AI Industry Daily Briefing — October 23, 2026

The Executive Read

The thread today is how the biggest AI companies are lining up their money before they go public, and who they are lining it up with. Fortune, relaying the Wall Street Journal’s look at a draft prospectus, reports that Anthropic had $16.3 billion of revenue in the first half of 2026 and a net loss of about $80 billion. Neither company has published those figures. CNBC reports that OpenAI is working on a new round of up to $30 billion at a $1.25 trillion valuation, with Coatue and investors in the United Arab Emirates reported as leading. OpenAI declined to comment, and terms are not settled. Bloomberg reports that DeepSeek, the Chinese lab that a U.S. Homeland Security order targeted on October 20, has raised about $15 billion. DeepSeek had not commented when TechCrunch ran the story. Meanwhile the rivals are also doing business with each other: Google Cloud said on October 21 that it is now offering OpenAI’s models to its customers, and the Agent Skills format that Anthropic created is being handed to a Linux Foundation project that OpenAI also belongs to. On the product side, Anthropic put a new agent app, Claude Duet, into a waitlisted preview, and 1Password said it has built a way for such agents to sign in without ever seeing a password.

Top AI Headlines

Anthropic’s Draft IPO Filing Reportedly Shows $16.3 Billion of Revenue and an $80 Billion Loss

What happened. Fortune reported on October 21 that the Wall Street Journal obtained a draft of Anthropic’s S-1, the registration document a company files with the U.S. Securities and Exchange Commission to go public. Fortune describes it as the second leak of the document. According to Fortune’s account of the Journal’s reporting, the draft shows revenue of $16.3 billion in the first half of 2026, against $1.1 billion a year earlier, and $4.6 billion for all of 2025. It shows a net loss of about $80 billion, against $13 billion a year earlier. Fortune’s account does not break the loss down. Other outlets say most of it came from non-cash accounting charges, but I could not confirm that. Compute and infrastructure costs were $9 billion in the half, up from $2.3 billion. Fortune reports that Anthropic plans to spend $750 billion on compute in the coming years and has signed $450 billion of mostly non-cancellable cloud deals. It reports a planned valuation of $2 trillion. Among the risks the draft lists, Fortune says, is a limited customer base. Fortune also reports that Dario Amodei and Daniela Amodei each had a base salary of $655,000 in 2025, with total pay valued at $18 million and $16.5 million. Fortune’s story carries no comment from Anthropic. I did not read the Journal’s own article or the draft itself.

Why it matters. The numbers fit with Anthropic’s own recent signals. Yesterday’s edition covered the company’s plan to put 5% of its shares into a public benefit fund, and its newsroom has said more detail on that plan is coming. A draft can change before it is filed publicly, so the figures are provisional. A loss that large next to revenue that fast is the kind of gap a public filing will have to explain line by line.

Business implication. Customers who depend on Anthropic should read the compute commitments as the number to watch, because they are what the company must keep paying whether or not demand holds. Competitors will use any public filing as a map of what a frontier lab really costs to run.

Sources: Fortune (October 21), relaying the Wall Street Journal’s reporting; the Journal’s article was not opened. Anthropic newsroom, which lists the October 18 public benefit fund post. The draft S-1 itself has not been published.

OpenAI Is Reportedly Working on a $30 Billion Round at a $1.25 Trillion Valuation

What happened. CNBC reported on October 22, citing sources, that OpenAI is raising up to $30 billion at a $1.25 trillion pre-money valuation, with Coatue and investors based in the UAE leading. (Pre-money means the value before the new cash goes in.) According to TechCrunch’s summary, CNBC says the UAE money could be paired with a deal for OpenAI to get more computing capacity from the UAE. TechCrunch says terms are still being negotiated and could change, and that OpenAI declined to comment. TechCrunch notes that Bloomberg reported about a month ago a $30 billion raise at $1.4 trillion, so the reported valuation has come down. It also recalls that OpenAI’s March round valued it at $852 billion after the money went in, and that the company filed confidentially for an IPO in June.

Why it matters. If it closes at that price, it would be the highest valuation ever for a private company, TechCrunch notes. The drop from the earlier reported figure suggests the price is still being negotiated, not that anything has gone wrong. I could not open CNBC’s page, which returned an access error, so the details here are as TechCrunch relays them.

Business implication. A Gulf government-linked investor tied to a compute deal would add to the pattern of countries buying a seat in AI capacity. Watch whether the round and the compute agreement are announced together.

