Anthropic's sales soar as compute commitments hit $895B
PLUS: Alibaba jumps 6% on Apple's December launch
Good morning. Anthropic's amended S-1 shows $17B in Q3 revenue, more than its whole first half. The catch: quarterly growth slowed to 47%, and the Claude maker has signed about $895.4B in compute and lease commitments while Wall Street worries that AI capacity can't keep up.
In today's issue:
📰 Today's Headlines
ANTHROPIC
Anthropic's $17B quarter comes with $895B in commitments
The Yuzu Squeeze: Anthropic just filed an amended S-1 showing roughly $17B in Q3 revenue, up about 1,127% year on year, and about $895.4B in purchase and lease commitments, nearly all for compute and infrastructure.
The details:
- Growth is cooling: Q3 revenue rose about 47% quarter on quarter, down from 144% in Q2 and 132% in Q1.
- Q3 net loss was $23.29B, versus $12.1B a year earlier. The nine-month loss of $104B was mostly a fair-value change on convertible notes and derivative liabilities.
- About $448B of the commitments is expected before end-2030 and roughly $70B by September 2027. Nine-month compute and infrastructure expenses rose $14.4B, or 329%.
Why it matters: Demand isn't Anthropic's constraint. Delivery is. Its capacity comes largely from Amazon, Google and SpaceX plus neoclouds, and hyperscalers' AI spending is already squeezing their cash flows. With a listing expected before US Thanksgiving and OpenAI aiming for 2027, Anthropic reaches public investors first, and slowing growth is the number to watch on the roadshow.
MISTRAL
Mistral's 1T open model edges 
GLM-5.3 on coding
The Yuzu Squeeze: Mistral just released the weights for Mistral Large 4 on Hugging Face. The 1-trillion-parameter model activates 52B parameters per task and has had three extra weeks of reinforcement learning since its preview.
The details:
- On Terminal-Bench 4.0, ML4 climbed from 28% at preview to 41%, edging GLM-5.3's 40%. It also rivals GLM-5.3 on Artificial Analysis' Coding Agent Index while using 47% fewer tokens per task.
- It scored 79% on CyberGym-E2E, which tests finding and patching real C/C++ memory-safety bugs. That matches MiMo-V2.6-Pro and beats DeepSeek V4.1 Flash's 76%.
- A quantised version runs on four Nvidia B200 or B300 GPUs. API pricing isn't published yet, and smaller distilled versions, including one for laptops, have no committed dates.
Why it matters: Mistral is pitching ML4 as an AI supply that can't be cut off, because once weights circulate, access can't be withdrawn. CEO Arthur Mensch says Mistral raised about $6B in three years while other labs raised hundreds of billions. ML4 tests whether that efficiency can keep a smaller lab at the frontier, but without API pricing its cost case is hard to judge.
Why it matters for Asia: Mistral's VP of Science argued that the Chinese labs dominating open models may not keep publishing weights, and he is pitching ML4 as the alternative.
🌏 Asia Headlines
ALIBABA
Alibaba jumps 6% as 
Apple dates Qwen-powered China launch
The Yuzu Squeeze: Alibaba shares just jumped 6% after Apple said on Saturday that Apple Intelligence, built on Alibaba's Qwen model, will launch in China in December. It is Qwen's biggest stage yet.
The details:
- Not all of the gain came from Apple: JD.com and PDD also rose on reports that Chinese regulators met banks and brokerages to support the market.
- Alibaba's AI Labs and Applications segment lost RMB 13.9B in adjusted EBITA in the June quarter, partly from marketing the Qwen app.
- Free cash flow was negative in five of the last six quarters, including RMB 44.7B in the June quarter. Shares last closed at 12x forward earnings, against a 13x five-year average.
Why it matters: A slot inside Apple's China rollout could win Qwen users far more cheaply than the app marketing that fed Alibaba's AI losses, and it closes Apple's feature gap with Huawei. The new Siri isn't in the first wave, though, so Qwen's real iPhone usage will take time to show. Alibaba's late-November report is the first check on whether those losses shrink.
