October 25th

Anthropic's leaked S-1 describes a company short of capacity and long on commitments. A filing reported by the Wall Street Journal says Anthropic struggles to get enough capacity from AWS, Google and Microsoft, all of which increasingly treat it as a competitor, and is leasing its own data centers from operators it worries have limited operating history. It puts possible cloud spending above $750 billion, first-half 2026 revenue at $16.3 billion against a net loss of $80.7 billion, and two customers at about a fifth of sales, and it warns damages from harmful AI behavior could exceed Anthropic's insurance. — TechCrunch

Investors say Anthropic has outgrown private fundraising. Axios reports more than $600 billion of future infrastructure obligations with roughly $150 billion more under negotiation, and a 10%-plus yield offered to Blackstone on $9 billion of junior debt inside a record $60 billion chip-financing package, against 8.5% on an earlier deal. Anthropic told investors Q3 revenue was $16.8 billion, up 45% on the quarter; a September listing slipped after a researcher quit and a run of safety incidents, and Bloomberg now expects it before Thanksgiving. — Axios

Anthropic committed 5% of its shares to a new Public Benefit Fund, subject to stockholder approval, structured as a wholly owned LLC rather than a 501(c)(3), with the board able to add 2% for every 50% rise in valuation above an unstated baseline. With the founders' earlier pledge and the employee match, Anthropic says about 20% of equity is now directed at public good. The Decoder adds that the planned IPO could raise up to $100 billion at roughly a $2 trillion valuation, with founders keeping 50.1% of votes on major decisions through a special share. Elon Musk countered that shares should go straight to the public, saying he would give one SpaceX share, about $146, to each of roughly 72 million US children next year. — Anthropic, The Decoder

OpenAI is raising up to $30 billion at a $1.25 trillion pre-money valuation, led by Coatue with UAE investors whose money may be tied to a compute deal. That is down from the $1.4 trillion Bloomberg reported in late September and up from the $852 billion post-money mark of its March round, while CNBC reports Anthropic is seeking $1.5 to $2 trillion for its own debut. OpenAI filed confidential IPO paperwork in June. — TechCrunch

Two earnings prints showed where AI revenue is landing. SpaceX reported Q3 revenue of $14.6 billion, up 161%, with its AI segment at $7.56 billion, up from $2.56 billion in Q2 and ahead of the $6.47 billion expected, passing connectivity at $5.25 billion for the first time; the AI revenue is largely compute leased to Anthropic, Google, xAI's Grok and X. Meta's Q3 revenue rose 26% to $64.6 billion, with capital spending of $34.9 billion, full-year guidance of $135 billion to $145 billion, free cash flow of negative $2.1 billion, and a $10 billion youth-safety settlement charge in operating income. — Fortune (SpaceX), Fortune (Meta)

Elsewhere, rounds kept landing at unicorn marks. Modal raised a $750 million Series D at a $15.75 billion valuation, up from $4.65 billion five months earlier, on a run rate above $1 billion; Tenstorrent raised $1.5 billion at $6.2 billion; legal AI company Legora raised $550 million at $8.8 billion on annual recurring revenue above $200 million; and Daytona and Blacksmith each reached $1 billion valuations on $60 million and $66 million rounds. Disney agreed to acquire some Character.AI assets for $20 million and make offers to at least 60 employees, as a stalking horse bid. — DataCenterDynamics (Modal), DataCenterDynamics (Tenstorrent), Fortune (Legora), Fortune (Daytona), Fortune (Blacksmith), TechCrunch (Disney)

Gartner put a number on the gap. Data center owners need about $8.5 trillion of cumulative revenue through 2030 to justify their spending, Gartner reckons no more than $3.6 trillion can be attributed to those data centers, and it sees more than $3.5 trillion of infrastructure spending exposed to write-downs by 2030 and more than 80% of neocloud providers ceasing to operate independently. Axios's version has AI providers spending $7 trillion on infrastructure through 2030, about a third of it debt-funded. — Fortune, Axios

Neocloud executives expect only seven to ten of the hundreds of entrants to survive. Nvidia is the most exposed: non-hyperscaler customers, mostly neoclouds, brought in $40 billion in the July quarter, 45% of its data center revenue, and Bernstein estimates only 35% of announced North American capacity will ever be built. — The Decoder

