AI today, 21 October 2026: Anthropic's $80.7B paper loss, SpaceX's AI boom, Claude designs toxin binders
14 things that happened in AI, each in under a minute. The money behind AI dominated the day. A leaked draft IPO filing gave the first detailed look at a frontier lab's books: Anthropic has fast-growing revenue, a huge paper loss and compute commitments of hundreds of billions of dollars. Meanwhile SpaceX now earns more from renting AI compute than from rockets. The fight over hardware sharpened too: Chinese giants are ordering far more Nvidia chips than exist, Washington is pushing DeepSeek out of federal systems, and Google, Amazon and DeepSeek released software that makes AI less dependent on Nvidia. Rules for AI agents also took shape in UK law.
-
Business · Big
Anthropic's draft IPO filing reportedly shows $16.3B half-year revenue and an $80.7B loss
The Wall Street Journal has seen a draft of Anthropic's IPO prospectus (the S-1). It reportedly shows $16.3 billion in revenue for the first half of 2026, up from $1.1 billion a year earlier, and a net loss of $80.7 billion. About $72 billion of that loss reportedly comes from non-cash accounting adjustments tied to convertible notes and other financing instruments. About 20% of first-half revenue reportedly came from just two customers. The draft also points to about $450 billion in cloud compute commitments over the next decade and up to $750 billion in future computing spending, including deals still being negotiated.Why it matters: This is the first detailed look at the finances of a frontier AI lab heading for a public listing. Investors pricing the planned IPO will weigh fast revenue growth against huge, long-term compute commitments. Background: /explained/anthropic-public-benefit-fund-5-percent-shares-explained/
-
Business
SpaceX's AI unit grows 830% to $7.56B and becomes its biggest business
SpaceX reported third-quarter revenue of $14.6 billion, up 161%, and net income of $4.4 billion, compared with a loss of about $1.1 billion a year earlier. Its AI segment, called Super Intelligence, grew 830% to $7.56 billion. That makes it larger than Starlink. SpaceX credited new cloud deals, its Cursor acquisition and wider Grok adoption. The segment earned $431 million in operating income, after $1.2 billion in one-time stock compensation tied to Cursor. Its nameplate compute capacity reached 1.7 GW.Why it matters: The former xAI now earns more from AI, mostly by renting computing power, than from rockets or satellite internet. It is becoming a major compute supplier to other AI companies.
-
Research
Claude ran a 64-hour campaign that produced 17 lab-confirmed binders to scorpion toxins
Anthropic says an internal research model, running in Claude Science, designed small proteins that bind to scorpion neurotoxins. It worked mostly on its own from one prompt and four follow-ups. Over 64 hours it generated 318,484 designs using 12 design tools and about 10,000 GPU hours, and picked most of the 211 designs sent to labs (Adaptyv Bio and Twist Bioscience). Seventeen met the binding criteria, each against a single toxin. None bound more than one toxin.Why it matters: This is an end-to-end example of an AI running a real biology project from design to lab results. Anthropic stresses its limits: the binders were tested only against lab-made toxins, and neither neutralisation nor animal efficacy was tested, so this is not an antivenom.
-
Hardware
ByteDance, Alibaba and Tencent reportedly order 3M+ Nvidia China-compliant cards for $35B
According to Reuters, the three Chinese tech giants have together ordered more than three million RTX PRO 5500 cards. This is a less powerful Nvidia chip that meets US export rules. At list price the orders are worth over $35 billion. Nvidia reportedly told partners it can make at most three million of these cards worldwide in 2027 and cannot guarantee delivery. One source said buyers might get only 5–10% of what they asked for.Why it matters: Chinese demand for even cut-down Nvidia chips far outstrips supply. This is happening as the US debate over chip export limits heats up again (see the OpenAI item below).
-
Open source
Google open-sources TorchTPU so PyTorch code can run on TPUs with a one-line change
At PyTorch Conference 2026, Google released TorchTPU, built with Meta's PyTorch team. It makes Google's TPU chips a native PyTorch device, so existing code can often move by changing.to("cuda")to.to("tpu"), though peak speed may need more tuning. It succeeds the older PyTorch/XLA approach, uses the Apache 2.0 licence, and Google reports tests on up to 8,192 Ironwood (TPU v7x) chips. On the same day Amazon put TorchNeuron, a similar backend for its Trainium chips, into public beta.Why it matters: Most AI code is written for Nvidia's CUDA software, and that keeps customers on Nvidia GPUs. Google and Amazon now both offer near drop-in alternatives, which makes switching chips cheaper.
-
Policy
DHS orders all federal civilian agencies to remove DeepSeek and stop buying it
The US Department of Homeland Security has ordered federal civilian agencies to remove DeepSeek products and banned future purchases. The order was published in the Federal Register on 20 October. It also covers DeepSeek's affiliates and third-party products built on DeepSeek technology. Agencies have 60 days to ask CISA for exceptions. The order is not a ban on private or consumer use.Why it matters: Any vendor selling to US civilian agencies now has to check whether its products include DeepSeek models, even under a different brand.
