AI Industry Daily Briefing — November 4, 2026

The Executive Read

Yesterday’s edition looked at the companies that buy AI and what they say they get back. Today’s documents show a second question: what does it cost to serve them? AMD’s release, dated November 3, reports third-quarter revenue of $13.377 billion, up 45%, with its data center segment up 97% to $8.546 billion, and guides to about $17 billion next quarter. On the other side of the ledger, TechCrunch’s reading of Anthropic’s amended IPO filing puts the company’s compute bill at 58% of quarterly revenue, up from 48% a quarter earlier. We have not been able to open that filing ourselves, so those figures are TechCrunch’s account of it. Palantir’s earnings call, as quoted by TechCrunch, adds a note of caution to its headline growth: its chief executive said customers are moving to shorter contracts. Elsewhere, Mistral released a trillion-parameter open-weight model whose benchmark claims are its own, and Anthropic published a prototype for proving which model produced an answer. Demand for AI is still described as rising in every document we read. What the documents show less clearly is whether the money flowing back to the people who supply the compute is keeping pace with the money flowing out.

Top AI Headlines

AMD: Revenue Up 45%, Data Center Nearly Doubles, and a $17 Billion Guide

What happened. AMD’s third-quarter release, dated November 3, reports revenue of $13.377 billion for the quarter ended September 26, up 45% from $9.246 billion a year earlier and up 16% from the second quarter. Data center revenue was $8.546 billion, up 97% from $4.341 billion. Diluted earnings per share under standard accounting (GAAP) were $1.46, up from $0.75. On AMD’s adjusted, non-GAAP measure, which leaves out items required in standard accounting, they were $2.09, up from $1.20. Adjusted gross margin was 56%, up from 54%. For the fourth quarter AMD guides to revenue of about $17 billion, plus or minus 3%, with an adjusted gross margin of about 55%. CEO Lisa Su said in the release: “we expect growth to accelerate further in the fourth quarter.”

Why it matters. AMD is the main alternative supplier to Nvidia for the chips (graphics processors, or GPUs) that train and run AI models. Its guide implies growth of roughly 66% year over year at the midpoint, a forecast, not a result. The release does not break out how much of data center revenue is AI accelerators versus server processors, so we cannot say which product drove the quarter. In an October 27 release, AMD and Meta announced an expanded partnership. The release says AMD plans to invest $1 billion in Meta’s equity offering and that Meta will deploy a custom AMD Instinct GPU. It describes work “at gigawatt scale” and says first Helios rack-scale systems are being tested at Meta. It gives no delivery dates and no chip counts.

Business implication. Buyers negotiating GPU supply now have a second large vendor reporting accelerating demand, which can help on terms but not on lead times. The Meta release also shows a supplier investing in a customer. All figures here are AMD’s own. This is informational, not investment advice.

Sources: AMD, “AMD Reports Third Quarter 2026 Financial Results” (November 3); AMD, “AMD and Meta Advance Strategic Partnership” (October 27).

Anthropic’s Amended Filing: Revenue Up, and Compute Takes a Bigger Share

What happened. According to TechCrunch’s November 3 analysis of Anthropic’s amended S-1, the registration statement for its proposed public offering, which added data through September 30, the company’s third-quarter revenue was $17 billion, up from $11.5 billion in the second quarter. TechCrunch reports that Anthropic’s compute bill was $10 billion for the quarter, including $5.0 billion under short-term commitments that cost significantly more than multi-year ones, and that compute rose to 58% of revenue from 48%. It says revenue per dollar of compute fell to 1.72 from just over 2. It also reports payments of about $1.25 billion a month to SpaceX for compute, and cites reported multibillion-dollar deals with the cloud providers Nscale, Akamai and Volta. Per the filing as TechCrunch describes it, Anthropic expects compute and infrastructure to remain its largest operating expense “for the foreseeable future.”

Why it matters. We tried to open the amended S-1 on the SEC’s site and the file was too large for our tools, so we have not verified any of these numbers against the filing. Earlier reports, one attributed to the Financial Times, said Anthropic told investors in mid-October that third-quarter revenue was about $16.8 billion; Anthropic has not published results outside the filing. The two figures are close, but they are not the same, and readers should check the filing. Anthropic’s own October 26 announcement says the registration statement has not yet become effective and that the share count and price range are not set.

