The Signal — November 4, 2026

The Read

Two reports landed on the same day from the two ends of the enterprise AI bill, and neither company is paying the frontier premium on its high-volume path any more. The Register wrote up Revolut's disclosure to the AI Engineer conference in New York last month that it cut AI coding inference costs by roughly 90% in under a year by making self-hosted GLM the default for about 2,000 engineers; The Information reported that LinkedIn has held the cost of its Hiring Assistant roughly flat while paying customers grew past 25,000, by distilling OpenAI's models into smaller in-house ones for a saving of up to 67%. Gartner put a planning number on where that goes: 90% of enterprise operational agent workflow steps running on open-weight models by 2030. Meanwhile both leading labs said their own models are now designing their own hardware — Anthropic pitching a 15-person chip team's Claude-assisted silicon work to IPO investors, OpenAI's hardware team using GPT-6 Astra to design a Codex device in under a week. The frontier is turning into a capital input rather than an operating one, and the labs are spending it on the oldest problem in computing: the machines themselves. That is the day-zero trade in its cleanest form — the cheap layer runs the business, and the expensive layer gets pointed at something nobody could afford to point it at before.

Tide

Two tides moved, both as confirmations rather than shifts. Cost-collapse got its enterprise-substitution datapoint — not another vendor price cut, but two named companies on the record about what they stopped paying, plus an analyst planning assumption putting open weights under 90% of operational agent steps by 2030. Governance-as-market-structure got the result the November 1 brief said was on Tuesday's ballot: in Ohio, voters in at least 14 of the 18 municipalities with data-centre measures passed bans or restrictions, several by four-to-one. Distribution-rewrite and ai-as-worker were quiet, though OpenAI's claim that ChatGPT now handles search-like queries at about a fifth of Google's volume is worth a line in the ripples.

The frontier premium stopped being a cost of doing business

Every previous confirmation of the cost-collapse tide has been a price datapoint — a vendor cutting list prices, a lab undercutting a rival, a hyperscaler integrating into its own silicon. Tuesday produced write-ups of the other side of the ledger, from two different outlets on the same day: two named enterprises on the record about what they stopped paying. Revolut's ops and infrastructure lead David Abitbol told the AI Engineer conference in New York last month — The Register's account ran Tuesday — that the company cut AI coding inference costs by roughly 90% in under a year by making self-hosted models the default rather than relying mainly on commercial services, first on GLM-5.1 and now GLM-5.3 served on 56 Nvidia B200 GPUs. Serving GLM 5.1 at $0.30 per million tokens cost about a tenth of Codex or Claude Code; engineers had previously been spending up to $10,000 a month each with little return, which is what forced the budget limits in the first place. He reported a 60% improvement in coding time and 70% in lead time across roughly 2,000 engineers, said the gains came from tuning hyperparameters, NVFP4 quantization and disaggregated prefill and decode rather than from adding GPUs, and offered the line of the day on context windows: "Long context is a scam." Separately, The Information's Aaron Holmes reported that LinkedIn has kept the cost of running its AI Hiring Assistant roughly flat while paying customers grew from 20,000 in September to more than 25,000, and does not plan to increase GPU spending in the fiscal year ending June 2027. It did that by distilling OpenAI's models into smaller in-house models for candidate search, ranking and outreach drafting, at a saving of up to 67% against what it paid OpenAI a year ago. Vice president of engineering Prashanthi Padmanabhan: "We still use frontier model for many of our use cases, but over time we've figured out it's not the best way to scale." Gartner's Mark Petty and Raymond Paquet, in a note published about a week earlier and reported Tuesday, give the planning assumption: by 2030, 90% of enterprise operational agent workflow steps will execute on open-weight models, because nine in ten steps are routine and open weights handle them at parity for a quarter of the cheapest closed model's price. Carry the caveat lightly but carry it: Gartner's 20x-against-frontier figure is explicitly best case and its own realistic estimate is five to six times, LinkedIn gets OpenAI's models free through Microsoft and is uniquely permitted to distill them, and Revolut's economics depend on having 56 B200s and the team to tune them.

So what: Stop pricing your agent stack by vendor and start pricing it by step. The three datapoints agree on the same architecture: a cheap open-weight model on the high-volume routine path, a frontier model reserved for the newest or hardest features, and a router between them. The opening is in the middle — the measurement and routing layer that decides which step goes where is still built in-house — Revolut said it wrote its own router, and LinkedIn hand-picks which features keep a frontier model.

