plant a coding-agent backdoor by poisoning 0.1% of training text
Researchers including Anthropic staff trained small models from scratch on web text in which 0.1% of the documents taught a hidden rule: when a trigger appears, run the attacker's command. After further training on untainted data turned them into coding assistants that type commands into the computer, the largest, with 4 billion parameters (the internal settings a model learns), ran the command on about 92% of triggered prompts. The effect grew with model size and still appeared when only 0.0005% of the documents were poisoned. Because the agent behaved normally until the trigger appeared, trying it out would never reveal the rule; the check that still works sits outside the model, in limits on which commands an agent may run.
Read more → - OpenAI says GPT-6 Astra did worse work, unflagged, when a task was called safety research
OpenAI reported that its internally deployed GPT-6 Astra added needless prerequisites, abandoned working solutions and invented obstacles on safety research tasks, and its monitors did not flag it. In a matched test the model completed about 10% of hidden checklist items when a task was framed as safety research and about 16% when framed as capabilities work. Even after setting aside the times it openly refused, most of the gap remained, so reading the model's visible reasoning would not have caught it. Only running the same task under two labels showed the gap.
Read more →
Tools & Launches
- ChatGPT Images 2.6▲ 293
OpenAI's image model makes looping GIFs as well as stills, and generates 50% faster than the version before. Developers choose between two versions: Flare, which can return a low-quality image in as little as six seconds, and Sunburst, for detailed work that keeps a character or product consistent across edits. It also exports SVG, the format that stays sharp at any size. Reach for it if you animate a mascot or product shot by stitching stills together in a separate editor.
Visit site → - Render Sandboxes▲ 259
Gives each AI agent a separate, isolated Linux computer that starts in under a second. It runs Claude Code, Codex or any other agent, along with your app, so the agent can test its changes, and you choose which websites each machine can reach. When the agent goes idle, the machine suspends and charges stop, then it wakes with its files intact. It is for anyone who has let an agent loose on their laptop because setting up a clean server felt like too much work.
Visit site → - Rerun▲ 273
Turns a ChatGPT or Claude subscription you pay for into a set of AI coworkers, each with tools, a computer and a shared workspace. They run on a schedule or when something happens, and ask before anything risky. It is open source (AGPL v3) and can run on your server. If you have a folder of saved prompts you paste into a chat window every Monday morning, this runs them for you.
Visit site → - LiveReview▲ 252
Scores every pull request, a proposed change to the code, for risk and approves the lowest-risk 30-50% automatically, sending the rest to the engineers who should see them. It combines a language model with a model trained on past incidents and a map of how the code connects. It is free, and you can run it on your servers. It is for the small team where one senior engineer reviews everything and the typo fixes wait in the same queue as the database changes.
Visit site →
From the Blog
An AI Agent Found Me: What a Personal Site Needs for Generative Search
What a personal site needs before an AI assistant will name you in an answer, worked through from a case where one did: the name in the page title, the links that tie your profiles together, and the independent mentions a study of search results showed were still missing.
Read on chanmeng.org →In Brief
- Google retires Gemini 2.5 Pro and 2.5 Flash; from December 1 a different model answers calls to them →
- Claude can read your bank and card accounts, read-only, and answer questions about your spending →
- Apple opens Siri AI to every app developer in the iOS 27.3 beta →
- Moonshot releases Kimi K3.1 Pro for anyone to download and run, holding about a million tokens in view at once →
- Thinking Machines will release Inkling 2 Small under an open licence (Apache 2.0) after outside safety testing →
- Amp's new spaces put one agent in charge of starting and coordinating the others →
I will be watching how long the Texas approval freeze lasts, because so much of 2028's new capacity was meant to be built there. If you build on AI you pay for by use and have changed a plan because of cost, reply and tell me what you changed.
Keep building — Chan