I read the primary posts on Sunday, 1 November, from Istanbul. Filings, earnings releases, security write-ups, a repo full of Lean. I ran nothing this week. I read.
The AI news this week comes down to one thing: the bill arrived. Anthropic made its S-1 public on 26 October and put $16.29B of first-half revenue next to an $80.7B net loss that is mostly an accounting charge. Microsoft, Alphabet, Amazon and Meta showed what compute costs in cash, and on Friday the Nasdaq Composite fell 2.9% while Intel had its worst day since the 1970s. Binance lost about $6.3B through one EIP-7702 authorization. Cathedral’s agents got into Anthropic’s back office in under 12 hours. Three labs priced new models three ways. Two labs say an AI refuted a 27-year-old hypothesis. And publishers got one robots.txt line to say no to training.
🧾 What does Anthropic’s S-1 actually show?
The SEC filing went public on 26 October, after a confidential filing in June, and proposes the symbol “ANT” on the Nasdaq Global Select Market. Share count and price range are blank. CNBC names Goldman Sachs and Morgan Stanley as lead underwriters. AP says it could top SpaceX’s record IPO, with a debut possibly next month. Could.
Revenue: $16,286M in H1 2026, about $16.29B, against $1,163M in H1 2025. FY2025 was $4,585M. FY2023 was $41M.
Now the loss, which will travel without its footnotes. H1 2026 net loss was $80,748M, about $80.7B. Of that, $72,388M is “change in fair value of convertible notes, derivative liabilities, and SAFEs”, plus $8,757M of other expense. The operating loss was only $693M. So Anthropic was close to operating breakeven, and the giant loss is mostly an accounting charge tied to its own soaring valuation. Compute and infrastructure cost $8,945M, and the filing says that line fell from 251% of revenue to 48%.
Net Revenue Retention was about 900% as of 30 September. More than 11,900 customers paid $100K+ of TTM revenue, up from more than 950 a year earlier, and about 250 paid $10M+. 48% of H1 revenue came through cloud marketplaces, up from 11% in 2023. Headcount is about 5,300. The line I reread most: “Over 90% of the contributions to our codebase in September 2026 were written by Claude.”
Compute commitments were about $265.6B as of 30 June, with terms of 1 to 20 years:
- Amazon: at least $110.0B, May 2026 to April 2036.
- Google: at least about $114.5B, April 2026 to July 2033.
- Microsoft: at least $44.9B, November 2026 to May 2033.
- SpaceXAI (formerly X.AI LLC, NVIDIA GPUs): up to $120.5B into 2029.
Another $125B to $150B sits outside that table. Class F stock, held by the seven co-founders’ Founder LLC, carries 50.1% of the vote, and the Long-Term Benefit Trust “generally has the power to elect and remove a majority” of the board.
For builders, this is the first audited look at a frontier lab’s API business: consumption pricing, nearly half of it flowing through AWS, Google Cloud and Azure marketplaces. Your model vendor is now a public-market story that warns about its own compute constraints, so I’d expect price and rate-limit changes to follow quarterly results. Caveat: with the price range blank, every “record IPO” headline is a guess.
💸 The compute bill arrives in cash
Microsoft reported on 28 October. Per CNBC, Azure, disclosed in dollars for the first time, hit $33.17B, up 48%. Capex including finance leases was $45.2B against StreetAccount’s $50.67B estimate. Not thrift. CFO Amy Hood: “Some third-party capacity deliveries we expected in Q1 were delayed.” On the call Microsoft guided calendar-2026 capex to about $170B. Free cash flow fell 27% to $18.63B, and about 75% of capex goes to short-lived chips and GPUs.
Alphabet’s release the same day showed Google Cloud up 103% to $30.7B and free cash flow at negative $9.07B. CNBC says capex guidance stays at $195B to $205B, higher in 2027, and the stock had its worst day on record Thursday, down more than 12%.
Amazon on 29 October: AWS up 39% to $46.0B, Q3 property and equipment purchases of $77.3B, TTM free cash flow an outflow of $42.4B, 2026 cash capex still about $220B.
Meta, whose free cash flow turned negative for the first time last quarter per CNBC, announced a raise of up to $95B on 26 October. CNBC, citing Meta, says Nvidia, AMD and Broadcom were interested in buying up to $5B. Suppliers buying into the customer that buys their chips. Sit with that.