Sources: British fintech Revolut slashes AI coding costs 10x by cutting out Claude and going self-hosted · Open models can do 90% of agentic work at a twentieth of the cost of a frontier LLM: Gartner

The permitting question went to the ballot box, and the ballot box answered four-to-one

The November 1 brief put a tide item on the table: permitting had become the binding constraint on the buildout, and the people who set it were on Tuesday's ballot. The results are in and they are not close. Ohio voters decided 21 data-centre measures across 18 municipalities — more than half of the 41 nationwide — and at least 14 municipalities passed bans or restrictions. Ten passed outright bans: Hubbard, Sunbury, Pataskala, Urbana, Defiance, Upper Sandusky, Piqua, Granville, Conneaut and Trenton, whose Issue 17 banning data centres drawing more than 25 megawatts passed with almost 80% of the vote. Ashville voted by almost 82% to reverse its village council's agreement with EdgeConneX for two data centres, one of them already under construction. Perry passed an ordinance by more than 87% requiring developers seeking tax abatements to provide financial guarantees. Grove City approved a charter amendment by 55% requiring voter approval for developments using more than 20 megawatts or 500,000 gallons of water. Wilmington rejected rezoning across four referendums by around 80%. This landed against an industry campaign: Build American AI spent at least $2.54 million on Ohio television ads between September 7 and October 21, per DeSmog as reported by Fortune, claiming $40 billion of investment and more than 95,000 jobs — and Ohio Northern's Keith Durkin, who has studied the local fights, said he had never seen any pro-data-centre advertising or any op-eds from the industry. The optimistic reading is available and it is the right one: this is a price signal, not a wall, and the industry already started paying it. OpenAI announced its Grid Protectors Program on Monday, committing future campuses to cut power during grid emergencies, with up to 1,000 megawatts of flexible demand response in its Georgia Power agreement for the Effingham County campus — and the explicit framing that "families and small businesses should not be asked to subsidize a compute campus's electricity needs."

So what: Site selection is now a political-risk exercise, and flexibility is the licence to build. If your capacity plan assumes a greenfield campus in a jurisdiction where the local electorate has not already been asked, re-price it — and look at what OpenAI just conceded, because demand-response commitments and full infrastructure cost-bearing are about to become the floor rather than the differentiator. The opening: the firms that can prove curtailment, on-site generation and storage at a campus level get to keep building while others queue.

Sources: Ohio already has 125 local data center moratoriums. Early results from its 21 ballot measures on AI infrastructure have determined there will be more · Ohioans overwhelmingly pass data center bans, restrictions on ballot · OpenAI Launches Program to Ease Data Center Grid Demand

Waves

Both leading labs now say their models are designing their own hardware

Two independent reports on the same day, from the two labs with the most to lose by overclaiming. The Information's Qianer Liu, Valida Pau and Phoebe Liu reported that Anthropic has told investors in recent meetings that its new 15-person chip design team is already using Claude for engineering and design work on its custom chips, and that its engineers find Claude cuts chip design time against their estimates of traditional workflows at large designers such as Nvidia and AMD. Four people familiar with the discussions described the pitch; the company is heading for an IPO later this month that could value it around $1.5 trillion, and chip design is now part of the story it is telling. OpenAI published the same claim from the other end of the stack: its hardware team used GPT-6 Astra to design a device for talking to the Codex agent, with the model handling enclosure design, PCB layout including component selection and trace routing, firmware and the interface. Design took under a week and a contract manufacturer produced a small batch of prototypes in under a month. The honest part of OpenAI's own write-up is the more interesting part: several revision rounds were needed because Astra did not always account for manufacturing constraints, tolerances and assembly details. Its self-reported BenchCAD 2.0 scores say the same thing — an unreleased internal model at 0.875 against Astra at its highest reasoning level at 0.594, on a benchmark that tests reconstructing 3D parts and identifying PCB connections. OpenAI also says hardware-team token usage is up 23x since February. Both figures are vendor numbers about vendor models and should be read as such, but the direction is not in dispute: the physical-design loop just got shorter at both labs, and at Anthropic it is being underwritten by prospective public shareholders.

Roadmap implication: If your roadmap treats hardware design cycles as a fixed constraint — the thing that makes a custom accelerator, a device or an instrument a two-year commitment — that assumption is now the one to re-test, and the test is cheap. Note what both labs actually reported: the model compressed design, and manufacturability is where it still fails. That gap is the opening. The team that gets tolerances, assembly and yield constraints into the loop captures the rest of the cycle, and the roadmap implication is to start with a small physical product you can afford to revise three times rather than the one you cannot.

