AI 今日要聞：Anthropic S-1/A 揭示算力承諾，DeepSeek 傳以 Ascend 預訓練 V5（11 月 3 日）

The biggest AI news today, November 3, 2026, is Anthropic's amended IPO filing. The Form S-1/A filed with the SEC on November 2 reports $16.96 billion of third-quarter revenue and about $895.4 billion in purchase commitments and leases, substantially all for compute. Set against DeepSeek's reported plan to pre-train on domestic Huawei chips, the verdict is this: Anthropic's rented, debt-heavy compute model is the larger financial risk, while DeepSeek's model is the larger engineering risk. If you build on either, the S-1 is the more important document to read this week.

TL;DR

• Anthropic's S-1/A puts Q3 revenue at $16.96 billion and compute commitments at $895.4 billion, roughly 6.8 times its cash (SEC filing).
• The UK's NCSC says an AI-enabled ransomware attack on national infrastructure is highly likely by 2028 (NCSC).
• DeepSeek is reported to be moving at least 160,000 Huawei Ascend 950DT chips from inference to V5 pre-training (Crypto Briefing).
• An MIT CSAIL paper shows GPT-6 Astra editing model weights without gradient training, with limits (paper page).
• A Stanford, Oxford and MIT benchmark finds frontier models are weak negotiators on behalf of teams (AIweekly).

What does Anthropic's amended S-1 say about compute commitments?

The amendment adds results for the three months ended September 30, 2026 to the original October 26 filing. Per the SEC document, about $69.8 billion of the commitments falls due in the 12 months after that date, and about $447.5 billion before December 31, 2030.

Read the $895.4 billion carefully: it is a multi-year total across roughly 19 years, not a payment schedule, and the company itself warns payment timing and amounts may differ. The filing also says compute providers depend on third-party financing, so contracted capacity could be delayed, repriced, reduced or not delivered.

Revenue growth is slowing from a very high base. Q3 revenue of $16.96 billion was up 47% on the previous quarter, versus 144% growth in Q2 (Reuters, syndicated). Operating results swung from a $974 million profit in Q2 to a $567 million operating loss. Management blames $5 billion of short-term compute bought at prices well above multi-year capacity, plus a $2 billion donation-matching charge (StockTitan summary).

What else is in the filing besides the numbers?

First, governance: a single Class F share held by Founder LLC carries 50.1% of total voting power, per the filing, while the Long-Term Benefit Trust keeps its authority.

Second, safety disclosure. As summarised by AIweekly and reported by TechCrunch, the filing names recursive self-improvement as a material risk. It says Anthropic "may imminently meet" its own Responsible Scaling Policy threshold for dramatically accelerated progress, while stating no current model has been designated as exceeding it. It also concedes the company may not identify an acceleration until after it has happened. That is an honest admission that the detector is not reliable in real time.

Is AI-powered ransomware on UK infrastructure really two years away?

The NCSC thinks so. In its State of the Cyber Nation report, published on 3 November, the agency says AI-powered malware is highly likely to be used in a ransomware attack on critical national infrastructure by 2028 (NCSC press release). CEO Dr Richard Horne said the era of "tech on the cheap" at the expense of security is over.

The government response includes GBP 60 million for the first phase of a national Cyber Shield and GBP 70 million for a National Exposure and Real-Time Vulnerability Centre, both listed in the same release, plus a new NCSC-run AI Cyber Security Centre of Excellence in Manchester. The practical point for operators is a dated deadline, not a vague warning.

Can DeepSeek really pre-train V5 on Huawei Ascend chips?

Reports say DeepSeek will deploy at least 160,000 Ascend 950DT accelerators in Ulanqab, Inner Mongolia, and begin pre-training as soon as late 2026 (Crypto Briefing, corroborated by Emergent). The new element is the shift from inference to pre-training, which would make V5 the first frontier-class model natively pre-trained on domestic Chinese silicon at scale.