Intel fell 29% on Friday, its worst day since the 1970s, per CNBC, after Q4 guidance of $15B to $16B missed the $17.1B expected. CFO David Zinsner: “We see the PC slowdown continuing at least through the first half of next year.” Nvidia fell 4%, Qualcomm 11%.
Nvidia also trimmed Vera Rubin. Per TechPowerUp, HBM4 bandwidth per GPU went from 22 to 19.2 TB/s, capacity unchanged at 288 GB, and NVLink 6 from 3.6 to 3 TB/s. Nvidia’s statement, via Wccftech, calls the old figures “preliminary architectural targets”.
As someone who pays per token: compute is scarce, expensive and now financed with debt and equity. Even Microsoft is waiting on third-party GPUs. A cheap model tier is a business decision, not a law of nature. Caveat: this is one quarter, and delayed capacity could change how these numbers read.
What the compute bill looked like, Q3 2026
| Company | 2026 capex outlook | Q3 capex | Q3 free cash flow | Cloud growth | Market reaction |
|---|---|---|---|---|---|
| Microsoft | about $170B (calendar 2026) | $45.2B incl. finance leases | $18.63B, down 27% | Azure +48% to $33.17B | up 4% after hours |
| Alphabet | $195B to $205B | $51.4B | minus $9.07B | Google Cloud +103% to $30.7B | worst day on record, down 12%+ |
| Amazon | about $220B (cash) | $77.3B (gross PP&E purchases) | TTM minus $42.4B | AWS +39% to $46.0B | down 1.4% Friday |
| Meta | $135B to $145B | not used | negative, first time | n/a | plans up to $95B stock raise |
Microsoft’s quarter is its fiscal Q1 2027 (July to September). Sources: CNBC, Motley Fool, Alphabet, Amazon IR, Meta IR.
Illustration generated with Higgsfield.
🔓 One Ethereum permission, $6.3B gone
It started on 29 October at about 20:05 UTC. TRM Labs counts about $6.36B taken across 10 chains, most of it in about 20 minutes, about four times Bybit’s $1.5B in February 2025. On 31 October the FBI named North Korea’s “TraderTraitor”.
Per CoinDesk, the attack used EIP-7702, shipped in the Pectra upgrade in May 2025, which lets an externally owned account delegate to contract code. One transaction carried 50 pre-signed authorizations, one per consecutive nonce. One matched. 3 minutes 24 seconds later, 167,464 ETH, 63 tokens and 1.615B USDT were gone. The delegation stayed live for about four hours.
It didn’t start on-chain. Binance says the entry point was a compromised DevOps engineer’s laptop, with “no evidence of any wallet private key leak”. TRM says the attacker still holds about $2.4B beyond any issuer’s reach. Tether, Circle, Ripple and WLFI froze about $2.67B.
Then Binance paid. Reuters reports it covered 95% of the $6.3B, about $6B, by buying replacement tokens with SAFU, insurance and treasury money. Withdrawals resumed Sunday in phases, bitcoin first, shortly after 06:00 GMT (09:00 Istanbul). The remaining 5% are less liquid and may take longer.
Treat pre-signed authorizations like private keys and alert on delegation changes on hot wallets. Caveat: phase one only.
Illustration generated with Higgsfield.
🕵️ Agents in the back office
Cathedral, which raised $160M at a $1.4B valuation in July per Decrypt, went at Anthropic through its bug-bounty-style program on HackerOne. The founders are former US government staffers. The Washington Post reported that after its humans failed for a few days, Cathedral’s agents got in in less than 12 hours through flaws in third-party business software, reaching meeting transcripts, salary details and Slack messages.
Anthropic confirmed access to “a limited set” of sales, HR and IT systems, said Claude conversations, source code and model weights were not reached, and that “we shut off their access the same day”. Jacob Krell of Suzu Labs, in TechRadar: “Security maturity is capped by the weakest vendor you integrate with.”
PHP2Shell, last week’s warning, is now exploited in the wild. TNW reports CISA added CVE-2026-109618 and CVE-2026-109619 to the KEV catalogue. CISA’s 28 October alert added the Linux kernel bug “Fragnesia”. TU Delft logged 68,499 exploitation requests from 652 IPs since 21 October. Dream says a campaign chained PHP2Shell with Fragnesia to root 25 servers. TNW did not verify it.
Patch to PHP 8.5.12, 8.4.27, 8.3.36 or 8.2.35. Every SaaS vendor wired into your Slack and HR stack is attack surface. Caveat: Cathedral’s account is a draft report, and Anthropic disputes parts of it.