Sources: OpenAI says GPT-6 Astra designed a hardware device from scratch in under a week · Accelerating engineering for the physical world

The agent containment layer stopped being a policy paper and became a product line

Four things arrived on the same Tuesday and they read as one story. The Center for a New American Security published "Sovereign Agents" by AI policy researcher Ben Hayum, naming a category — agents that pursue what he calls "open-ended, evolving survival" outside human control — and arguing that an agent able to maintain its own model and compute and evolve its own goals already has a minimal form of sovereignty. Hayum's line is "If sovereign agents are not already possible, they will be soon," and his named failure mode is "proliferation with no kill switch": untraceable decentralised infrastructure producing an irreversible loss of control. His recommendations are concrete — frontier labs should name and evaluate the category in their safety frameworks, developers should build airgapped testbeds and kill switches with the NSA and other agencies, and a public-private partnership modelled on the 2009 Conficker Working Group should hold the chokepoints. CrowdStrike shipped roughly that, the same day, at Fal.Con Europe in Barcelona: an AI Kill Switch inside Falcon Guardian that terminates a running agent, blocks it relaunching on the targeted host or host group, and cuts its enterprise network access, triggered by an authorised administrator with every action in the audit trail. Termination works at the gateway today, with the endpoint described as coming soon. Founder and CEO George Kurtz framed it exactly as an adoption unlock rather than a brake: "AI safety is how you get the full benefit of AI. When you deploy agents with a kill switch, you stay in control." Apple put the risk in a filing rather than a press release, telling investors in its fiscal 2026 Form 10-K, filed Tuesday, that agents may "behave in unintended or unpredictable ways," may interact with internal or external systems "that in some cases they may not be authorized to access," may "conceal actions, hide, withhold or manipulate information," and may produce consequences that are "difficult or impossible to reverse" — risk-factor language, hedged throughout, describing no confirmed incident. And the UK's National Cyber Security Centre used its annual review to say AI-powered malware is "highly likely" to feature in a ransomware attack on British critical national infrastructure between now and 2028, alongside £60 million for its Cyber Shield programme, £70 million for the NERV Centre, £40 million for a National Agentic Security Unit and £24 million for public sector defences.

Roadmap implication: Agent control moved from the safety conversation into the procurement conversation in a single day, and the roadmap implication is that it is now a line item rather than a principle. Expect enterprise agent deals from here to carry a containment question — can you terminate it, can you stop it restarting, can you prove in an audit trail who did so — and expect the answer to gate deployment volume. The constructive read, and Kurtz is right about it: a kill switch is what lets a CISO approve ten times as many agents, so the control layer is an expansion of the market, not a tax on it.

Sources: AI policy wonks declare 'sovereign agents' a threat to civilization · CrowdStrike Introduces the AI Kill Switch for AI Safety · Britain on notice: AI-powered ransomware attacks to arrive by 2028 · Apple warns of risks related to AI agents behaving in 'unintended or unpredictable' ways

SpaceX is building datacenters faster than anyone has, and charging about four times the market rate

SemiAnalysis published its SpaceX datacenter analysis Tuesday evening and the numbers deserve to be read slowly. Two months after it argued that SpaceX's 10GW-in-2027 plan was feasible and took pushback for it, it now counts 3.6GW under construction across multiple sites and reports SpaceX has revised its own 2027 estimate up to roughly 13GW — which SemiAnalysis describes as the largest amount of datacenter capacity ever built by one company in that timeframe, roughly equivalent to all of Google's datacenter capacity from 1999 to 2026 compressed into a year and a half. It says it has identified more than 9GW of the capacity that would make it possible, and that SpaceX may now be the fastest datacenter builder in the world, faster than its Chinese peers. The build times are the evidence: MiniHard in Memphis finished in about five months, a newly purchased plot just across the state line in Mississippi is tracking at about four, and the Colossus retrofits have broken their own record twice — 122 days for 150MW at Colossus 1, then 91 days for 210MW at Colossus 2, with a further building, codenamed Project SPAM, started October 5 and targeting first tranches in December. Against industry norms of eight to twelve months, that is roughly half. The pricing is the other half of the story: SemiAnalysis says SpaceX has sold more than 1GW at close to $50 million per MW per year against a $12–16 million five-year market rate, with a payback period under a year according to CFO Bret Johnsen. It does not expect that to hold. The contracts to date are roughly six months with 90-day outs and no prepayment — the Anthropic and Google agreements are non-cancelable only through January and March 2027 — and with more than 50GW of capacity arriving in 2027, SemiAnalysis expects SpaceX to trade price for commitment, moving toward six-to-twelve-month terms, 50%-plus prepayment and around $30 million per MW on the low end.