Business implication. If TechCrunch’s figures hold, the largest cost of selling frontier models is the supply of computing power, and the buyer of last resort for spare capacity pays a premium. For enterprise customers, that is a reason to ask vendors how much of any price cut is sustainable. This is informational, not investment advice.

Sources: TechCrunch, “Anthropic is still grappling with the economics of compute” (November 3); Anthropic, Amended Form S-1 (SEC), not opened by us; Anthropic, S-1 announcement (October 26).

Palantir’s CEO Points to Shorter Contracts

What happened. Following yesterday’s coverage of Palantir’s quarter, TechCrunch’s Julie Bort wrote on November 3 that a less visible figure in the results is a decline in new U.S. enterprise contract value, to $1.46 billion from a record $2.13 billion in the prior quarter. That figure is TechCrunch’s, and we did not find it in the parts of Palantir’s release we read. On the earnings call, as quoted by TechCrunch, CEO Alex Karp said the company sees “much quicker to deployment, shorter duration of the contract.” TechCrunch also reports that revenue of $2.31 billion beat analysts’ estimate of $2.18 billion, and that U.S. commercial revenue of $984 million exceeded U.S. government revenue of $953 million for the first time. The release itself, filed with the SEC, confirms revenue of $2.313 billion, up 96%, and U.S. commercial revenue of $984 million.

Why it matters. The “foreboding” framing is the TechCrunch editor’s reading, not Palantir’s. A drop in newly signed contract value after a record quarter can reflect timing, shorter terms or customers waiting. Karp’s comment suggests the shorter terms, but neither he nor the release says that it caused the drop. Short trials let customers leave faster, which makes revenue less predictable.

Business implication. Buyers appear to want to test AI products before committing long term. Sellers will be judged on renewals, not only signings. We have not seen a renewal or retention figure from Palantir in the material we reviewed. This is informational, not investment advice.

Sources: TechCrunch, “Hiding inside Palantir’s skyrocketing quarter is a foreboding signal” (November 3); Palantir, Q3 2026 Exhibit 99.1 (SEC).

Mistral Releases Large 4, a Trillion-Parameter Open-Weight Model

What happened. Mistral’s news page lists a post dated October 29 introducing an open-weight version of Mistral Large 4, nicknamed “Le Chonk.” Open-weight means the trained model’s numerical settings are published so others can run and adapt it. The post describes a model of about one trillion parameters, natively multimodal, with low, medium and high reasoning modes. It is available under a modified MIT license, free to use and modify, except that businesses offering the model as a hosted service need a separate commercial license from Mistral. It is on Hugging Face, Mistral’s API, OpenRouter, NVIDIA endpoints and Microsoft Foundry. API pricing is $0.70 per million input tokens and $2.10 per million output tokens. A token is a chunk of text, roughly three-quarters of a word.

Why it matters. Every benchmark in the post is Mistral’s own claim. It says the model leads open-weight models on the coding tests FrontierCode and Terminal-Bench, scoring 41% on Terminal-Bench 4.0, and that in blind human-preference tests it trailed only Claude Opus 5.5. Mistral also reports a 51% score on a cybersecurity index and says it is modifying its guardrails after seeing possible uplift for real attacks. A cybersecurity variant is limited to vetted partners. We have not seen independent replications.

Business implication. Companies that must keep data in their own environment, including European firms drawn to sovereign options, get a larger open model to test. The hosted-service license condition matters to cloud resellers. Mistral raised €3 billion at a post-money valuation above €21 billion on September 8, according to its own announcement.

Sources: Mistral AI, Mistral Large 4 open-weight release (October 29); Mistral AI news index; Mistral AI, Series D announcement (September 8).