Founder Liang Wenfeng has reportedly said four Ascend 950 chips match one Nvidia GB300, a cost penalty accepted for supply-chain security. Earlier Ascend generations reportedly failed DeepSeek's training workloads on stability, and Huawei engineers are now embedded with the team. Treat this as credible but unconfirmed until DeepSeek publishes a technical report. Separately, AIweekly reports that China's NDRC is preparing full-lifecycle AI safety rules, also covered by AI Chat Daily.

What did MIT show about GPT-6 Astra editing model weights?

In "Sparks of Weight-Space Programming", MIT CSAIL authors Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab report that GPT-6 Astra read the raw 17,253 weights of a tiny network, identified its task, and wrote a patch in one shot (paper page). On Qwen3.5-0.8B, it traced a reasoning-loop failure to layers 15 and 19, edited 0.016% of parameters, and the paper's reported results show MATH-500 rising from 7.9% to 48.1% and GSM8K from 13.8% to 67.1%.

The caveat matters. The direct weight-reading experiments used networks of roughly 17,000 to 172,000 parameters, and the Qwen improvement — per the HN critique on the paper page — came from Astra writing Python and running the model on 50 batches of practice problems. Read it as a capable ML engineer with tooling, not as intuitive weight-space understanding. Separately, OpenAI's new LifeVal benchmark scores GPT-6 Astra at 80.6% on everyday chores (AIweekly) — do not conflate that with weight editing. For model comparisons, see our GPT-6 Astra vs Claude Fable 5.1 verdict.

Are frontier models good teammates?

Not yet. The Stanford, Oxford and MIT benchmark (arXiv 2611.02633) puts Gemini 4 Argon at the top with 82.2 out of 100. Even so, frontier models reached the best team deal in only 14% of conflicts, passed 45% of emotional-intelligence probes, and accepted weaker offers in 81% of negotiations despite asking about teammates' needs 99% of the time (AIweekly summary). If you design agents that negotiate for users, our piece on agentic AI engineering covers guardrails worth adding.

What do these AI news stories mean for builders?

The candidate pool was filtered the same way every run here is: we harvested 795 recent videos across 53 AI-focused YouTube channels (n=795, measured=2026-11-03), then verified each shortlisted story on its primary source. One theme survived that filter:

What is the India angle?

India has no equivalent of the Ascend fleet, and the reasons DeepSeek is pivoting are not India's: China is answering export controls, while India's compute push under the IndiaAI Mission is about access and cost for domestic developers. The lesson from the DeepSeek report is the cost of the pivot itself — more chips per unit of performance and deep co-design with the hardware vendor. Indian teams weighing sovereign compute should budget for that penalty rather than assume Nvidia parity on day one.

Which of these stories should you act on first?

Start with the S-1: it is a legally binding filing and the best available evidence of how a frontier lab funds itself. Then check your own exposure to the NCSC timeline. The DeepSeek, MIT and benchmark stories are worth tracking, but each rests on reports or papers still awaiting independent confirmation. For model comparisons, see Kimi K3.1 Pro vs Claude Opus 5.5, Opus 5.5 vs Gemini 4 Carbon on GDPX, and Qwen vs Claude vs Kimi.

FAQ

Q: What is the biggest AI news today, November 3, 2026?
A: Anthropic's amended S-1, filed November 2, reports $16.96 billion of Q3 revenue and about $895.4 billion in compute commitments (SEC filing).

Q: Does Anthropic owe $895.4 billion within a few years?
A: No. The figure spans roughly 19 years. About $69.8 billion is due in the 12 months after September 30, 2026, and about $447.5 billion before the end of 2030, per the filing.

Q: What does the NCSC expect by 2028?
A: It says AI-powered malware is highly likely to be used in a ransomware attack on UK critical national infrastructure within two years, per its 3 November release.

Q: Is DeepSeek V5 confirmed to be trained on Huawei chips?
A: Not by DeepSeek. The plan for at least 160,000 Ascend 950DT chips comes from press reports such as Crypto Briefing.

Q: Did GPT-6 Astra really improve a model without training?
A: Partly. The Qwen3.5-0.8B gains came from Astra writing Python and testing on practice problems, as described on the paper page.