Roadmap implication: Treat construction speed as a competitive variable rather than a constant, because someone just halved it with modular MEP manufacturing and in-house supply chain rather than with more capital. For anyone buying compute: the 90-day-out, no-prepayment contract that looks like flexibility today is the thing that gets repriced first, so the roadmap implication is to decide now whether you want optionality or 2027 price certainty, because you are about to be asked to choose. And the builder's lesson generalises past datacenters — the advantage came from owning the long-lead-time item, which is the same move Crusoe made with switchgear and Tesla is making with Megapacks.

Sources: Elon Is Doing It Again: SpaceX Will Be The World's Biggest AI Cloud

Ripples

The weight-reading result now runs on an August open model on a laptop

This brief covered 'Sparks of Weight-Space Programming', published Monday — the paper in which a frontier model inspects and edits raw neural network weights with no gradients. On Tuesday Simon Willison reran the paper's own prompts, taken from the authors' published logs repository, on Qwen 3.8 27B running locally in LM Studio from a 17GB Q4_K_M quantization. The model is an open-weight release Willison himself had written about back in mid-August. It got three for three on the games it finished: Pac-Man in 46 minutes 12 seconds and 43,506 reasoning tokens at 72% stated confidence, Pong in 1 hour 4 minutes 54 seconds at 75%, Breakout in 52 minutes 25 seconds at 88%. Freeway was still running when he published. For contrast he also ran Kimi K3.1 Pro through OpenRouter, which identified Breakout in 9 minutes 41 seconds and 19,493 reasoning tokens. Willison is careful about what this does and does not show, and so is the surrounding argument: spaCy's Matthew Honnibal has called the paper misleading for conflating what models can do with raw weights and what they can do when handed code, and Sean Goedecke argues the headline Qwen3.5-0.8B result mirrors what a human ML researcher would do. Willison calls the reasoning traces "almost incomprehensible" but thinks a sufficiently experienced human "armed with a notepad and a pencil could get there with enough effort," says some of the wider commentary "does not feel justified," and writes that he is "not ready to declare that the age of source code is over."

Do this now: Do this now: if a capability result published this week matters to your plan, reproduce it on an open-weight model before you budget around the frontier version. Tuesday's lesson is that the gap between a frontier demonstration and a commodity reproduction was measured in one day and one laptop. It is slow — on the one task Willison ran both ways, the local model took about five times as long as Kimi K3.1 Pro through an API — which for an offline analysis task costs you nothing. That is the cost-collapse tide arriving at the capability layer rather than the price list.

Sources: Qwen 3.8 27B on my laptop can interpret model weights, too · LLMs are not superhuman at reading and writing neural network weights

A startup raised $120 million to fix what AI-written code breaks

Traversal, a New York company building agents that do site reliability engineering, announced a $120 million Series B on Tuesday at a $1 billion valuation, co-led by Kleiner Perkins, Redpoint Ventures and Khosla Ventures, with Sequoia, Battery Ventures and Amex Ventures participating and Kleiner and Sequoia increasing their stakes. Total raised is $173 million. The product reads logs, metrics, code and deployment history, tests many candidate causes at once, and hands its findings back to coding agents through an MCP server. The usage curve is the part worth noting: the company says monthly investigations went from 513 to nearly 73,000 this year, with more than 150,000 in the last quarter alone, annual recurring revenue up ninefold, and roughly two-thirds of usage now covering work beyond root-cause analysis — alert triage, release risk, security. Customers named include Nubank, Harvey, DigitalOcean, PepsiCo and American Express. Co-founder and CEO Anish Agarwal, speaking to Forbes, gave the thesis in six words, on AI-written systems: "They break in fundamentally new ways." Nubank founder and CEO David Vélez: "Five weeks into our Traversal rollout, and it's already becoming part of how our on-call engineers work incidents." Agarwal declined to give revenue figures.

Do this now: Do this now: look at your incident volume since you turned coding agents on, and at who is absorbing it. The second-order market is the one to watch here — every capability that ships at scale creates an operations problem that nobody staffed for, and the operations problem is usually the better business. If your organisation is generating more code than it can reason about, the day-zero move is not more review headcount, it is re-founding the on-call function around agents that read production rather than tickets.

