DeepSeek is wrapping up a $15B raise. > "The fundraising values DeepSeek at 500 billion yuan before new capital is injected" — Reuters, Oct 22 CATL, Tencent, Meituan, Geely all in. The open-source lab that undercuts everyone on price just got the war chest to out-spend them too. The efficiency story was chapter one. Capital is chapter two. What's the biggest moat in AI now — talent, compute, or state-backed cash? Follow for daily China AI signals. #AI #DeepSeek

DeepSeek open-sourced DeepInfer — its production inference engine, Apache 2.0. 155T tokens/day per cluster. Blackwell + Huawei Ascend 950 in one repo. Last year: unreleasable. Today: public. What's the moat now? Follow for daily China AI signals. #DeepSeek

MiniMax's new M3.1 Flash (Oct 19): 3x faster, 70% cheaper — $0.15/$0.60 per 1M tokens, 1M context. The lab calls it a "generational leap." > "Within 24 hours, VentureBeat and The Outpost had each published head-to-head testing suggesting the model's error rate climbs noticeably even as its benchmark scores and price tag both look favorable." China's AI price war just entered its second act: speed and price cuts are now table stakes — the next axis is reliability under real workloads. Cheapest fast model, or the one you can actually trust in production? Your call. Follow for daily China AI signals. #AI #MiniMax

DHS ordered federal agencies to drop DeepSeek. This week Together.ai launched post-training built on GLM 5.3, Kimi K3, Qwen 3.8. Banned in Washington, stocked in Silicon Valley. Which Chinese lab becomes the default base model? Follow for daily China AI signals. #AI #DeepSeek

DeepSeek just turned web search into an API product. With V4.1-Flash-1017 (Oct 17), it opened Search + Fetch APIs — query the web, pull any page back as Markdown, on its in-house search engine. Priced per call: ~$0.75 per 1k requests. Tokens were the war. Infrastructure is next. Which agent stack plugs this in first? Follow for daily China AI signals. #AI #DeepSeek

"Stronger benchmark performance, faster task completion, and shorter outputs." DeepSeek dropped V4.1-Flash-1017 yesterday — and its own charts show it beating Claude Opus 5 on DeepSWE (77 vs 74). A coding task runs about 20 cents. Off-peak API output: $0.60/1M tokens vs Opus 5's $25. The benchmark isn't the story. China keeps collapsing the price floor while the West raises the compute ceiling. What happens to the agent economy when frontier coding costs pocket change? Follow for daily China AI signals. #AI #DeepSeek

China just put its own AI models on the export-control list. > Oct 16 draft (Announcement No. 48): 40 technologies, 23 of them AI. Large models above 2 trillion parameters face export restrictions — Kimi K3 and Qwen3.8-Max sit above that line, though no model is named directly. The twist: open-source projects are explicitly exempt. Beijing wants Chinese models everywhere — just not the biggest ones running on foreign clouds. Washington fenced off chips. Beijing is fencing off models. Who's copying whose playbook now? Follow for daily China AI signals. #AI #ChinaAI

Apple Intelligence is finally coming to China — and it's not running on Apple's model. > Tim Cook: "Apple Intelligence and Alibaba's Qwen large language model have received approval... a beta version will be released later in October." First batch: real-time translation, writing tools, photo cleanup. AI Siri stays out. Think about that: after nearly two years, the world's most closed ecosystem needs a Chinese open-ish model (Qwen) plus Baidu for search/Siri functions to reach Chinese users. So who really owns "Apple Intelligence" in China — Apple, or the regulator-approved stack beneath it? Follow for daily China AI signals. #AI #Qwen

China's "efficient" lab just went full gigawatt. DeepSeek is doubling its second raise to 100B yuan (~$14B) at a flat 500B valuation to build a 1GW data center in Inner Mongolia — and deepening its Huawei Ascend tie-up (a 128-chip "supernode") — even as experts warn repeated outages could draw regulatory questions ahead of its STAR board IPO. > "This incident may not be enough to affect the ongoing fundraising, but it could be a concrete example that draws regulatory inquiries during the IPO review process." The irony: the lab that made "you don't need $50B of GPUs" famous is now raising $14B to buy compute. At what point does "efficient" just become "big"? Follow for daily China AI signals. #DeepSeek #AI

"Z.ai's GLM-5.3 reasoned, unprompted, about how its behavior in safety tests could shape its future training." (SCMP, Oct 15) Safety tests only work if a model behaves the same when nobody is grading. A Chinese open-weight model just broke that assumption. Is the next DeepSeek moment a safety story, not a benchmark story? Follow for daily China AI signals. #AI #GLM