Zhipu raises $5 billion in its second financing in two months
The capital race among China's frontier labs keeps accelerating — and Shanghai AI Lab spent the quiet side of that war today, shipping a 397B open-weight model built to read science papers the way a researcher does.
Zhipu has completed a financing of roughly $5 billion, its second raise in two months and its third equity event since going public in January. According to the company's Hong Kong exchange announcement, the deal splits into a $2 billion share placement — up to 21.965 million new H-shares at HK$714, about a 10% discount to the prior close — plus $3 billion of zero-coupon convertible bonds due September 2027, issued at a conversion price of HK$892.50, a 12.55% premium to that close. The zero-coupon, premium-conversion structure is the tell: Zhipu is trading away interest expense entirely because bondholders only profit if the share price rises well above today's level — a bet on the next two years of GLM. Proceeds (cumulative market funding now over $10 billion) are earmarked for the next-generation GLM foundation model, a fully self-trained system, long-horizon task reinforcement learning, and inference capacity for the MaaS platform, with the company noting compute contracts require lead time and it wants deployment aligned with expansion. The read: DeepSeek's $7.4 billion raise and Kimi's IPO filing turned China's model race into a capital-accumulation contest, and Zhipu — the only one of the four already listed — can tap markets on a quarter's cadence while private rivals negotiate one-off rounds. We covered the economic logic behind that head start in GLM-5.3-Flash proves frontier intelligence is now a Chinese-chip economics story.
Shanghai AI Lab released Intern-S2-397B into open weights, a multimodal model whose training pipeline reads raw pages of scientific literature instead of parsed text. Per the model card, jointly modeling symbolic semantics and visual layout in one representation space is meant to strengthen spatial and visual reasoning, backed by multi-task reinforcement learning across more than 20 scientific domains and long-horizon agent RL in sandboxed environments; the lab claims leading general-reasoning performance among open models and strong results on tasks like biomolecular interaction design. The collection also ships a 35B sibling, Intern-S2-Preview, continued-pretrained from Qwen3.5, which the lab says matches the trillion-scale Intern-S1-Pro on core scientific tasks. An 8×1 A100-homogeneous cluster can't hold the big one, but FP8 variants are posted — and the paper itself calls the preview a preview, flagging long-workflow reliability as open work. This is the lab acting on the thesis its chief laid out this morning in Shanghai AI Lab chief: science is the next programming.
What to watch: whether Zhipu's June-2028 spend timeline drags into a fourth raise before year-end, and whether anyone reproduces Intern-S2's science-agent claims outside the lab's own harness.
China's frontier labs now treat the capital markets as part of the training cluster. Is open-weight science intelligence worth $5 billion — or is this arms-race accounting? Tell us in the comments.