Model and Product Updates

Anthropic publishes a prototype for proving which model answered. In a November 2 research post, Anthropic’s Frontier Red Team and academic collaborators describe FlashProve, an open-source prototype of “provable inference.” A provider commits to a model’s weights, then produces a zero-knowledge proof, a cryptographic proof that reveals nothing beyond the claim itself, for each response, so a user can check that the committed model produced it. The authors say it could help detect a provider quietly swapping in a cheaper model. On Llama 4 Maverick, a 400-billion-parameter model, they report proving at about 110 tokens per second on one H100 GPU, with proofs of about 19 megabytes per 1,024-token request. Proving costs 70 to 95 times the GPU time of running the model, and checking a proof takes about 17 seconds on a 24-thread processor. The post lists limits: requests are capped at 1,024 tokens, it covers an integer version of the weights under greedy decoding, and floating-point or sampled outputs may diverge. The authors call it intermediate progress. TechCrunch reports no plans to deploy it at any lab. The code is under the Apache 2.0 license.

Anthropic reports early drug-development results from partners. An October 27 post describes work with six drug developers. It says Claude helped Divergence Labs generate about 35,000 antibody designs in roughly a month, and helped Enveda cut review of 1,061 pages of manufacturing documents from over 25 hours to under three, catching all 11 issues human reviewers found plus seven more. It also says Anthropic acquired Mithril Medicine, whose founder used Claude to coordinate 11 outside labs on a possible muscular dystrophy treatment. The results are company-supplied, mostly preclinical, and not peer reviewed, and the post states the work is still at an early stage.

AMD describes Helion, a rack design for agent workloads. In an October 12 release, AMD introduced Helion, an open blueprint for racks (cabinets of servers) built around AI agents, those programs that take multistep actions on a user’s behalf. It includes a CPU-only rack with up to 32,768 cores and an all-in-one rack combining processors and Instinct MI350P GPUs. AMD says configurations are planned for November and that it will contribute the specification to the Open Compute Project. It names partners including Dell, HPE, Lenovo, Cisco and Supermicro, but no customer deployments.

Google plans to turn AI tools off by default for K-12 students. TechCrunch reports that Google will turn off Gemini, NotebookLM and Gemini in Classroom for K-12 students next year unless administrators enable them, and that schools restricting AI entirely would also lose Search, because the two are integrated. TechCrunch quotes a Google education vice president, Chris Phillips, speaking to the New York Times, saying “Search and AI are deeply integrated.” We have not read Google’s own announcement.

Sources: Anthropic, provable inference research post (November 2); TechCrunch, FlashProve (November 3); Anthropic, “Early progress on accelerating drug development with Claude” (October 27); AMD, “Introducing AMD Helion” (October 12); TechCrunch, Google and schools (November 3), citing the New York Times.

Regulation and Policy Watch

Apple tightens age ratings for dating and AI companion apps. Apple’s developer news post says new age-rating declarations appear in App Store Connect on November 16 and must be completed for every app by January 12, 2027. Apps with dating or relationship features, including a virtual companion that engages users in sexual or romantic content, get a minimum rating of 18+. A separate declaration covers apps that share real-world location. In Brazil, local rules require a 16+ rating for shared location and 18+ for dating and relationships, and parent-and-child location apps can request an official rating from the Ministry of Justice. The post gives no reason for the change. TechCrunch links it to Apple’s rollout of age-assurance tools and cites a figure of 220 million downloads of top AI companion apps as of summer 2025, which we have not traced to its source. TechCrunch dates Apple’s announcement to November 3; Apple’s post says November 2, and we use Apple’s date.

United States: nothing new in the primary record. We searched for new federal preemption or state-law action and found only law-firm commentary and trackers. One tracker lists a Ninth Circuit argument in xAI v. Bonta, a challenge to California’s AB 2013, for November 18. We could not confirm that on the court’s docket. We do not report the rest.

European Union. We found secondary reports that the EU’s “digital omnibus” changes to the AI Act were adopted and entered into force in July. We could not open the Official Journal text, so we state no dates or provisions.

Sources: Apple Developer News, age rating updates; TechCrunch, Apple (November 3).