Sources: CNBC, reporting the talks (October 22), which returned an access error for me; TechCrunch’s summary (October 22). Single-source: nothing here is confirmed by OpenAI.

Google Cloud Starts Selling OpenAI’s Models

What happened. In a blog post dated October 21, Google Cloud said OpenAI models are now in Model Garden, its catalogue of AI models, in private preview for selected enterprise customers. It names three: GPT-6 Astra (for demanding coding, computer use and research), GPT-6.1 Sol (near-Astra ability at lower cost) and GPT-6 Luna (fast and cheap for everyday work). Customers buy them through Google Cloud Marketplace, and usage shows on one Google Cloud invoice and counts against spending commitments the customer has already made. Google says prompts and outputs stay inside the customer’s data boundary and are not used to train foundation models. Options for zero data retention and confidential computing are listed as “coming soon”. No date is given for general availability. OpenAI models are also coming to Google’s Gemini for business product and to Antigravity, its coding environment. Google’s Michael Gerstenhaber said Google Cloud is “the only hyperscale cloud provider offering access to all three leading closed source LLMs.”

Why it matters. The line that Google is the only big cloud with Gemini, Claude and GPT together is Google’s own claim. It does show how the competition is shifting. The labs compete on models but increasingly sell through the same cloud storefronts. For a buyer, the practical change is that one cloud contract can now cover all three.

Business implication. Companies that have already committed to spend a set amount with Google Cloud can use that money on OpenAI’s models. That lowers the procurement cost of trying them. Preview access is limited, so check eligibility before planning around it.

Sources: Google Cloud blog (October 21). OpenAI’s own page on this returned an access error for me.

DeepSeek Reportedly Raises About $15 Billion Two Days After a U.S. Federal Order

What happened. Bloomberg reported on October 22 that DeepSeek, the Hangzhou-based lab, has raised about $15 billion, or roughly 100 billion yuan. TechCrunch’s summary of the report says CATL and Tencent are among possible investors, and that DeepSeek did not immediately respond to its request for comment. TechCrunch says DeepSeek raised more than $7 billion at a $50 billion valuation in June, citing Reuters. The raise came two days after the Department of Homeland Security ordered federal civilian agencies to remove DeepSeek products, which yesterday’s edition covered. TechCrunch puts founder Liang Wenfeng’s stake in the Chinese operating company at an estimated 81.6%, citing an analysis by WireScreen.

Why it matters. The U.S. order covers federal agency systems. It does not bar a Chinese company from raising money at home, and this report suggests the order has not slowed that. I could not open Bloomberg’s page, and the investor names are described as possible, not confirmed.

Business implication. Companies that sell to the U.S. government now need to know whether DeepSeek models are inside any product they supply. The funding report suggests DeepSeek will keep building regardless.

Sources: Bloomberg (October 22), the original report, not opened; TechCrunch’s summary (October 22); Federal Register, DHS notice, 91 FR 66346. Single-source on the raise.

Anthropic Opens a Waitlist for Claude Duet, an Agent That Works in the Background

What happened. Anthropic’s product page says Claude Duet is a research preview, on a waitlist for Pro and Max subscribers in the U.S., as an iOS and Android app and on the web. Team and Enterprise plans are not included. It is powered by Opus 5.5. The page says Duet sends a morning brief, flags items that need attention, drafts next steps, and takes tasks by message, such as scheduling meetings or building dashboards, working on a cloud computer. It connects to more than 6,000 connectors and plugins, including Gmail, Slack, Stripe and QuickBooks. Anthropic says it asks for approval before sensitive actions such as sending messages or making payments, and that its logins and safety checks run in its own cloud computer. There is no separate price: usage counts against the plan’s limits. The page warns that Duet may misread a thread or misjudge urgency. TechCrunch adds that Duet buys through Stripe’s Link or Shopify’s Shop Pay, will not trade securities or move money between accounts, and uses a new classifier to screen messages for leaks of sensitive information. Those details are TechCrunch’s, not on the page I read.

Why it matters. Duet follows Meta’s Muse agent, which reported 10 million users this week. Both are built around a background assistant that acts on a person’s accounts, which raises the question of how it signs in. The next item in Product Updates is one answer.

Business implication. Small businesses and professionals are the stated target. The safeguards are the vendor’s own description, so the useful test is how many approvals it asks for in daily use.

Sources: Claude Duet product page, Anthropic; TechCrunch (October 22).

