Product1 publisher2 min readPublished
Z.ai's zero-coupon bond converts 12.5% above where the shares traded before the raise
Buyers of $3bn of Z.ai paper accepted yields as low as minus 0.5 percent, roughly $15m a year on that size, in exchange for a conversion price only about an eighth above the last close before the placement.
The Product Desk · Product desk
What happened
- Z.ai, the Chinese model developer listed in Hong Kong as Zhipu, raised about $5bn at once, roughly $2bn of it from 21.97 million new shares placed at HK$714, about 10 percent below the previous close.
- The other $3bn came as RMB 20.14bn of convertible bonds maturing in September 2027 with no coupon, priced in the filing Reuters reported to yield between minus 0.5 percent and zero.
- Z.ai already runs a data centre on Chinese-made accelerators instead of Nvidia, and the raise is earmarked for research, compute and infrastructure, expansion, investments, acquisitions and working capital.
- The placement was priced against Friday's close, two days after the NSA, FBI and CISA named Z.ai in an advisory on industrial-scale distillation; Beijing called the advisory unfounded.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- cost A paid inference tier is now priced against weights from a company approaching $1bn in sales while giving its strongest models away, and the discount that clears free comes out of the vendor's margin.
- exposure A security reviewer can cite a US federal document when asking why Chinese open weights sit in the build, and the answer falls to whoever shipped the feature, not to the model's author.
- constraint Plans that assumed hardware limits would slow Chinese model releases have to budget for funded output instead, because the rules on chips do not reach a Hong Kong bond book.
- contradiction Accepting a negative yield looks like conviction in the Nvidia-free build, but the compute allocation stays undisclosed, so the terms evidence appetite for the equity and say nothing measurable about the silicon.
Work the strike back to where the shares actually traded. The placement went at HK$714, about 10 percent under the previous close, so that close was near HK$793 [1]. Bonds convert at HK$892.5 [6]. Against the pre-deal close, that is a move of about 12.5 percent [2], half the 25 percent premium the terms quote off the placement price [6]. Holders have until September 2027 [3].
Minus 0.5 percent on $3bn is about $15m a year of certain loss [3], set against exposure to a stock that rose roughly 2,000 percent between its January listing and June [10].
Z.ai has already built a data centre running on Chinese-made accelerators instead of Nvidia [8]. The company did not say how much of the $5bn goes to compute or what it will buy with it [9]. So the book measures appetite for the equity of a company that has built one such data centre, at a strike an eighth above the market.
Z.ai is approaching $1bn in sales while giving its strongest models away [12], and its GLM line was the stealth model that beat DeepSeek earlier this year [13]. No one has published retention or usage depth for GLM deployments.
The advisory alleges that Z.AI pulled billions of tokens from GPT-5.5 and Claude Opus by mid-2026 [15]. Neither that document nor the export rules reach a Hong Kong bond book: the controls were written about Chinese access to chips, and capital was never covered [17]. Moonshot is reportedly raising $5bn of its own [18].
The people who have to act on this sat outside the book. They are the teams with GLM weights already inside a shipped feature, and the ones whose paid tier is priced against a model that is now both free and funded.
Two tests sort the roadmaps this reaches. One is whether a buyer's security review will accept weights from a company three US agencies have named [14]. The other is whether the paid tier holds up against a strong free model whose owner just took in about $5bn [1]. A plan that clears both can ignore the bond terms. For a plan that fails either, the funding calendar is the thing to price: analysts said in June that Zhipu would stay deeply unprofitable and would need to keep raising for about three years [11].
What to watch
- Any disclosure of the compute split, or an order for domestic accelerators. An order would turn one Nvidia-free data centre into a procurement commitment.
- Whether Moonshot's reported $5bn raise closes on comparable terms, showing sector-wide appetite.
- Whether the three US agencies follow the distillation advisory with anything that reaches buyers and integrators of Chinese open weights.