What Alibaba is up to now

New 09-22/23: Alibaba unveiled the Zhenwu V900 chip (three times the M890, clusters of up to 500,000), Qwen 4 in training and a 20GW cloud-capacity target for 2032 at Apsara, and opened its first cloud regions in Finland, the Netherlands and Turkey. Its shares fell about 4% on 09-23 on the report that China's regulator is probing DeepSeek and Moonshot. New 09-10: named by Anthropic as the operator of the largest distillation campaign it has ever measured, with Qwen 3.5, 3.6 and 3.7 identified as the models trained on reasoning harvested from Claude Opus 4.6/4.7 via prompt injection. This moves the distillation dispute from volume claims to named products for the first time. The MAP GAP flagged in the 08-24 note below is now CLOSED — `alibaba` is a watchlist entity alongside `alibaba- qwen`, so the issuer can be tagged. New 08-23/24: the parent company raised $10.2bn (HK$80bn, 710m shares at HK$112.70, a 3.6% discount) in Hong Kong's largest-ever follow-on and put all of it into AI — then FELL ABOUT 10% once the terms became tradeable on Monday, which is holders reading a discounted raise that size as dilution first and AI ambition second. ⚠️ MAP GAP: `alibaba` the company is still not a watchlist entity — only `alibaba-qwen`, the model — so the largest AI capital raise in Hong Kong's history cannot be tagged to its own issuer. No new movement since 08-15; standing picture below holds. New 08-15: Qwen's open-weight family passed 3 billion cumulative downloads across 460+ variants — more than Google's ~418M and Meta's ~227M open-weight downloads combined — per Hugging Face's "state of open models" report, with 300,000+ derivative fine-tunes and distribution into Southeast Asia/Africa through Alibaba's own cloud that most US open-weight competitors lack. First hard usage-share number this map has on China's stack winning real adoption outside China, distinct from the benchmark-parity claims (Zhipu's GLM-5.3 vs. Anthropic's Mythos 5, 08-14) the china-stack-independence thread has tracked so far. New 08-12→08-14: Qwen3.8-Max's open weights landed on Hugging Face 08-12 (2.4T total/~95B active MoE, 1M-token context), resolving the "due ~08-10" flag on the prior entry after a brief slip, paired with the revenue-share monetization policy going live as announced — large commercial cloud hosts generating over $20M/year from the model now owe a negotiated cut, mirroring Moonshot's Kimi K3 terms; Alibaba shares rose 4.5% (US premarket) and 7% (Hong Kong) on the release. New 08-10: China's regulators ordered Alibaba, alongside ByteDance and Tencent, to pull AI-companion apps over emotional-dependence harm — a consumer- product action distinct from the Qwen model line, but the same domestic regulatory environment Alibaba operates inside. New 08-14: Apple confirmed a proprietary China-specific AI model built with Alibaba's support, ending Apple's reliance on third-party models for Apple Intelligence in mainland China — Qwen now integrates across iOS/iPadOS/macOS/visionOS there, reportedly making Apple the first foreign company Beijing has cleared to deploy its own proprietary AI model domestically, China's cyberspace regulator having registered Apple Intelligence — a real distribution win for Qwen beyond Alibaba's own products. The BIS offshore-Nvidia-rental review naming Alibaba (08-07) stands unchanged and unresolved.

as of 2026-09-23