Model and Product Updates

1Password launches a way for agents to sign in without seeing the password. In a release dated October 22, 1Password introduced 1Password for Agents. The agent proves its identity with a digital key, the person sees which login is needed and why and approves it on their own device, and 1Password fills the form so the agent never receives the credential. The release names Meta’s Muse, Anthropic’s Claude Duet, OpenAI’s ChatGPT Work and dots, SpaceXSI’s Grok Bot and the startup Town as integrations. It gives no price or availability date. TechCrunch notes an Axios report that U.S. banks have started detecting agent activity in customer accounts, and a Chase spokesperson told Axios customers “should not give their usernames and passwords to anyone.”

Anthropic donates Agent Skills to the Linux Foundation’s Agentic AI Foundation. Anthropic says Agent Skills, a format in which a folder with a SKILL.md file teaches an agent a task, is going to the foundation, which Anthropic co-founded with Block and OpenAI. It says more than 70 agent platforms, including ChatGPT, Cursor, GitHub Copilot and VS Code, have integrated it, and that about 1 million skills have been published and nearly 280 million installed. Those are Anthropic’s own counts. The standard, name, logo, GitHub organization and agentskills.io domain move over, and the planned roadmap includes versioned releases and security guidance for people who write and install skills.

OpenAI’s Decisions API goes to all customers. According to TechCrunch, OpenAI made the API widely available on October 22, after announcing it on September 29. It is a “decision” model, meaning one that quickly picks among options and reports how confident it is, a category created by TypeSafe’s Jev. TechCrunch reports OpenAI built it by reconfiguring its Luna model, cutting response times from roughly 1–3 seconds to about 100–180 milliseconds, with Box, Datadog, Ramp and Speak as early customers. On two independent indexes it cites, OpenAI’s model ranks 8th and 14th, behind several rivals including Jev. TypeSafe’s chief executive urged users to test on their own workloads. OpenAI’s page returned an error for me.

Microsoft publishes its medical diagnosis system in Science. Microsoft AI says its MAI-DxO, a panel of AI agents that question a simulated patient and weigh test costs, reached 83% accuracy on 304 published New England Journal of Medicine cases, against 36% for physicians who could consult reference materials and 22% for those who could not. On emergency-department cases it reports 80% against 47%. These are Microsoft’s results from a simulation, with estimated test costs. TechCrunch notes specialists and physician teams were not in the comparison and that gains were small on emergency cases for newer base models. A real-world test with Beth Israel Deaconess Medical Center is planned but has no results yet.

Canva releases Canva Craft, its first image model. According to TechCrunch, it plans layout first, then generates images up to 2K resolution in 30 aspect ratios, was trained on Canva’s own templates, and is available in Canva’s Dream Lab and API. TechCrunch reports no benchmarks, and I could not find a Canva page to link.

Uber puts about 30,000 employees on ChatGPT Work. TechCrunch reports an expanded partnership announced October 22, after a pilot of about 1,000 people in which roughly 80% of early users were non-engineers. Uber did not disclose cost. The details are Uber’s statements to TechCrunch.

Sources: 1Password press release (October 22); Anthropic, Agent Skills donation (October 22); TechCrunch on Decisions API (October 22), linking OpenAI; Microsoft AI, “Building Medical Superintelligence” (October 22); Science paper, linked by TechCrunch and not opened; TechCrunch on MAI-DxO; TechCrunch on Canva (October 22); TechCrunch on Uber (October 22).

Regulation and Policy Watch

U.S.: the DeepSeek order and its first test. The Department of Homeland Security notice published October 20 in the Federal Register orders civilian agencies to exclude and remove DeepSeek products. The Bloomberg-reported raise above is the first sign of how little the order touches DeepSeek’s finances. I found no new compliance deadline.

Labor and military AI: a DeepMind engineer resigns. The Intercept reported on October 21 that Sofia Liguori, a research engineer at Google DeepMind, resigned over Google’s military contracts with the U.S. and Israel, citing Project Nimbus and Google’s classified Pentagon AI contract. Her letter urged colleagues to “stay human.” Google spokesperson Tom Hewett told the Intercept that Google follows a consistent review process for public-sector contracts, including compliance with its acceptable-use policy and AI principles. It is one resignation, not a wave, and the account is largely the engineer’s own. The Intercept notes that DeepMind employees voted to unionize in May.

Governance: Meta plans a non-voting share class. On its earnings call, CFO Susan Li said Meta will create a new class of non-voting capital stock and distribute it to existing shareholders as a stock dividend, for uses including raising capital, paying employees and acquisitions, while keeping voting power with Class A and B holders, TechCrunch reports. Semafor first reported the plan. I did not find the formal filing.