🏷️ Gated, cheap, open: three ways to ship a model
OpenAI’s GPT-6.1 Terra (26 October) is the cheap one: $1 per 1M input tokens, $0.05 cached and $5 output on the API, which OpenAI says is 50% below GPT-5.6 Terra and GPT-6.1 Sol. OpenAI’s own numbers: more than 3x the token efficiency of Claude Opus 5.5 at max effort, about 9x lower cost per task, and Terminal-Bench 4.0 at 50.3% against 21.5% for the predecessor. OpenAI also rates it as meeting the “Critical” cybersecurity threshold under its Preparedness Framework (TNW).
Google gated Gemini 4 Carbon (28 October), covered in my Gemini 4 Carbon roundup. Google says it ranks #1 on the Artificial Analysis Intelligence Index v5.0 with 59, and hits SEC-bench Pro 88.2% without guardrails. It is not generally available. Cyber defenders get it first, then trusted partners through the “Fairwind Program”. Intro price $2/$10 per 1M tokens, then $4/$20.
Mistral opened Large 4 (see my open weights post), a 1T-parameter multimodal MoE with 49B active per token, per vLLM, at $0.70 in and $2.10 out per 1M. The license is a modified MIT, and Model-as-a-Service operators need a separate commercial license. vLLM’s day-0 recipe serves the FP8 checkpoint on 8 GPUs. Rack-class, not laptop-class.
The strongest cyber model sits behind vetting. The cheap tier is what most agents will actually run on. Open weights are real but need a rack. Route by task and keep a fallback. Caveat: every benchmark here is the vendor’s own, so don’t line up Terminal-Bench numbers across vendors as if they ran on one harness.
🧮 Did an AI just break complexity theory?
OpenAI (26 October) and Anthropic (27 October) both claim an AI found a deterministic subexponential algorithm for counting k-SAT solutions, both formalized in Lean, both found on 16 October. Anthropic’s repo says “The result was discovered by an internal research Claude model”, and that it “is purely asymptotic and says nothing about solving formulas in practice.”
Ryan Williams (MIT) is organizing an open review with Russell Impagliazzo and Ramamohan Paturi, who proposed ETH in 1999. A review, not a certification. Williams told TechCrunch “there are two or three alien-looking parts”, and that OpenAI was “reasonably confident” the result breaks no crypto systems. More in my ETH deep dive. Lean checks the proof, not whether the statement matches the paper, and that gap grows as AI writes more of our math. Caveat: unreviewed claims, and nothing here says P = NP.
🤖 One line in robots.txt
On 30 October Google described a shared robots.txt token, __usage_ai-training, based on the IETF AIPREF working group’s AI training definition. The Verge and SEJ say the draft is an individual Internet-Draft by Kevin Kelley, Anthropic’s standards lead, with no formal IETF standing yet. The example from coverage of the draft:
User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/
It controls training use, not crawling, and a * wildcard group does not match it. Google plans support “in 2027”, and treats an opt-out as a Google-Extended disallow when Google-Extended isn’t specified. OpenAI says “in the coming months”, with GPTBot controls still in force. Bing honors it from 1 February 2027 and adds a page-level AI Grounding preference that defaults to Allow. The Verge says Anthropic already adopted the signal. Microsoft says it sets up no licensing or payment.
I run WordPress sites, and one line beats a dozen bot names. I’m reading the draft and considering adding the line to my own sites. I haven’t added it. Caveat: it is voluntary, two of the big three wait until 2027, and it is a preference signal, not a paywall.
Illustration generated with Higgsfield.
📌 More AI news this week
- GitHub Universe (28 Oct): Copilot app, Copilot Fusion routing, code review Max.
- OpenAI Model Spec (29 Oct): five values, and the model is now “GPT”.
- Harmonic Aristotle (27 Oct): formal verification for C and Verilog in CI.
- Factory Droid 1.0 and Pika H3 Flex, both covered here.
Everything else lives in the AI hub.
What I’m watching next week
Moonshot said Kimi K3.1 Pro weights are due by 4 November. Anthropic’s IPO pricing fills in the blanks. Binance owes phase two. And Microsoft’s delayed GPU capacity decides a lot of next quarter’s pricing math.