DEEPSEEK
DeepSeek reportedly bets V5 training on 
Huawei chips
The Yuzu Squeeze: DeepSeek is reportedly planning to train its roughly 8-trillion-parameter V5 model on Huawei's Ascend 950DT SuperPoDs, according to X tipsters cited by Wccftech. It would be China's first training run on a wholly domestic hardware stack.
The details:
- DeepSeek has ordered 160,000 Ascend 950DT chips for a 1GW Inner Mongolia data center. At roughly $16,000 per chip, the order is worth up to about $2.56B, with shipments ramping in Q4 2026.
- Each chip is made on SMIC's 7nm-class N+3 process and carries 144GB of HBM. Huawei's Atlas 950 SuperPoD links up to 8,192 of them for 8 EFLOPS of FP8 compute.
- DeepSeek's stack, including its CUDA-like TileLang compiler, now runs natively on Ascend and lifts utilisation (MFU) to 45%. About 30% was previously reported for Huawei's 910C.
Why it matters: As one X commentator argues, the real test isn't whether Ascend can train big models. It's whether TileLang can carry a full frontier pre-training run without CUDA. If V5 succeeds, Nvidia's software lock-in in China weakens further, and that commentator says Nvidia's share of China's AI chip market has already fallen from 95% to 8%.
🗞️ Everything Else in AI
Anthropic's amended S-1 warns that it may soon meet its own definition of recursive self-improvement, adding that it has already used AI to build models whose capability leaps it didn't recognise at the time.
Apple reported results for its fiscal fourth quarter ended Sept 26. Going in, the market expected about $113.4B in revenue, up 10.7%, and iPhone sales of around $56.6B, up roughly 15%.
London's 
Fractile has reportedly closed a roughly $700M round at about a $6.5B valuation, with the UK Sovereign AI Fund joining and a reported $250M Anthropic order, though it hasn't announced the deal.
Salesforce agreed to buy 
Cresta, whose voice agents handle customer calls end-to-end, for an undisclosed price. It is Salesforce's third acquisition announcement in about two weeks, after 
Braze and Rune Technologies.
📌 Everything Else in Asia Tech
Nvidia has verbally told cloud customers worldwide that at least 51% of their service revenue must come from non-Chinese clients, according to the Financial Times. The rule applies only to new contracts.
ByteDance plans to bar external AI coding tools such as Claude and Codex from its L4-level code repositories and Feishu documents starting this month, even on accounts employees pay for themselves.
Samsung has reportedly committed the entire output of its second Taylor, Texas fab, a 2nm-class plant, to HBM base dies for 
AMD and Nvidia, with mass production starting in 2030.
Singapore's 
Grab raised its 2026 forecasts to revenue of US$4.15B–4.2B and adjusted EBITDA of up to US$800M, from US$740M. It also added about three million monthly users, reaching 57 million.
India's MeitY unveiled three IndiaAI Mission foundation models: Socket AI's 26B-parameter EKA-1, trained on 7.5 trillion tokens, and 
ZenteiQ's Dhīr 2B and 5B, which run offline on tested mobile hardware.
Japan's TTC and South Korea's TTA signed a letter of intent with Europe's ETSI in Seoul to explore a 3GPP-style global AI specifications body, the AI Partnership Project, without the US or China.
🔥 Hot on Social
GPT-6 ASTRA
Critic says Astra's Qwen fix wasn't raw weight-reading
On Hacker News, a post with 914 points argues that Astra's headline Qwen result, which lifted MATH-500 from 7.9% to 48.1%, used Python and practice runs, not just raw weights. It adds that the raw-weights-only tests covered only networks of 17K–172K weights.
⚡ Quick Link
Engineer Divyansh Agrawal open-sourced Qwenputer, which compiles a 32-bit RISC-V CPU into Qwen 3.5 4B's weights to run C, MicroPython and DOOM, at about 19GB.
That's it for today. See you tomorrow!
