Z.ai GLM-5.3-Flash Lifts Shares 8% With All-Chinese Chip Serving

A model reveal that moved the stock

In late August, Z.ai, the Beijing AI developer known as Zhipu AI inside China, confirmed that a mystery model called "Ox Alpha" was its own GLM-5.3-Flash. The anonymous model had already reached the top of usage charts on OpenRouter and OpenCode before the company named it.4 The company described a natively multimodal mixture-of-experts system with 320 billion total parameters, 18 billion of them active per token, and a context window of just over one million tokens. It released the model under the permissive MIT license.4

The detail that caught investors' attention was not a benchmark score. Z.ai said that all online requests to the model were being handled by 100,000 Chinese-made chips.34 Its Hong Kong-listed shares rose more than 8% on August 27.4 Independent rankings added to the case: GLM-5.3-Flash placed 10th on the Artificial Analysis Intelligence Index, ahead of DeepSeek's V4 Pro Max.4

Seven weeks on, the episode tells us a lot about where Chinese AI development is heading. Domestic hardware is moving from a stopgap forced by sanctions to a selling point in its own right. The episode also shows how far Z.ai's new models now drive its share price, in both directions.

What "runs only on Chinese chips" actually means

The wording matters. Reports of the August launch say the domestic chips were used to serve the model's online requests, which is inference, the daily work of answering users.310 Founder Tang Jie repeated this on the company's August 31 earnings call. He called GLM-5.3-Flash the first GLM model served entirely on domestic-chip clusters of the 100,000 class, and said per-token inference costs had fallen 80% since the start of 2026.4

Training is less clear, and the coverage does not agree. Several accounts say the earlier flagship GLM-5 was trained entirely on Huawei Ascend hardware. One describes a 100,000-chip Ascend cluster and 28.5 trillion training tokens.1314 Huawei itself has pointed to GLM-5 as proof that its chips can handle native training of large foundation models.12 Other reporting describes Chinese frontier labs, including the team behind GLM-5.3, as running a hybrid setup: Nvidia hardware for training where they can get it, Huawei for inference, and open weights for worldwide distribution.

Our reading is that the August claim is narrower than it first sounds, but more important in commercial terms. Training is a large, one-off expense. Inference is a cost that grows with every user. A lab that can serve a top-ten model at scale without Nvidia has dealt with the part of the export-control problem that hits its revenue most directly.

The model cadence behind the rally

GLM-5.3-Flash was the smaller sibling of a busy release schedule. The full GLM-5.3 came out on August 14. It reuses the base model of June's GLM-5.2, with roughly 743 billion parameters, and adds more post-training.11 Z.ai promoted it for coding and cyber defense, called it the strongest open-weights coding model available, and published the full weights on August 28.4 Unlike the Flash variant, those weights carry a custom GLM-5.3 license rather than MIT.4 That looks like a deliberate split between a freely usable workhorse model and a more restricted flagship.

The company has released new models at a fast pace all year. GLM-5 arrived in February, GLM-5.1 was open-sourced in April, and GLM-5.2 followed in mid-June with a one-million-token context window, up from GLM-5.1's 200,000.111 Reported figures for GLM-5.2 differ slightly between outlets, with total parameters given as 744 billion or 753 billion and about 40 billion active.1318 The direction is the same either way: very large mixture-of-experts models with long context, aimed squarely at agentic coding.

In independent rankings, Z.ai sits at the front of the Chinese pack but behind the US leaders. On the Artificial Analysis index, GLM-5.3 scores 44. Anthropic and OpenAI models score between 50 and 53, Alibaba's Qwen3.8-Max scores 40, and MiniMax-M3 scores 30.5 Z.ai has also published cyber-security tests in which GLM-5.3 beat Anthropic's Mythos 5 and OpenAI's GPT-5.6.5 That is a vendor claim, and it should be judged against the broader index results, which still favor the US labs.

Why the domestic-chip story resonates

Z.ai's interest in independent hardware comes from its history. The US Commerce Department put Zhipu and nine affiliates on the Entity List in January 2025, making it the first major Chinese LLM developer to be blacklisted.4 Since then the company has steadily built its hardware independence. It announced compatibility with Huawei Ascend in August 2025 and released GLM-4.6 with FP8 and Int4 quantization on Cambricon chips the following month.1 In January 2026 it released GLM-Image, which it said was trained end to end on Huawei's Ascend servers and MindSpore framework.1516 GLM-5.2 was adapted for Huawei Ascend, Moore Threads and MetaX hardware.16

This summer the effort became physical infrastructure. In July, Z.ai began partial operation of a data center of roughly one gigawatt built only on Chinese chips. It also bought XCore Sigma, a developer of compiler and runtime software that helps workloads use different domestic accelerators efficiently. That news lifted the stock 37% in one session.4 The company has also reportedly talked with Chinese chip designers about a custom processor for the GLM family.16

The hardware is still behind. Huawei's next-generation Ascend 960DT, due in early 2027, offers about one-eighth of the FP8 throughput of Nvidia's Rubin R200, and arrives roughly a year later.12 The current Ascend 950PR is limited by memory bandwidth, low yields on SMIC's 7nm process, and a software ecosystem still catching up with CUDA.14 Huawei's answer is to link thousands of chips together, with systems that connect up to 4,096 processors.12 That helps explain why a figure like "100,000 chips" appears in Z.ai's announcements: scale makes up for weaker individual chips.1219

Z.ai is not alone. DeepSeek is reportedly preparing to install at least 160,000 Ascend 950DT chips at a gigawatt-scale site in Inner Mongolia, mainly for inference. Alibaba unveiled its Zhenwu V900 chip in September, saying it delivers three times the performance of its predecessor, and its shares rose.27 Across the sector, investors are rewarding any credible sign that China's AI stack can grow without Nvidia.

The money behind the models

The enthusiasm faded quickly. Less than three weeks after the August jump, Z.ai announced plans to raise about $5 billion through a share placement and zero-coupon convertible bonds. The shares fell more than 10%.36 It was the company's second large raise in two months, after a roughly $4 billion placement in July.69 About 60% of the new money is set aside for next-generation model research and what the company calls a fully self-training system.9

The financial picture explains the need for cash. First-half revenue rose almost 400% from a year earlier to about RMB 954 million. Over the same period the company recorded a net loss of more than RMB 2 billion and spent more than RMB 2.1 billion on research.4 A Macquarie analyst has warned that Z.ai and rival MiniMax could stay unprofitable into the next decade.5 Estimates of the company's value vary widely depending on the date and source. One profile puts the market capitalization at about $62 billion in August.1 Another puts it near $47 billion by mid-September, down from a peak of around $128 billion in July.4

There is also a geopolitical cost. In September, US agencies named Z.AI among China-based labs they accuse of running large-scale distillation campaigns against US frontier models.4

The takeaway

The 8% rise is best understood as the market pricing in proof of deployment, not proof of capability. GLM-5.3-Flash did not overtake US models. It showed that a capable, openly licensed Chinese model can be served at large scale and falling cost on Chinese hardware. For Z.ai that is the bigger achievement, because inference economics will decide whether its fast-growing revenue ever turns into profit.

The next test is GLM-6.0. Tang named it on the August earnings call as the next base model, built around recursive self-improvement, but gave no release date or specifications.4 The question to watch is whether Z.ai can say that this model was trained, not just served, on domestic silicon. That claim would answer the uncertainty that the August launch left open.