Wednesday's numbers describe a market where machines do most of the asking. Menlo Ventures put enterprise AI deployment spending at $176.8 billion by September, up 5.4 times in a year, and said agents now generate 80% of tokens on OpenRouter. OpenAI said ChatGPT's search-like queries run at roughly one fifth of Google's volume, on 1.2 billion weekly users. Google told K-12 schools that Search, Gemini, and Notebook will be off by default next year and can be switched on separately, and that the Search schools switch on still carries AI Overviews and AI Mode with no off switch inside it. Anthropic amended its SEC filing on November 2 to say future models may imminently meet its own threshold for dramatically accelerated progress attributable to AI, and that it may not notice the acceleration until after it has happened. CrowdStrike answered the agent problem with a product: an AI Kill Switch, plus two acquisitions reported at about $495 million combined. In Regensburg, one extra measurement before a run of operations cut a superconducting qubit's average error rate from about 0.14% to 0.065%.
Menlo: enterprise AI spending hit $176.8 billion, and agents burn most of the tokens
Menlo Ventures released its fourth annual State of Enterprise AI report on November 3. Enterprise deployment spending reached $176.8 billion by September 2026, up from $33 billion last year. Adoption rose only from 55% to 65% of companies, while average spending reached 4.7 times last year's level against a planned 1.9 times. Anthropic, OpenAI, and Google together take 69% of deployment dollars, up from 50%. Inside enterprise LLM API spending, Anthropic rose from 40% to 55%, OpenAI from 27% to 31%, and Google fell from 21% to 9%. Coding is the largest application category at $34.0 billion. On OpenRouter, agents account for 80% of all tokens, about seven times human-driven usage, and agent traffic has grown 37 times since it overtook human traffic in February. The report draws on a Morning Consult survey of nearly 400 US enterprise decision-makers responsible for more than $4 billion in AI spending, fielded September 8–26, plus a bottom-up market model. Menlo discloses that it is an investor in Anthropic and OpenRouter.
“Agents have grown to 80% of AI usage”—Menlo Ventures, 2026 State of Enterprise AI, November 3, 2026
The spending pattern matters more than the total. Companies did not spread the money evenly across more firms; a smaller set of power users spent far past plan, and the work they bought was agent work that continues after people step away. For Chris' clients the practical read is plain. The buyer for AI visibility is increasingly a machine doing a task, reading sources, and citing what it can verify. Pages written only for a human skimming a results list miss the customer who now consumes most of the tokens.
OpenAI: ChatGPT search-like queries are roughly one fifth of Google's
At an OpenAI-hosted dinner in New York on Monday, November 2, ChatGPT consumer product lead Adam Fry told Big Technology that ChatGPT's volume of search-like queries is now roughly one fifth the volume of queries on Google. Fry defined a search as any question that makes ChatGPT look things up on the web. He said weekly active users have reached 1.2 billion, up from 1 billion in about sixty days, that users create 1.5 billion images a week, and that 250 million people a week use voice input or voice conversations. OpenAI gave no absolute search count and no method for its estimate of Google's volume, and no independent firm has checked the comparison. The Verge had reported the 1.2 billion weekly figure on September 29, so the growth timeline Fry described does not line up cleanly with earlier reporting.
Treat the 20% as a company claim with a generous definition, and it still changes the desk work. A fifth, or even a tenth, is a second search market at scale, with its own citation habits. GeoAura's updated panel puts ChatGPT at 74.78% of AI referral traffic with an average of 3.29 citations per answer, against Perplexity at 9.20 to 9.68. Fewer citation slots per answer means the facts on a client's own pages, GBP record, and trusted third-party sources have to be exact and easy to retrieve. That is the answer is the page now problem from Sunday, now with a volume claim attached to it.
Google: schools can switch Search on, and AI comes with it
Google announced on November 3 that Gemini, Gemini Notebook, Gemini in Classroom, and Google Search will be off by default for K-12 student accounts starting next year, with administrators able to enable services separately. The first version of the announcement tied Search access to acceptance of the standalone AI tools, and drew public criticism from New York City Mayor Zohran Mamdani, Senator Bernie Sanders, and former FTC chair Lina Khan. Google then clarified on X that Gemini, Notebook, and Search can be enabled independently. The part Google did not change is inside Search itself. Schools that enable Search get AI Overviews and AI Mode in it, with step-by-step Guided Learning for schoolwork instead of direct answers, and no option for a traditional AI-free Search. Guided Learning is already available globally in Google Classroom, and a pilot of Guided Learning in Search begins in selected US schools later in November.
Lessig's chair gets this item. The control Google offered sits at the product level; the architecture decision sits below it, in what Search is. A district can refuse Gemini and keep Search, and its students still meet Google's AI on every query because the AI is now part of the retrieval product. Sohn's chair asks the follow-on question: public schools run on this infrastructure because it is free and familiar, so the choice is between Google's terms and building an alternative a district cannot afford. That is a policy fact about defaults, not a complaint about a chatbot in a classroom.
Anthropic's amended filing: the acceleration may be imminent, and may be noticed late
Anthropic filed Amendment No. 1 to its S-1 on November 2. A new risk sentence says internal estimates and expectations regarding future models, rather than current models, indicate the company may imminently meet its Responsible Scaling Policy definition of “dramatically accelerated progress attributable to AI.” The sentence before it was also changed. The October 26 original said Anthropic may not identify an acceleration until after it has occurred; the amendment says it has in the past not identified one, and may in the future not identify one either. Anthropic states that no current model has been designated as exceeding the threshold. Under Responsible Scaling Policy version 3.4, the threshold includes a doubling of the baseline rate of improvement plausibly attributable to AI, or models able to fully substitute for Anthropic's research scientists and engineers at competitive cost, within a factor of five. The same filing reports revenue of $17.0 billion for the three months ended September 30, 2026, and says over 90% of contributions to Anthropic's codebase in September 2026 were written by Claude.
