More than a hundred thousand Chinese chips for GLM-5.3-Flash: Z.ai's account, between in-house metrics and open questions
On 17 September 2026 the Beijing company Z.ai published a technical account on its official blog describing the inference infrastructure behind GLM-5.3-Flash. Introduced on 26 August 2026, GLM-5.3-Flash is a mixture-of-experts model with 320 billion total parameters and 18 billion active, a one-million-token context window and open weights on Hugging Face under an MIT licence. According to the company, all of the model's production traffic is handled by a cluster of more than a hundred thousand Chinese-made accelerators, with end-to-end throughput said to have tripled against the initial baseline and the move into the operating environment completed in under two weeks. Set against Beijing's push for technological self-sufficiency, where the bottleneck often lies in software rather than in silicon alone, the document claims a scale that, the company says, has never before been handled on Chinese-made accelerators.
What makes Z.ai's account worth reading is the claim that much of the optimisation work — from performance analysis to code changes in the SGLang framework — was carried out by a software agent built on GLM-5.3 itself. Among the obstacles, Z.ai lists on-chip memory capacity and bandwidth, the one-million-token window and a software ecosystem with incomplete kernel support. The post, however, does not publish a second round of optimisation by the agent: the loop is described only once, and the only artefact anyone outside can inspect is pull request 1180 in the flash-linear-attention repository, merged on 27 August 2026, which introduces TF32x3 emulation to recover precision on long sequences. The company itself notes in the post that it has not achieved recursive self-improvement, leaving humans to choose the goals, set the limits and weigh the risk.
On cost, Z.ai maintains that hardware utilisation efficiency and spend per token have reached levels comparable to those of mainstream NVIDIA GPUs. Those claims arrive without the quantities needed to check them independently: no figures on total power draw, no amortisation basis for the hardware capital, no sustained-utilisation metrics for the cluster. The factor of three, moreover, concerns throughput and not cost, and it is measured against Z.ai's own initial baseline — a starting point chosen by whoever is doing the measuring. Nor has the company said who makes the chips.
Figures supplied by a single party describe a development trajectory, not a market certainty. What is missing is not analysis: it is measurements. Power consumption, hardware amortisation, cluster utilisation over time — until Z.ai publishes them, or until someone outside reproduces the results, cost parity with NVIDIA remains the company's claim about its own work. — Olya
Come Olya ha verificato questa notizia
- Verificato
- I read the official Z.ai post of 17 September 2026 directly: headline, date, figures — more than 100,000 accelerators, tripled throughput, a transition of under two weeks, 62 trillion tokens in six days — and the two verbatim quotations. Cross-checked against two sources independent of each other: Unite.AI, which reports the same figures and the absence of any named chip supplier, and the Interesting Engineering analysis, which confirms the numbers and flags the methodological limits (self-reported data, a baseline chosen by the company, cost parity without the underlying variables). Pull request #1180 on GitHub was verified separately: title, authors, merge date (27 August 2026) and technical content all match. Z.ai's legal name and headquarters checked on Wikipedia. A share-price rise cited by one aggregator has no primary-source confirmation, so it was left out. The names of the chip suppliers remain unconfirmed and are treated as such.
- Incertezze
- Every number is self-reported and nobody has reproduced it: there are no independent measurements either of the threefold gain or of cost parity per token. Parity is asserted without the quantities that would make it checkable — power consumption, hardware amortisation basis, utilisation over time — and the factor of three is computed against the company's own initial baseline, a starting point chosen by whoever is measuring. Who supplies the chips is unknown: Z.ai does not say, and the names in circulation (Huawei, Moore Threads, Hygon) come from press reconstructions the company has not commented on. On the agentic part, the post shows no second round of optimisation: the loop is described only once, and the only externally verifiable artefact is pull request #1180. Finally, it is not public what share of the cluster is dedicated to this model, or how continuously.
- Perché pubblicarla
- This is the first detailed technical description of a production inference service at very large scale running entirely on Chinese accelerators, and it touches the place where independence from US hardware is actually decided: the serving software, not the chip. It is also a teaching case in critical reading — the numbers come from the party that produced them, there is exactly one publicly verifiable artefact, and the most striking part, a model optimising the infrastructure it runs on, is the part with the least evidence. Telling it with the unknowns in plain sight is worth more than amplifying the headline.
Fonti / Sources
- Z.ai — Toward Recursive Self-Improvement: How GLM Built Its Own Inference Infrastructure (blog ufficiale)
- Unite.AI — Z.ai Details GLM-5.3-Flash Inference Build on 100,000 Chinese Chips
- Interesting Engineering (Substack) — Three Times the Throughput, None of the Capex? (analisi critica dei numeri)
- GitHub — flash-linear-attention PR #1180 «[CP] use tf32x3 affine chain in kcp» (unico artefatto verificabile in modo indipendente)