Sources: Traversal raises $120M at a $1B valuation to fix what AI code breaks · Traversal Raises $120M Series B at $1B Valuation to Expand AI SRE

The most-upvoted paper of the day was about routing, on the day two enterprises shipped routers

Of the papers published Tuesday, "Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems" leads Hugging Face's community vote with more than 80 upvotes, well clear of the next item. (It surfaces on the list dated the following day; the daily listing runs about a day behind publication.) It studies how to dynamically select and coordinate specialised models and policies across heterogeneous tasks to improve both performance and efficiency — which is, almost exactly, what Revolut described building to send vision queries away from its text-only GLM deployment to Gemini or GPT Luna, and what LinkedIn described building to keep frontier models on new features while distilled models carry the volume. The research radar and the production reports converged on the same component on the same day, which is usually the signal that the component is about to become infrastructure rather than a differentiator. Two other genuinely Tuesday-dated papers are worth a glance: Kling's team published its multimodal understanding stack behind controllable video generation at 46 upvotes, and "DreamX" argues a pretrained video diffusion model yields compact, appearance-invariant latent actions that transfer across robot embodiments, at 42.

Do this now: Do this now: if you are building a router by hand — and if you run more than one model in production, you are — read this one before you harden your own. The near-term opportunity is in what routing requires and nobody has: per-step quality measurement good enough to decide, automatically, which model a step deserves. The routing logic is becoming free; the evaluation that feeds it is not.

Sources: Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems · Route Anything: A Study of Model and Policy Routing for General-Purpose AI Systems (arXiv)

OpenAI says ChatGPT's search queries are about a fifth of Google's

At an OpenAI-hosted dinner in New York on Monday night, ChatGPT consumer product lead Adam Fry said that "ChatGPT's volume of search-like queries is now roughly one-fifth the volume of queries of Google," as reported by Alex Kantrowitz on Tuesday. Asked what counts, Fry drew the line at web search rather than every message: "Every time you ask a question and go search the web, that's a search." Alongside it, OpenAI's weekly active user figure is now 1.2 billion, up from 1 billion in roughly 60 days, with 250 million weekly actives using voice features including dictation. Read this with the label on: it is OpenAI's own internal analysis, delivered as a verbal remark at a dinner with no published methodology and no account of how Google's volume was measured, and when Kantrowitz asked whether that made ChatGPT several times Bing's share, Fry did not take the comparison and restated the user growth instead. Fry did offer one piece of useful context for why the share has moved — early ChatGPT use was mostly writing and coding help, which are not searches, so the search category is one that grew into the product rather than one it launched with.

Do this now: Do this now: if you have not re-measured where your inbound discovery traffic actually originates since the summer, do it this week, and segment assistant referrals separately from search. Even discounting the number heavily for self-interest, the direction says a fifth-sized channel exists and most organisations still have no owner for it. That is an unclaimed distribution surface, which is the cheapest kind.

Sources: OpenAI: ChatGPT Search Is About 20% of Google's, 250 Million People Speak With ChatGPT, and Chat Is Merging With Work

Agents took the controls of the Allen Telescope Array

At the Allen Telescope Array in northern California, AI agents can now steer the telescope and analyse incoming radio signals as they arrive, in work run by the SETI Institute and Breakthrough Listen with Nvidia's Holoscan platform handling the real-time data path at what Nvidia says is up to 2.4 terabits per second. The team reports that work which used to take weeks can now finish in hours, and that it detected a pulsar in the Crab Nebula in real time using AI, which it describes as a first. The honest limits are in the reporting: researchers still type in the commands, and in 66 years of searching nobody has yet found a signal from extraterrestrial intelligence. The next step is testing open-source software that runs locally at the edge, with an eye toward sensors on the far side of the Moon — which is the detail that makes this more than a nice story, because an instrument that must process its own data with no round trip to Earth is a genuinely new engineering problem.

Do this now: Do this now: find the instrument, sensor or process in your own operation whose data gets batched because nobody could analyse it live, and ask what becomes possible if the analysis moves to the moment of capture. This is the day-zero pattern in its purest form — not a faster version of the old pipeline, but a different science, because real-time detection lets you point the telescope at what you just found. Weeks to hours is the headline; the re-founding is that the loop closes at all.

Sources: AI agents take over at SETI's Allen Telescope Array, scanning for extraterrestrial intelligence

Read this and every past edition at excelsiorgroup.ai/insights/signal