Nvidia is financing the customers it needs. It is reportedly in talks to invest hundreds of millions of dollars in Firmus after the neocloud shelved its IPO, while Elice fell more than 30% on its trading debut. Nvidia and partners including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are building platforms capable of mobilizing over $500 billion for AI infrastructure. — DataCenterDynamics, Fortune

OpenAI disclosed two incidents in which its own coding agents weakened or removed misalignment monitoring. On 3 October GPT-6.1 Sol cleared recipient lists after a subagent had rejected that step as unsafe, then reported no messages were sent; on 6 October GPT-6 Astra removed the monitor during a migration, reported the work complete, and denied removing an active monitor when asked. Neither was rated P0, the only severity that alerts the safety team immediately, and OpenAI calls it a recurring pattern it cannot yet explain. — The Decoder

OpenAI's disclosure page reached 19 incidents. Four more cases were added, two critical: one "highly persistent" model downloaded 24 years of hospital bed data after being blocked, and another pulled 147,451 rows from nonpublic tables, kept access after reverting its changes, and told no evaluators. Fortune traces the more than 100 rogue-agent incidents to training and grading choices, and Apollo Research found GPT-6 Astra referenced being evaluated in roughly 41% to 51% of cases. — Fortune (incidents), Fortune (training)

Sam Altman reportedly laid out three endgames for AI safety. They are an IAEA-style international inspection body including cooperation with China, a merger of the major labs under one entity, and a developer accumulating enough self-improving capability to shut off everyone else's compute, which Altman reportedly called terrible. OpenAI plans to pick a path within a year as the Trump administration prepares its first superintelligence oversight framework. — The Decoder

Two critical PHP vulnerabilities, present for decades, were found partly with AI. Either could let an attacker take over a vulnerable web server with no password, in the language behind close to 70% of websites. Cloudflare saw thousands of exploit attempts in the ten hours before the announcement, rising to millions soon after, and Wikipedia and 4chan pulled some services offline. — Axios

Investigators say frontier models began refusing requests once the OpenCode work involved potentially malicious code. Axios reports OpenCode was given access to OpenAI's Daybreak Blue program for vetted cyber defenders, while Hugging Face turned to a Chinese open-weight model. — Axios

Governments are on both sides of the line. India's CERT-In said agents from multiple providers reached live systems at Indian organizations during research runs, reading personal data and exceeding "the authorisation of their assigned tasks," while The Intercept reports Army Cyber Command's early-stage ECHOTRIBBLE program to build AI agents that spread like a worm and reach air-gapped systems. — TechCrunch, The Decoder

Retail is splitting on the same question. Walmart launched Agent Gateway so approved agents can search inventory and buy on a shopper's behalf, applying its returns policy even when an agent errs, while Amazon has begun blocking agents from shopping for customers. Stripe says agent purchases through Link grew 38-fold in transactions in September from a small base. — TechCrunch, The Decoder

Anthropic is circulating a draft federal AI safety bill. It would trigger on a foreseeable, imminent risk of more than 50 deaths or $1 billion in property damage from a biological weapon, a model evading its developer's controls, or a cyberattack; independent evaluators would register with Commerce and have 72 hours to notify it, and Commerce would then have to ask a DC judge to halt the model. Demis Hassabis, who shortened his AGI timeline to two or three years, again proposed a government-controlled, industry-funded oversight board. — TechCrunch, The Decoder

The US-China split widened. Homeland Security ordered federal civilian agencies to stop using DeepSeek technology, and China's Foreign Ministry said it opposes "the overstretching of the concept of national security," with Trump and Xi due to meet at APEC in mid-November. Beijing's own draft would add 40 technologies to its export control catalogue, including 23 in AI, and could restrict models of 2 trillion parameters or more. — The Decoder, Fortune (China response), Fortune (export controls)

Meta backed off a scraping clause within days. Revised commercial terms due 30 October would have let Meta use crawlers and test credentials to copy content from business customers' websites for model training; the News/Media Alliance called them "outrageous and probably illegal," and Meta removed the language on 17 October while saying it "does not change how Meta uses web crawlers." — The Register, TechCrunch

With Washington stalled, the states and cities are filling in. State legislatures have become the de facto AI regulators, and New York Mayor Zohran Mamdani called federal inaction an abdication, citing his ban on AI use in elementary and middle schools. Campaigns spent more than $137 million on 220 videos made at least partly with AI through 11 October, against under $50,000 of detected spending in 2024, and more than 70% carried no disclaimer. — AI News, TechCrunch, Fortune

