--- title: "Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI" description: "The tutorial walks enterprises through serving GLM-5.3 on two DGX Stations, roughly $190,000 at listed reseller prices, with no per-token bill to Z.AI. For a stock carrying 118.6 times sales, that reads as a price list for everything Z.AI sells." publisher: "The Inference" section: "Model" published: 2026-11-15T16:20:37.721Z modified: 2026-11-15T16:20:37.721Z canonical: https://theinference.org/article/nvidia-shows-how-to-run-z-ai-s-glm-5-3-on-190-000-of-nvidia-hardware-with-no-per-token-bill-to-z-ai language: en keywords: "Z.ai, Nvidia, Open Weights, Model Pricing, Inference Costs" --- # Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI > The tutorial walks enterprises through serving GLM-5.3 on two DGX Stations, roughly $190,000 at listed reseller prices, with no per-token bill to Z.AI. For a stock carrying 118.6 times sales, that reads as a price list for everything Z.AI sells. ## The arithmetic the tutorial left out A reader on Hacker News did the sums [Nvidia](https://theinference.org/markets/companies/nvidia)'s post never shows. Commenting under [the thread on serving GLM-5.3 across two Nvidia DGX stations](https://news.ycombinator.com/item?id=50499366), a user called liuliu pulled [Z.AI's own price list](https://docs.z.ai/guides/overview/pricing), $1.4 per million input tokens and $4.4 per million output, put two stations near $200,000, and judged that the hardware pays for itself within a year of avoided token bills [1][2]. His own caveat: the post never separates prefill from decode [1]. ## HK$199 billion on HK$1.49 billion of sales [Z.AI](https://theinference.org/markets/companies/z-ai), the Beijing lab formerly known as Zhipu, closed November 13 at HK$408.40, down 10.04% on nearly double its usual volume [3]. That is HK$199.13 billion of market value on HK$1.49 billion of trailing revenue, 118.6 times enterprise value to sales, carried over minus HK$12.38 billion of annual levered free cash flow, HK$4.58 billion of cash, and a share price 86% below its 52-week high [3]. Tokens are the product; The Inference [tracked that demand](https://theinference.org/article/alibaba-starts-selling-ai-tokens-on-taobao-october-20-after-a-z-ai-storefront-lifted-token-sales-160) in October, when a Z.ai storefront lifted token sales 160% [3]. *Chart: **HK$199B of market value on HK$1.49B of trailing revenue** Bar chart comparing Z.AI's HK$199.13 billion market capitalization with its HK$1.49 billion trailing revenue, minus HK$2.95 billion trailing net income, minus HK$12.38 billion levered free cash flow and HK$4.58 billion of cash* *Hong Kong dollars, billions. Market capitalization at the November 13, 2026, close; revenue, net income and levered free cash flow are trailing twelve months; cash is the most recent balance sheet. Source: Yahoo Finance quote page. [3]* ## Nvidia hosts the weights and writes the recipe The model outgrew the machine. GLM-5.3 carries 753 billion parameters, 40 billion active per token, and a 464 GB checkpoint against the 252 GB of memory in one GB300 [4]. [Nvidia's tutorial](https://developer.nvidia.com/blog/when-one-gpu-isnt-enough-serving-glm-5-3-across-two-nvidia-dgx-stations-with-vllm/) chains two stations into one endpoint, hosts its own quantized copy of the weights on [Hugging Face](https://theinference.org/markets/companies/hugging-face), and asks only for two license agreements: no payment to Z.AI appears anywhere in the recipe [4]. Zhipu launched the model in August, promised the weights after a safety review, and called its security skills "public goods"; Nvidia's pitch is iterating on prompts "without per-token costs" [5][4]. ## 607.8 tokens a second, then the wall Benchmarked on coding conversations, the pair peaked at 607.8 output tokens per second with 16 users in flight and fell to 385.0 at 32, where the average request waits 453.5 seconds for its answer [4]. The hardware is listed, not hypothetical: [one reseller prices the GB300 DGX