Zhipu Raises About $5 Billion to Fund Next-Gen GLM, Self-Training and Compute

Zhipu AI (02513.HK) announced on September 13 that it has completed approximately $5 billion in financing. According to the company's announcement, the round consists of about $2 billion in share placement and about $3 billion in convertible bond issuance, with the convertible notes carrying a zero coupon - meaning no interest cost before any conversion into equity.

The stated use of proceeds is specific: next-generation GLM foundation models, a fully self-training system and related compute infrastructure, covering large-scale training, production inference, adaptation to domestic chips and inference optimisation.

The most notable phrase in the announcement is fully self-training. Zhipu describes it as training the next generation of GLM inside environments built by the previous generation, creating a recursive self-improvement loop. The concrete work includes automatically generating and filtering training data and constructing task environments, so that the model keeps iterating inside environments it builds itself.

The structure of the raise matters too. Pairing a share placement with zero-coupon convertible bonds keeps near-term cash costs very low, while pushing dilution out to a future conversion date - a reasonable trade for a company facing long, capital-heavy compute spending.

Several outlets described the deal as one of the largest post-listing financings by a Chinese AI company. In the current market context, the money points at the same bottleneck as everyone else's: compute and training data.

The timing is a striking contrast in tempo. Abroad, leading labs are publicly debating whether to slow down for safety, with Anthropic's Dario Amodei framing a six-to-twelve-month window to act; at home, a leading lab is using capital to build out compute and close its self-training loop, writing the next round of the arms race into a corporate filing.

What to watch next: how quickly the money is deployed into actual compute purchases, whether domestic chip adaptation and inference optimisation deliver, how the next GLM performs under the self-training regime, and how conversion of the notes affects the shareholder base.