Sources
- SEC, Anthropic S-1: https://www.sec.gov/Archives/edgar/data/2130357/000162828026067168/anthropicpbc-sx1.htm
- CNBC, Anthropic: https://www.cnbc.com/2026/10/26/anthropic-ipo-live-updates-ai-s-1-filing.html
- AP: https://apnews.com/article/anthropic-ipo-shares-public-ca2f4c282bef3201a392e59e3df2300c
- CNBC, Microsoft: https://www.cnbc.com/2026/10/28/microsoft-msft-q1-earnings-report-2027.html
- Motley Fool: https://www.fool.com/earnings/call-transcripts/2026/10/29/microsoft-msft-q1-2027-earnings-call-transcript/
- Alphabet: https://www.sec.gov/Archives/edgar/data/1652044/000165204426000080/googexhibit991q32026.htm
- CNBC, Alphabet: https://www.cnbc.com/2026/10/28/alphabet-googl-earnings-q3-2026.html
- Amazon IR: https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Third-Quarter-Results/
- CNBC, Meta: https://www.cnbc.com/2026/10/26/meta-stock-offering-ai.html
- Meta IR: https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-Launch-of-Offering-of-Depositary-Shares-Representing-Series-A-Mandatory-Convertible-Preferred-Stock-and-Offering-of-Class-A-Common-Stock-in-Connection-with-Forward-Sale-Agreements/default.aspx
- CNBC, Intel: https://www.cnbc.com/2026/10/30/amazon-intel-earnings-big-tech-ai-spending-concerns.html
- TechPowerUp: https://www.techpowerup.com/354133/nvidia-cuts-hbm4-speed-and-nvlink-bandwidth-targets-for-vera-rubin
- Wccftech: https://wccftech.com/nvidia-official-statement-on-vera-rubin-spec-change/
- TRM Labs: https://www.trmlabs.com/resources/blog/north-korea-steals-usd-6-36-billion-from-binance-in-the-largest-hack-in-history
- FBI: https://www.fbi.gov/investigate/cyber/alerts/2026/north-korea-responsible-for-6-billion-binance-attack
- Binance: https://www.binance.com/en/support/announcement/detail/c5a8729a97ac486d963784d96c5642d7
- CoinDesk: https://www.coindesk.com/tech/2026/10/30/here-s-how-one-ethereum-permission-let-an-attacker-drain-usd2-5-billion-from-binance
- Reuters: https://www.investing.com/news/stock-market-news/binance-to-resume-withdrawals-after-recovering-bulk-of-6-billion-in-stolen-funds-4989388
- Washington Post: https://www.washingtonpost.com/technology/2026/10/26/researchers-cybersecurity-start-up-cathedral-got-inside-anthropic/
- Decrypt: https://decrypt.co/381860/ai-agents-hack-anthropic-business-systems-12-hours
- TechRadar: https://www.techradar.com/pro/security/ai-startup-claims-it-hacked-anthropic-in-12-hours-gains-access-to-employee-details-and-internal-messages
- TNW, PHP2Shell: https://thenextweb.com/news/php2shell-flaws-exploited-cisa-kev
- CISA: https://www.cisa.gov/news-events/alerts/2026/10/28/cisa-adds-two-known-exploited-vulnerabilities-catalog
- OpenAI, GPT-6.1 Terra: https://openai.com/index/introducing-gpt-6-1-terra/
- OpenAI API: https://developers.openai.com/api/docs/models/gpt-6.1-terra
- TNW, Terra: https://thenextweb.com/news/openai-gpt-6-1-terra-computer-use
- Google, Gemini 4 Carbon: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
- Mistral: https://mistral.ai/news/mistral-large-4-oss/
- vLLM: https://vllm.ai/blog/2026-10-29-mistral-large-4
- Anthropic, ETH repo: https://github.com/anthropics/formal-math/tree/main/eth
- TechCrunch, ETH: https://techcrunch.com/2026/10/27/frontier-ai-labs-say-their-models-refuted-the-exponential-time-hypothesis-what-does-that-mean/
- Google Search Central: https://developers.google.com/search/blog/2026/10/ai-training-control
- OpenAI, training preferences: https://openai.com/index/helping-websites-express-ai-training-preferences/
- Bing: https://blogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding/
- The Verge: https://www.theverge.com/tech/1023763/openai-google-microsoft-ai-scraping-standard-opt-out
- SEJ: https://www.searchenginejournal.com/google-microsoft-openai-shared-ai-training-opt-out/595387/
- GitHub: https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/
- OpenAI, Model Spec: https://openai.com/index/updated-model-spec
- Harmonic: https://www.harmonic.fun/news/aristotle-software-hardware-verification/