Station](https://www.msn.com/en-gb/money/technology/nvidia-s-gb300-powered-dgx-station-desktop-tower-listed-for-nearly-100000-online/ar-AA2aL1M9), Nvidia's deskside AI computer, from $94,930 to $108,350, which lands two of them on liuliu's guess [6]. *Chart: **Throughput peaks at 16 concurrent users, then drops 37%** Column chart of GLM-5.3 output throughput on two DGX Stations: 107.4 tokens per second at one user, 386.0 at eight, 607.8 at sixteen, 385.0 at thirty-two* *AIPerf on the SPEED-Bench Coding set, 80 conversations, four speculative draft tokens per step, FP8 KV cache, prefix caching on; concurrency is configured concurrent requests. Source: NVIDIA Developer Blog, November 13, 2026. [4]* ## The selloff priced chips, not the tutorial Nothing ties the drop to a post that shared its date. [Jensen Huang](https://theinference.org/markets/people/jensen-huang) resigned as Nvidia's chief executive on November 12 after reports the board weighed removing him [7][8], and a draft Commerce Department rule, not yet final, would cut Chinese AI firms off from Nvidia compute rented through data centers in seven countries including Malaysia, Singapore and Thailand [8]. Z.AI runs its own data center on Chinese-made accelerators [9], and its GLM-5.3-Flash answers at $0.50 per million output tokens against the flagship's $4.4 [2]. What no counterargument restores is the per-token revenue a self-hoster stops generating. ## Two segments, no revenue per token Z.AI reports two segments, on-premise deployment and cloud, and discloses no revenue per token [3]. The next set of results will show whether open weights distributed GLM-5.3 or replaced its bill; the Commerce rule dates the other fork [8]. Nvidia's recipe prices the hardware. Only Z.AI's income statement can price what it gave away. ## Takeaway Two deskside machines can serve GLM-5.3 with no token bill to anyone, which is the one number a 118.6 times sales multiple cannot absorb. ## Sources 1. [Hacker News, comment by liuliu in the thread "Serving GLM-5.3 at 607.8 tok/s across two Nvidia DGX stations with vLLM", November 2026](https://news.ycombinator.com/item?id=50499366) 2. [Z.AI developer documentation, "Pricing", accessed 15 November 2026](https://docs.z.ai/guides/overview/pricing) 3. [Yahoo Finance, Z.AI Co., Ltd. (2513.HK) quote, profile and key statistics, close of 13 November 2026, accessed 15 November 2026](https://au.finance.yahoo.com/quote/2513.HK/) 4. [NVIDIA Developer Blog, Fatima Fall, "When One GPU Isn't Enough: Serving GLM-5.3 Across Two NVIDIA DGX Stations with vLLM", 13 November 2026](https://developer.nvidia.com/blog/when-one-gpu-isnt-enough-serving-glm-5-3-across-two-nvidia-dgx-stations-with-vllm/) 5. [South China Morning Post, "Zhipu launches flagship model GLM-5.3 as China seeks Mythos-level edge in cyber defence", 14 August 2026](https://www.msn.com/en-xl/technology/artificial-intelligence/zhipu-launches-flagship-model-glm-5-3-as-china-seeks-mythos-level-edge-in-cyber-defence/ar-AA2a7wbP) 6. [Tom's Hardware, "Nvidia's GB300-powered DGX Station desktop tower listed for nearly $100,000 online", 23 August 2026](https://www.msn.com/en-gb/money/technology/nvidia-s-gb300-powered-dgx-station-desktop-tower-listed-for-nearly-100000-online/ar-AA2aL1M9) 7. [Investing.com via Yahoo Finance, "China AI stocks tumble as Nvidia fallout fuels fears of tighter U.S. chip curbs", 13 November 2026](https://finance.yahoo.com/technology/ai/articles/china-ai-stocks-tumble-nvidia-064035259.html) 8. [Investing.com via Yahoo Finance UK, "U.S., Europe chip stocks dip as Nvidia board reportedly weighs ousting Huang", 12 November 2026](https://uk.finance.yahoo.com/news/china-chip-stocks-rally-u-063831379.html) 9. [The Next Web, "Z.ai raises $5bn in Hong Kong including $3bn of zero-interest bonds", 13 September 2026](https://thenextweb.com/news/z-ai-5bn-hong-kong-zero-interest-convertible-bonds) --- Nvidia shows how to run Z.AI's GLM-5.3 on $190,000 of Nvidia hardware, with no per-token bill to Z.AI — The Inference. Canonical: https://theinference.org/article/nvidia-shows-how-to-run-z-ai-s-glm-5-3-on-190-000-of-nvidia-hardware-with-no-per-token-bill-to-z-ai