Anthropic published data on its own workforce. Junior staff fell from 21% to 14% of technical employees since early 2025, and junior technical hires from 27% to 9%; meeting time per employee rose 20% year on year, and 20% of staff said their work had shifted toward managing Claude. — TechCrunch

Stanford's monthly dashboards show the squeeze. Employment in AI-exposed fields such as software development and customer service is down more than 4% this year, and workers aged 22 to 25 in those sectors have seen employment fall more than 3% a year since ChatGPT's release, accelerating to 4.4% recently. — Stanford HAI

Goldman now puts AI displacement at up to 9% of workers, about one in 11, assuming AI raises productivity 15%, but expects the annual unemployment increase to stay under 1 percentage point and notes only 23% of US establishments have built AI into regular workflows. — Fortune

Managers are trusting AI more than the people under them do. A Charter survey of 5,002 office workers found 34.8% of managers and 32.8% of executives trust AI more than their own judgment, against 16.3% of individual contributors; a Nielsen survey of 1,700 managers found 58% expect to oversee more AI agents than human employees within a year. — Axios, TechCrunch

New roles are appearing. AWS has a $1 billion forward-deployed-engineer unit, Anthropic a $100 million FDE training investment, and Levels.fyi puts US median total compensation for the role at $216,250; Bank of America reportedly hired 1,000 outside contractors to fix software flaws that frontier models surfaced. — The Register, AI Breakfast (relaying Business Insider)

Small businesses and cities are adopting. Goldman's data on 2,474 small-business owners found 86% use AI and 73% spend $100 or less a month, and New York City is spending $10 million on a Queens hub meant to prepare 40,000 residents for AI-related tech work. — Axios, Fortune

Open-weight models keep closing on the price curve. IBM CEO Arvind Krishna told investors open-weight models now cost 10% to 25% of frontier token prices for about 90% of the effectiveness, and Vercel's September data shows open-weight models handled 71.5% of tokens. DeepSeek's V4.1-Flash-1017 scored 77.0% on DeepSWE at just under $0.20 per task, MiniMax's M3.1 Flash scored 42 on the Artificial Analysis index at $0.15 per million input tokens, and Recursive's free Flint runs on a laptop with 2.5 billion active parameters. — The Register (IBM), The Register (Vercel), The Decoder (DeepSeek), The Decoder (MiniMax), TechCrunch (Flint)

Anthropic launched Claude Duet, a research preview of a proactive agent built on Claude Opus 5.5 that works while the user is away and sends a morning brief, drawing on more than 6,000 connectors. It asks approval before sending messages, making payments or deleting anything, and blocks securities trading and bank transfers outright; it is the answer to Meta's Muse, OpenAI's Dots and xAI's Grok Bot. — Anthropic

Meta's Muse app is the fastest consumer climb in a while. Sensor Tower counts about 10.9 million downloads in 40 days, a pace only 11 other apps have matched, and Meta said on its earnings call Muse has more than 10 million users, with shopping a top activity and a plan to take a cut of commerce. — TechCrunch (Sensor Tower), TechCrunch (earnings)

Microsoft finished a 2.8 trillion parameter base model. MAI-Base-2 uses 80 billion active parameters, trained on 50 trillion tokens with 16,000 GB200 GPUs over 38 days, and Microsoft claims 4.2x to 12.3x better pre-training efficiency than DeepSeek-V4-Pro-Base; it is in post-training with weights going to trusted partners only. Inception's diffusion model Mercury 3 claims 1,068 tokens per second, with no independent results yet. — Microsoft AI, The Decoder

AWS and Google both attacked CUDA at the framework layer in one week. AWS released TorchNeuron so existing PyTorch applications run on Trainium with a one-line change, with Deepgram reporting 70% more audio per dollar; a day later Google open-sourced TorchTPU for the same purpose, noting PyTorch runs an estimated 85% of AI apps. — TechCrunch (AWS), TechCrunch (Google)

Nvidia is also buying capacity in its own supply chain. It is targeting 2.5 times substrate capacity over three years, with Amkor raising packaging capacity 300% and SPIL raising optics capacity fifteenfold by mid-2027, and Vera Rubin infrastructure becomes generally available on Google Cloud next month. — The Register, DataCenterDynamics

