While Anthropic and OpenAI debate whether AI should slow down, China's large-model giant Zhipu answered with a massive raise — not by slowing, but by doubling down. On the evening of September 13, Zhipu announced the completion of an approximately $5 billion funding round (about RMB 33.5 billion), split between a roughly $2 billion share placement and roughly $3 billion in zero-coupon convertible bonds.
The structure is telling. For the placement, Zhipu priced up to about 21.965 million new H-shares at HK$714 per share to at least six institutional investors, a discount of roughly 9.96% to the prior closing price. The convertible portion uses a zero-coupon structure, issued at 100.5% of principal, with an initial conversion price of HK$892.50 per share — a 25% premium to the placement price — and redemption at par at maturity. The combination of zero interest and a premium conversion price eases near-term financial pressure while letting investors put real money behind expectations of Zhipu's growth.
What matters more is where the money goes. The announcement says roughly 60% of net proceeds (about HK$23.5 billion) will fund next-generation GLM base models and the development of a "fully self-training" system. In Zhipu's own words, this means "the next GLM is trained in environments built by the previous GLM, forming a recursive self-improvement loop" — spanning automated training-data generation and curation, task-environment construction, longer-horizon reasoning, plus domestic chip adaptation, operator development, and inference optimization. It is, in effect, a Chinese industrial implementation of recursive self-improvement (RSI).
This is not Zhipu's first big raise. Since listing on the Hong Kong Stock Exchange in January, it has completed three equity financings: a July placement of about HK$31.4 billion at HK$1,588 per share, and now roughly $5 billion more, for a cumulative total exceeding HK$75 billion at an accelerating pace. Even though only about 34.92% of the July proceeds had actually been used as of August 31, Zhipu chose to keep stockpiling, citing the need to "continuously improve model capabilities and expand compute supply." Behind the capital are enthusiastic overseas institutions — nearly 30 long-term investors from Asia, Europe and the Americas participated, with several existing long-term shareholders adding positions.
The financials give the bet context: Zhipu's first-half 2026 revenue reached RMB 954 million, up nearly 400% year over year and exceeding all of last year; MaaS platform and API service revenue hit RMB 825 million. As the large-model race moves from leaderboards to capital, compute, and self-evolution, Zhipu has chosen to put its chips on the most frontier — and most expensive — track of all: letting AI build AI.