Negative-Yield Convexity: Why Hong Kong Just Handed Z.AI $5 Billion for Free

On September 13, 2026, Z.AI — the Beijing-based model developer listed in Hong Kong as Zhipu — filed the paperwork for a capital raise that would have seemed impossible for an AI lab even a year ago: roughly $5 billion raised in a single move, split between a $2 billion share placement and $3 billion of convertible bonds that pay no interest at all.

The structure itself is the story. The 21.97 million new H-shares were placed at HK$714, a deliberate 9.96% discount to Friday’s close of HK$793 — a modest concession that institutional buyers effectively auctioned for. But the bond tranche is where the signal lives. The RMB 20.14 billion of convertible bonds, maturing September 2027, carry a zero coupon and were priced to yield between minus 0.5% and zero. In plain terms: professional investors are accepting a small, certain loss on the debt itself in exchange for the right to convert into Z.AI equity at HK$892.50 — a 25% premium over what the placement investors just paid, and roughly a 12.5% premium to where the stock traded before the announcement.

Lending money at a negative return only makes sense if the equity underneath is what you actually want. On these terms the bond is less a loan than a ticket, and the size of the book — $3 billion of it — suggests plenty of institutions wanted one. That is a stronger statement about belief in Chinese frontier AI than any benchmark score.

Where the money goes

According to the filing, proceeds break down along three lines:

• 60% to next-generation GLM foundation models — training runs, inference infrastructure, and the compute contracts behind them. This is the line with a hard constraint behind it: Z.AI has already built a data centre running entirely on Chinese-made accelerators, without Nvidia, and scaling that approach is expensive in ways that simply buying the market leader’s hardware is not.
• 15% to business expansion and strategic investments, including possible acquisitions.
• 25% to working capital — a reminder that this is a company analysts expect to stay deeply unprofitable for years while it gives its strongest models away as open weights.

The share placement is expected to settle September 16. It comes barely two months after a July follow-on that raised about $4 billion, meaning Z.AI has now pulled close to $9 billion from public markets in a single quarter.

The context that makes this extraordinary

Three threads collide in this raise, and each one sharpens the others.

First, the stock. Zhipu went public in Hong Kong in January 2026 as the world’s first listed large-model company. By June its shares had risen roughly 2,000% from the listing price. Even after pricing this placement at a discount, the company is raising from a position of extreme strength — a public-market validation that OpenAI, Anthropic, and Google DeepMind, all still private or buried inside larger entities, cannot replicate.

Second, the advisory. Two days before the placement priced, the NSA, FBI, and CISA published a joint advisory naming Z.AI among six Chinese firms accused of industrial-scale “distillation” — extracting capability from American frontier models by pulling billions of tokens from GPT-5.5 and Claude Opus through mid-2026. Beijing called the allegations unfounded. Hong Kong investors put $5 billion in three days later. Export controls were written to restrict Chinese access to chips; capital was never covered. Washington writes the rules on hardware, Hong Kong writes the cheques, and nothing in this week’s filing suggests the two are about to meet.

Third, the competition. Moonshot is reportedly preparing a $5 billion raise of its own, and the wider “AI tigers” cohort — MiniMax included — is queuing for Hong Kong listings. Z.AI’s success in pricing negative-yield paper effectively lowers the cost of capital for every Chinese lab that follows, because it demonstrates the depth of institutional demand behind this asset class.

Why open weights and a rising share price proved compatible

The puzzle Z.AI poses for Western observers is economic, not technical. The company is approaching $1 billion in annual revenue while publishing its strongest models as open weights — the stealth line that beat DeepSeek earlier this year was GLM, free for anyone to download. Conventional wisdom said giving away the crown jewels would cap the multiple; instead, open weights became the growth channel; the paid tier, enterprise contracts, and agentic products built on top of the open ecosystem do the monetisation. The market has now voted, twice in three months, that this model works.

The $5 billion question — literally — is what the next generation of GLM models trained on a substantially larger compute budget does to that gap. If efficiency-focused Chinese labs have been closing the frontier using less, the marginal value of a sudden multi-billion-dollar compute injection is not a linear improvement. It is a step change in how fast the open-weight ecosystem iterates, because every capability Z.AI bakes into GLM lands in thousands of downstream applications simultaneously.

The uncomfortable read

For Washington, this filing is a quiet policy failure rendered in numbers. The premise of export controls was that starving Chinese labs of top-end chips would slow frontier progress. Z.AI’s answer was to build without Nvidia, monetise through Hong Kong’s deep institutional pool, and now raise $5 billion at terms that assume the equity compounds faster than the debt decays. Whether or not the distillation allegations hold, the capital markets have decided the strategy is working.

For the frontier-lab pacing debate that dominated this same news cycle — Amodei, Altman, and Musk publicly calling for deliberate slowdown while Beijing rejects the framing as “fearmongering” — Z.AI’s raise is the counterexample that keeps the pressure on. A lab that raises $9 billion in a quarter is not pacing. It is accelerating, and it is doing so with other people’s money priced at negative yield.

---

Cover: dark cyberpunk financial-district aesthetic, trading floor and AI motif.