Z.AI Raises $5 Billion in Shares and Bonds to Fund Next-Generation GLM Models

Z.AI — the Beijing lab behind the GLM family of foundation models, listed on the Hong Kong Stock Exchange under stock code 2513 — announced on Sunday that it has raised approximately $5 billion through a concurrent share placement and a zero-coupon convertible bond issue, according to a September 13 filing with the Hong Kong Stock Exchange. The share placement is expected to settle on September 16, subject to regulatory approval; the bond issue has its own separate closing conditions. Neither transaction should be treated as complete until those conditions are satisfied. Combined with the company's July share placement of approximately $4 billion and its January IPO proceeds of $624 million, the September raise brings Z.AI's total capital raised in public markets since listing to roughly $9.5 billion in under nine months — making it the most capital-intensive Chinese frontier AI lab by money raised, if not by revenue.

That gap between capital raised and revenue generated is the story's structural tension. Z.AI reported first-half 2026 revenue of approximately $142 million, a 400 percent year-over-year increase but still well below its US peers and, critically, far below its 2.13 billion yuan (about $300 million) in first-half research and development spending. The company's singular use for new capital is spelled out in the filing: 60 percent of net proceeds are earmarked for research and development of its next-generation GLM foundation models, large-scale training infrastructure, and production inference systems — all of which it intends to run on domestically manufactured Chinese chips.

Two Transactions Built Around a Single Strategic Purpose

The raise consists of two independent instruments. The first is a placement of up to 21.97 million new H shares at HK$714 each — a 9.96 percent discount to the September 11 closing price of HK$793 — generating approximately $2 billion in net proceeds. The second is a zero-coupon convertible bond issue with a principal of RMB 20.14 billion (approximately $3 billion), maturing in September 2027 and settled in US dollars. The bonds carry no interest coupon and were issued at 100.5 percent of face value, meaning buyers effectively accepted a small negative yield. That willingness to accept a guaranteed loss on the debt reflects the bonds' real value: they carry an initial conversion price of HK$892.50, allowing holders to take equity at that level if Z.AI's share price recovers sufficiently before maturity.

The bond structure is itself a financial encoding of Z.AI's thesis. Buyers are not lending money expecting interest income; they are purchasing an option on the company's stock recovering to roughly double its current level within twelve months. If the shares trade at or above 130 percent of the conversion price for at least 20 of 30 consecutive trading days before February 2027, Z.AI can redeem the bonds early. If the stock does not clear that threshold, the company must repay approximately $3 billion in cash when the bonds mature in September 2027 — a material refinancing risk given that Z.AI held approximately $600 million in cash at the end of June and had fully deployed its IPO proceeds.

Read more: Z.AI's GLM-5.3-Flash: the 320B MoE that ran on Chinese chips as Ox Alpha

How Z.AI's GLM Architecture Defines Its Infrastructure Needs

The next-generation GLM models this capital is intended to support build on an architecture already visible in Z.AI's most recent production release, GLM-5.3-Flash. That model uses a mixture-of-experts design with approximately 320 billion total parameters, but only 18 billion parameters are active for any given input token. In a standard transformer, every parameter participates in every forward pass. In a mixture-of-experts model, each token is routed through a small subset of specialist modules — in GLM-5.3-Flash's case, 8 of 288 expert modules plus one shared expert that always runs. The result is that the forward pass consumes compute equivalent to roughly a 22-billion-parameter dense model, even though the full checkpoint is over 15 times larger. This is the architectural mechanism that allows Z.AI to offer GLM-5.3-Flash at $0.15 per million input tokens — approximately one-tenth the cost of its own full GLM-5.3 flagship — without simply sacrificing capability.

GLM-5.3-Flash also integrates a hybrid attention design combining sparse and linear attention layers that achieves a 4.44-fold reduction in KV cache size at million-token context lengths compared to its predecessor, and supports up to 1,048,576-token contexts natively. The model runs inference on a cluster of more than 100,000 domestic chips; the company has not officially confirmed the supplier, though analysts cited in industry reporting identified the hardware as likely from Huawei's Ascend series.

Z.AI's flagship GLM-5, a 744-billion-parameter model trained in early 2026, was confirmed to have been trained entirely on Huawei Ascend 910C chips with no NVIDIA hardware involved — a first at that parameter scale. The training run required roughly 15 percent more compute time than an equivalent NVIDIA-based run, and inference throughput on Ascend hardware runs approximately 17 to 19 tokens per second against 25 or more on NVIDIA-backed systems. These efficiency gaps are structural costs that the company absorbs in exchange for supply-chain independence under US export controls.

China's AI Capital Wave and Z.AI's Place Within It

Z.AI's raise is the most prominent recent transaction in a sustained acceleration of Chinese AI infrastructure financing. Competitor MiniMax, which also listed on the Hong Kong Stock Exchange in January 2026, raised $2 billion in July. Moonshot AI, developer of the Kimi model series, closed a $2 billion round in May at a $20 billion valuation. DeepSeek — widely regarded as the most technically rigorous Chinese lab — raised $7.4 billion in its first-ever external financing in June. Alibaba has pledged more than $10 billion toward AI infrastructure; ByteDance has announced approximately $23 billion in 2026 AI-related capital expenditure.

Within this field, Z.AI occupies a distinctive position. Its open-weight strategy — releasing MIT-licensed model weights on Hugging Face — has driven API user registrations toward approximately 7 million by mid-August 2026. Independent analysis described the company's on-premises deployments to Chinese state-owned enterprises as carrying approximately 40 percent gross margins, structurally similar to Palantir's enterprise model, which differentiates it from peers relying primarily on commodity API revenue. The Stanford HAI AI Index 2026 estimated the performance gap between US and Chinese top models had narrowed to approximately 2.7 percent, and Z.AI's GLM-5.2 topped the Artificial Analysis Intelligence Index among open-weight models in June.

What the September 2027 Maturity Date Will Measure

The hardest constraint on Z.AI's plan is not the chip supply chain or the model architecture — it is time. The zero-coupon convertible bonds mature in September 2027. That single date will reveal whether the sovereign AI compute thesis produced enough commercial momentum to justify a stock recovery to the conversion level, or whether Z.AI must find $3 billion in refinancing capacity while simultaneously funding its next training run. The company has stated it expects to deploy all proceeds by June 2028, implying the training infrastructure investment is a multi-year project. But the bond clock runs on a much shorter timeline.

The settlement of the current share placement, expected on September 16, is the nearer milestone — and the more straightforward one. The bond maturity in September 2027 will be the actual verdict on whether building frontier AI models on domestic Chinese silicon, at a cost premium, at a revenue base a small fraction of OpenAI or Anthropic's, remains a viable long-term strategy or whether it requires a structural rethinking of how Chinese frontier labs raise and deploy capital.

Read more: Alibaba raises $10.2B in Hong Kong's record share sale to fuel AI infrastructure