Lei Jun's fund and Huawei's Hubble back DiffuSpace's record round
Chinese money went after a non-autoregressive architecture this morning, while one of the world's most-watched investors delivered his sharpest warning yet that the AI trade is closer to the exit than the entrance.
Shenzhen's DiffuSpace has closed two back-to-back rounds totaling several hundred million yuan — according to reports, the largest funding ever raised by a diffusion language model startup, and the company's first disclosure since it was founded this May. Matrix Partners China, Shunwei Capital (Lei Jun's venture firm) and Legend Capital co-led, with CAS Star, Huawei's Hubble fund and Horizon Robotics following on; no exact figure or valuation was disclosed. What the money buys is a training schedule: the founders — HKU NLP lab co-director Professor Kong Lingpeng and his PhD students Gong Shansan and Ye Jiacheng — are training a larger dLLM and plan to open-source it this month. The team has been in this architecture since DiffuSeq in 2022 (ICLR 2023), and its Dream 7B was the first model of its size to beat same-size autoregressive models — matching the 671B DeepSeek-V3 on planning tasks while pulling more than 2.5 million downloads on Hugging Face — with Dream-VL, Dream-VLA and DreamOn variants following. Gong claims inference roughly 10 times faster than mainstream autoregressive models on equal hardware, and a partnership with agent platform Acrab claims fivefold speedups for on-device agents; both are the company's own numbers, and every outlet covering this round is working from the same first disclosure. The take: strategic money from two silicon makers says the field is being underwritten as a product bet now, not a research curiosity — and we have explained why the architecture might earn that in AI 101 — What is a diffusion language model?, with the incumbent benchmark to beat still Inception ships Mercury 2.5, the largest diffusion LLM yet.
Ray Dalio now calls AI a "classic bubble" that is "close to" bursting — his sharpest timing claim yet. Speaking at the Forbes Global CEO Conference in Singapore on October 7, the Bridgewater founder said rising interest rates and the need to turn unrealized wealth into cash are the usual pin: "you have to sell wealth in order to get money — and so the bubble usually pricks at that." That escalates his September framing from a bubble forming to one approaching its bursting point, and it lands straight after the session in which OpenAI's revenue is $20B below prior reports, AI stocks sink. Dalio has called AI bubbles before and been early; the new part is the clock, not the warning.
What to watch: whether DiffuSpace actually ships the open-source model it has promised for October — the first real test of whether a record round converts into weights anyone can run.
Is strategic money from Huawei and Horizon enough to make diffusion models a serious rival to the autoregressive stack, or is this a research premium? Tell us in the comments.