Sources: Federal Register, DHS notice, 91 FR 66346; The Intercept (October 21); TechCrunch on Meta’s share class (October 21), citing Semafor, which I did not open.

Emerging Startup Radar

TypeSafe AI, maker of Jev, valued at $7.5 billion. TechCrunch reported on October 9 that TypeSafe raised $870 million at a $7.5 billion valuation, led by Andreessen Horowitz with Sequoia and DCVC participating, weeks after Jev launched on September 15. Jev is not a language model. It outputs probabilities, which TypeSafe calls calibrated decisions, and the company claims it is faster and uses far fewer tokens than language models. The speed claim is TypeSafe’s. Today’s Decisions API item shows OpenAI now competes with it directly.

Sources: TechCrunch (October 9). Single-source; I found no TypeSafe page I could open.

AI Infrastructure and Market Signals

Nvidia packs a Vera Rubin data center into a shipping container. At GTC Berlin on October 21, Nvidia introduced DSX POD, a reference design for complete Vera Rubin AI factories that partners build inside modular containers and test before delivery. Nvidia says this cuts deployment at a prepared site from six months to as little as 14 days. That is the company’s own figure. Jensen Huang is quoted: “An AI factory creates value when it starts producing tokens — not when construction is complete.” Partners include Dell, HPE, Lenovo, Foxconn’s Ingrasys, Schneider Electric, Vertiv, Crusoe and Bull. Deliveries are expected worldwide in early 2027. The release names no European customers and gives no price. The customer still has to supply power, cooling and network connections, Nvidia’s developer material indicates, so the 14 days start only once a site is ready.

Micron’s Taoyuan union strikes over bonuses. Reuters reported on October 20 that the Taoyuan union, whose plant makes DRAM and high-bandwidth memory used in AI servers, planned a Thursday strike in solidarity with a second union in Taichung. The dispute is over profit-sharing; Reuters says mediation ended without agreement in September and members authorized a strike in early October. Micron employs about 15,000 people in Taiwan. I found no reliable report of how the October 22 action went or whether output was affected, and accounts of the Taichung union’s status conflict. The union has described it as a warning strike.

Meta’s range sets the floor, and now a new share class. The Meta share-class plan above, and its guidance of $135–145 billion in capital spending this year, which yesterday’s edition covered, belong together. Meta is asking shareholders for a way to sell stock without diluting voting control.

Sources: Nvidia investor relations, DSX POD (October 21); Reuters via The Star (October 20); Meta Investor Relations. The Star republishes Reuters’s story; I did not open Reuters’s own page.

Public Investment Watchlist

Informational only, not investment advice.

• Texas Instruments (TXN). Reported October 21 after the close: revenue $6.06 billion, up 28% from a year earlier; operating profit $2.87 billion, up 73%; earnings per share $2.62. It guided fourth-quarter revenue to $5.80–6.30 billion. CEO Haviv Ilan cited growth in industrial, automotive and data center markets. The release gives no data-center revenue figure. Third-party reports of a specific data-center growth rate are not in the filing.
• Micron (MU). The Taiwan strike above bears on supply of memory chips for AI servers. I have not verified any share-price reaction.
• Meta (META). Plans a non-voting share class, per its CFO on the earnings call.
• Anthropic and OpenAI. Neither is listed. Reported IPO plans for both rest on press accounts, not public filings.
• Intel (INTC). Reports October 29 after the close, per its investor relations page.
• Federal Reserve. The FOMC meets October 27–28.

Sources: Texas Instruments Form 8-K, Exhibit 99 (SEC); Intel Investor Relations; Federal Reserve, FOMC calendar.

Watchlist

• Anthropic’s public S-1. Whether a public filing matches the draft figures Fortune describes, and how the loss is broken down.
• OpenAI’s round. Whether CNBC’s reported terms hold, and whether a UAE compute deal is announced with it.
• Google Cloud and OpenAI. A general availability date for the OpenAI models in Model Garden.
• DeepSeek. Confirmation from the company or investors of the reported $15 billion, and any U.S. follow-up to the DHS order.
• Mistral Large 4 weights. Mistral promised them for late October. One outlet gives October 27, which I could not verify.
• Micron. Whether the Taiwan dispute settles or widens to the Taichung plant.
• Claude Duet. Who gets off the waitlist, and what independent testing finds on its approval prompts and payment handling.