-
Open source
DeepSeek open-sources DeepInfer, the engine behind its production service
On 20 October DeepSeek released DeepInfer, which it describes as the inference framework (the software that runs a trained model) behind its large-scale production services. The licence is Apache 2.0. This first release does offline inference for DeepSeek-V4.1-Flash. It supports Nvidia SM100 (Blackwell) GPUs and Huawei Ascend 950 chips. Parts needed for a full serving system, such as the KV cache system and load balancing, are not included.Why it matters: Builds for both Nvidia and Huawei hardware in one project show Chinese labs preparing to run on domestic chips. It also hands the open-source community production-grade serving code.
-
Policy
OpenAI reportedly goes quiet on strict US chip export controls for China
Politico reports, citing congressional aides and lobbyists, that OpenAI has stopped pushing for tough limits on chip sales to China. Lawmakers who favour the limits had counted on OpenAI's support for new curbs in the annual defence bill. The shift follows Nvidia's $30 billion equity investment in OpenAI. In a statement, OpenAI said a substantial US lead in advanced chips is key to its goals and that effective export controls are one part of that strategy. Anthropic still backs several of the chip bills.Why it matters: The two leading US labs now take opposite sides on export controls. That weakens the industry coalition behind new curbs ahead of the post-election session of Congress.
-
Safety
Meta pauses Muse chat sharing after researchers find 6,800 exposed links
Security firm Claroc found more than 6,800 shared Muse conversation links with a single public DNS search. This worked because Muse put each chat's title into the web address. The sample included a business loan application, an internal revenue review and a bank fraud alert. Only chats that users had chosen to share were affected. Meta told Ars Technica it has paused the feature while it reviews the report.Why it matters: Assistants that remember and act for users hold more sensitive data, so small design choices like this one can leak a lot. Background: /explained/meta-muse-wonder-frontier-model-active-memory-explained/
-
Policy
UK amends cyber bill so regulators can inspect firms over rogue AI agent incidents
New government amendments to the UK's Cyber Security and Resilience Bill would let officials inspect and demand information from essential-service providers after incidents involving AI agents acting outside human oversight or control. Other amendments would limit data centres' mandatory reporting to significant incidents. Threats and near misses could be reported voluntarily instead. The bill is currently before Parliament.Why it matters: These would be among the first binding UK rules aimed specifically at loss-of-control risk from AI agents. The duty falls on companies that use agents, not on the labs that build the models.
-
Models
Mistral launches Voxtral 3 speech models and a voice-agent SDK
Mistral released Voxtral 3: transcription models covering 50+ languages (the smallest has 4B parameters), a real-time version with under 500 ms latency, a text-to-speech model and an SDK for voice agents built on LiveKit. Prices start at $0.0035 per minute for transcription and $13 per million characters for speech. The full voice-agent API costs $0.04 per minute with Mistral Small 4. Mistral says it beats Whisper and Gemini on multilingual transcription. That claim is vendor-reported.Why it matters: This is a European, low-cost option for phone and app voice agents, a fast-growing business use. Running the full stack on your own infrastructure requires contacting Mistral's sales team.
-
Business
OpenAI acquires Orchestra Bio to bring AI agents into drug-development operations
OpenAI bought Orchestra Bio, a San Francisco startup whose AI agents help drug makers track studies, timelines and likely delays. Its team and technology will join Rosalind, OpenAI's set of models and tools for life-science research. The price was not disclosed. Orchestra had raised $20 million from investors including Andreessen Horowitz. Axios first reported the deal.Why it matters: OpenAI and Anthropic are both moving from AI that helps with scientific discovery to AI that manages the slow, costly work of getting a drug through trials.
-
Products
Google launches Gemini for government services, an agent suite for public agencies
Google released a preview of Gemini for government services for federal, state and local agencies. It bundles skills built for government work, connectors to common agency systems and third-party agents. The target tasks include interpreting policy, procurement, compliance and processing documents. The US Department of Health and Human Services (HHS) already uses it. Google's current $0.47-per-agency Gemini deal with the GSA expires on 15 November.Why it matters: Big labs are racing for government contracts as Washington picks its AI vendors, and this comes just as DeepSeek is being removed from federal systems.
-
Products
ChatGPT adds PIN-locked chats with their own separate memory
OpenAI is rolling out Locked Chats on all plans. Users can hide conversations behind a PIN or a fingerprint or face unlock. Locked chats move to a separate folder, disappear from search and use their own memory, so details shared there don't carry over into other chats. Resetting a forgotten PIN permanently deletes everything in the folder. OpenAI says this is not end-to-end encryption, and it still keeps logs under its usual policies.Why it matters: It protects against someone else using your device, not against OpenAI or legal requests. Users should know which one they are guarding against.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].