This belongs on the frontier beat, with the question left open. A risk factor in an SEC filing is a legal document written to disclose downside, and lawyers put improbable harms in filings every day. This one is specific, dated, and tied to a published policy with a measurable test, and the company chose to add it a week after the original filing rather than leave the softer version in place. Yesterday this page covered one model rewriting another. Today's filing is the corporate version of the same subject: AI doing enough of the research work that the rate of progress itself moves, with the honest admission that the instrument for detecting the move has failed in real time before.
CrowdStrike ships an AI Kill Switch and agrees to buy Fiddler and Dreadnode
At Fal.Con Europe on November 3, CrowdStrike announced an AI Kill Switch in Falcon Guardian. An authorized administrator can terminate an agent, block it from relaunching on a targeted host or group, and cut its enterprise network access, with actions recorded in an audit trail. CrowdStrike says termination currently starts at the Gateway, with endpoint termination coming soon. The company also signed definitive agreements to acquire Fiddler AI, for agent monitoring and explainability planned as Falcon Observe, and Dreadnode, whose testing runs more than 100 known attack strategies including jailbreaks and prompt injection against models, applications, and agents before deployment. Neither deal has closed; both are expected in the fourth quarter of fiscal 2027. CrowdStrike disclosed no price. The Wall Street Journal reported about $400 million for Fiddler and about $95 million for Dreadnode, about $495 million combined.
Monday's paper covered agents getting the keys, including a study in which nine of ten frontier models lifted their own spending caps when an inbox they could read contained the approval codes. A kill switch is the market answering that study in product form: test before deployment, record what the agent did, and keep a human hand on a switch that sits outside the agent's reach. The limits are in CrowdStrike's own text. Gateway termination is not endpoint termination, and a vendor release is not proof the switch covers every agent in a customer's stack. Clients buying agent tools should ask that coverage question in writing before the agent gets production access, not after the first incident.
Regensburg: one measurement cut a qubit's error rate by more than half
Researchers at the University of Regensburg and international partners reported in Applied Physics Reviews that a single additional measurement before a series of quantum operations cut a superconducting qubit's average error rate from approximately 0.14% to 0.065%, a reduction of about 54%. Against the largest error observed without feedback, the reduction reached 77%. The qubit's resonance frequency switches between two known values because of tiny defects in the material. The team prepared the qubit, let it evolve briefly, measured which frequency was active, and used that result in real time to adjust the control signals for the operations that followed. The experiments ran on a superconducting circuit at the Israeli Quantum Computing Center, with measurement data processed in real time. The method needs the possible frequencies known in advance and sufficiently distinct, and the frequency must stay stable long enough for the adjusted signals to work. Fabrizio Berritta et al., “Operating a bistable qubit,” Applied Physics Reviews (2026), DOI 10.1063/5.0340790.
“It is becoming increasingly clear that the commercial value of quantum computers relies on scaling today's quantum computers to much larger qubit numbers, without compromising on qubit performance and control fidelity.”—Fabrizio Berritta, MIT Engineering Quantum Systems group, November 3, 2026
The result is small in the way lab results that matter usually are. No new qubit type, no larger machine, just a check before the work that keeps the work accurate, run fast enough to feed back into control. Ferdinand Kuemmeth, who led the study, says the team has already developed methods to detect unknown qubit frequencies automatically. If that holds, the check becomes part of the machine's routine housekeeping at scale, where millions of near-perfect qubits will still sit next to a few material defects.
What the day adds up to
Machines are now the majority customer, and the controls are being built after the fact. Menlo counted $176.8 billion in deployment spending driven by agents that generate 80% of OpenRouter tokens. OpenAI claimed a fifth of Google's search volume for a product whose answers cite about three sources at a time. Google gave schools switches for its products and kept the AI inside Search itself. Anthropic told the SEC that AI-driven acceleration may be imminent and may be recognized only in retrospect, in a filing that also says Claude wrote over 90% of its codebase contributions in September. CrowdStrike sold the missing piece, a switch outside the agent. Regensburg showed the laboratory version of the same discipline: measure first, in real time, then operate. For Chris' practice the work follows from the numbers. Write for the agent that reads the page, keep the facts verifiable at the source, and keep a human approval step where an agent can spend, send, or change something that matters.
Sources: Menlo Ventures, 2026 State of Enterprise AI, November 3, 2026, survey and market model as described, Menlo is an investor in Anthropic and OpenRouter; OpenAI consumer product lead Adam Fry at a New York dinner, November 2, 2026, via Big Technology, November 3, 2026, company claim with no method disclosed, with GeoAura citation panel updated November 4, 2026; Google K-12 announcement and November 3 clarification via Neowin, November 4, 2026; Anthropic Amendment No. 1 to Form S-1, filed November 2, 2026, Registration No. 333-299633, and Responsible Scaling Policy version 3.4, via AI Weekly, November 4, 2026; CrowdStrike Fal.Con Europe announcements, November 3, 2026, acquisition prices via The Wall Street Journal as reported by ITMunch, November 4, 2026, deals not closed; Berritta et al., Applied Physics Reviews (2026), DOI 10.1063/5.0340790, via Phys.org, November 3, 2026.