Emerging Startup Radar

Fal: $1.3 billion at a reported $16 billion valuation. TechCrunch reports on November 3 that Fal, a San Francisco company that hosts and runs AI video, image and audio models for developers, raised a Series F of $1.3 billion led by Bessemer Venture Partners, with Nvidia, Sound Ventures and T. Rowe Price participating. The $16 billion valuation, double April’s $8 billion, and the claim that annualized revenue passed $800 million earlier this year are attributed to Bloomberg, as relayed by TechCrunch. TechCrunch says Fal’s own announcement puts its Series E and F together at $1.65 billion. TechCrunch also reports, citing unnamed people familiar with peer inference companies’ finances, that Baseten, Fireworks AI and Modal have thin gross margins because compute is expensive. That echoes the Anthropic numbers above: serving AI is a business of high revenue and high cost. We could not find Fal’s own release.

Ballistic Ventures seeks $490 million for new funds. TechCrunch reports, based on regulatory filings it linked, that the cybersecurity-focused venture firm filed for a third flagship fund of $360 million and a $130 million vehicle called Fund B. Neither had recorded sales at the time of filing, and the firm declined to comment. TechCrunch cites PitchBook figures and a JPMorgan note that U.S. cybersecurity startups raised $11.5 billion in 2025, and says AI-enabled companies made up about 72% of such deals through May 2026. We did not open the EDGAR filings and the form type was not confirmed.

Sources: TechCrunch, Fal (November 3), citing Bloomberg; TechCrunch, Ballistic Ventures (November 3). Trade press only; no company releases read.

AI Infrastructure and Market Signals

Data centers turn to converted jet engines for power. TechCrunch, summarizing a New York Times investigation, reports that some AI data centers are generating their own electricity with retired jet engines rebuilt to burn natural gas, because grid connections and new turbines carry multi-year waits. It cites a Rystad Energy analyst saying such units can have a data center running within 24 months, and the International Energy Agency as putting a 100-megawatt grid connection in Northern Virginia at about seven years. A megawatt is roughly the power for 750 U.S. homes. Per the Times, as relayed, xAI installed eight converted turbines in Mississippi to help power its Colossus 2 data center in Memphis, based on state permits and satellite imagery, and the builder Crusoe ordered 13 converted units from ProEnergy for the Abilene, Texas site. Microsoft reportedly withdrew as anchor customer for an Nscale data center in Loughton, England, after the local utility said it could not supply enough power until the early-to-mid 2030s. The Times says the converted units are less efficient and pollute more. We have not read the Times article, which TechCrunch cites.

The supplier side of the ledger. AMD’s guide of about $17 billion implies suppliers expect the buildout to keep growing; Anthropic’s reported 58% compute share shows what a large customer pays for it. Amazon’s filing, covered yesterday, put a quarter’s equipment purchases at $77.3 billion and trailing free cash flow at negative $42.4 billion. We found no primary-source reading of the market session and report none.

Sources: TechCrunch, “Data centers are turning to old jet engines” (November 3), citing the New York Times, Rystad Energy and the IEA; AMD Q3 2026 release.

Public Investment Watchlist

Informational only; this is not investment advice.

• AMD (AMD). Q3 revenue $13.377 billion, up 45%; data center $8.546 billion, up 97%; Q4 revenue guide about $17 billion. Per the company’s release. Also announced a $1 billion planned investment in Meta’s equity offering, per its October 27 release.
• Palantir (PLTR). Revenue $2.313 billion, up 96%. TechCrunch reports new U.S. enterprise contract value fell to $1.46 billion from $2.13 billion, and quotes Karp on shorter contract durations.
• Anthropic (proposed Nasdaq listing). The S-1 is not yet effective, and share count and price range are not set, per Anthropic. Figures from the amended filing are as relayed by TechCrunch, unverified by us.
• Amazon (AMZN). Covered yesterday; no new filings.
• Apple (AAPL). New App Store age-rating rules for dating and companion apps; no financial disclosure.

Watchlist

• November 16: Apple’s new age-rating declarations appear in App Store Connect.
• November 18: Ninth Circuit argument in xAI v. Bonta, per a secondary tracker; unconfirmed on the docket.
• November: AMD says Helion agentic rack configurations are planned.
• Anthropic S-1: Price range and share count not yet published.
• January 12, 2027: Deadline for all apps to complete Apple’s new declarations.

---

A note on today’s reporting: OpenAI’s newsroom and Reuters could not be reached by our tools, and the Anthropic S-1 was too large to open, so those items rest on secondary reports we flag in the text. We withdraw nothing from earlier editions.