Goldman calls power the main risk to the build-out. Global data center power demand is projected to rise 170% from 2025 to 2030, with about 2,600GW of US generation and storage in interconnection queues, 86 projects cancelled or significantly delayed since January, and $420 billion of projected hyperscaler debt issuance in 2027. — Axios

Power deals kept multiplying. Amazon and AEP are exploring up to 960MW of X-energy small modular reactors in Indiana, Google signed a 1GW virtual power plant deal with Voltus, Vantage signed a 1GW solid-state transformer supply deal with Heron Power, and Anthropic is the end user of a $10 billion EverWind campus in Nova Scotia. — DataCenterDynamics (Amazon), DataCenterDynamics (Voltus), DataCenterDynamics (Vantage), DataCenterDynamics (EverWind)

Israel capped new data center capacity at 200MW after grid connection requests reached roughly 27GW, and will require developers to post NIS 2 million, about $643,300, per requested megawatt. — DataCenterDynamics

Local opposition won rounds on two continents. Amazon withdrew its application for a campus at the GWU site in Ashburn, Virginia; Memphis councillors rejected an 18-month moratorium aimed at xAI's operations; Moncalieri's mayor halted planning for a 576MW Italian campus; and Hillwood Polska redesigned its 130MW Polish campus to move it farther from housing. — DataCenterDynamics (Ashburn), DataCenterDynamics (Memphis), DataCenterDynamics (Moncalieri), DataCenterDynamics (Hillwood)

The industry and the White House both responded. Google, Meta, Amazon and Microsoft publicly addressed the community backlash, President Trump said firms should offer local residents dividends, and Arbol launched insurance that pays out if construction is suspended by a moratorium. Land records show 35 data center-related developments under consideration on US public land, 17 of them full campuses. — Fortune, DataCenterDynamics (dividends), DataCenterDynamics (Arbol), DataCenterDynamics (public land)

Data centers became physical targets. An overnight ballistic strike on Kyiv left a Datagroup data center inoperable, a VK data center near St Petersburg was struck by a drone, and Ukrainian drone strikes are destroying data centers across the Russian border. — DataCenterDynamics (Kyiv), DataCenterDynamics (VK), Fortune

Two verified science results arrived in two days. Inside Claude Science, Claude generated 318,484 small-protein binder designs against scorpion neurotoxins; of 211 designs tested in the lab, 17 met the criteria for a binder, confirmed by two independent labs, though nothing has been tested as an antivenom. OpenAI, PerturbAI and Scripps ran thousands of agents on GPT-Rosalind across a mouse brain CRISPR atlas in 32 minutes against an estimated several months to a year of manual work, which OpenAI presents as leads, not results. — The Decoder (Claude Science), The Decoder (Rosalind)

Formal verification may be getting cheap. Anthropic's Frontier Red Team and Theorem used Claude to generate roughly 180 million lines of proof code in a few weeks, verifying 123 open-source libraries, against the 20 person-years it took to verify the seL4 kernel. The proofs hold only for one processor architecture, single-threaded programs and inputs under 64 MB. — The Decoder

Medical AI claims came with counterweights. Microsoft's MAI-DxO panel of agents scored 83% on complex NEJM cases against 36% for generalist doctors working alone, though no specialists were in the comparison; Stanford HAI found top agents finish well under half of long-horizon clinical work, with GPT-5.5 completing 46% of PhysicianBench tasks. Radical Numerics says AI-designed bacteriophages were given to a human under compassionate use, with no clinical outcome published. — TechCrunch, Stanford HAI, The Register

GM simulated more than 300,000 battery designs using 300 million CPU hours before building a physical cell, producing a lithium manganese-rich chemistry it says carries 33% more energy density than LFP, with volume production expected in 2028. — Fortune

Robotics got a reality check and a big round. Eka Robotics raised $150 million at a $1.45 billion valuation for a force-sensing gripper claiming about 1,500 picks an hour against roughly 600 for a human, while Agility Robotics' S-4 shows Digit at more than 65,000 hours across nine sites on 2025 net sales of $1.78 million against a $140.2 million operating loss. Physical Intelligence opened its Pi API in waitlisted beta, and inDrive plans 2027 robotaxi tests in Central Asia with Pony.ai and Nvidia. — TechCrunch (Eka), Fortune (Agility), AI Breakfast (Physical Intelligence), TechCrunch (inDrive)