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.
Vincent Jiangthree-minute read
The arithmetic the tutorial left out
A reader on Hacker News did the sums Nvidia's−3.21% — Nvidia, down 3.21 percent today post never shows. Commenting under the thread on serving GLM-5.3 across two Nvidia DGX stations, a user called liuliu pulled Z.AI's own price list, $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 12. His own caveat: the post never separates prefill from decode 1.
HK$199 billion on HK$1.49 billion of sales
Z.AI−10.04% — Z.AI, down 10.04 percent today, 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 in October, when a Z.ai storefront lifted token sales 160% 3.
HK$199B of market value on HK$1.49B of trailing revenue
The 118.6x EV-to-revenue multiple rides on minus HK$12.4B of annual cash burn
Data
| Value | |
|---|---|
| Market cap | HK$199.13B |
| Trailing revenue | HK$1.49B |
| Trailing net income | -HK$2.95B |
| Levered FCF, TTM | -HK$12.38B |
| Cash, latest quarter | HK$4.58B |
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 chains two stations into one endpoint, hosts its own quantized copy of the weights on 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" 54.
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, Nvidia's deskside AI computer, from $94,930 to $108,350, which lands two of them on liuliu's guess 6.
Throughput peaks at 16 concurrent users, then drops 37%
Past 16 users the pair serves less in total and waits far longer per answer
Data
| Output throughput (tok/s) | |
|---|---|
| 1 user | 107.4 |
| 8 users | 386 |
| 16 users | 607.8 |
| 32 users | 385 |
The selloff priced chips, not the tutorial
Nothing ties the drop to a post that shared its date. Jensen Huang resigned as Nvidia's chief executive on November 12 after reports the board weighed removing him 78, 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.
Get Model by email, free
The AI labs, their models, and the business behind them. In our daily email, every weekday.
More about NVIDIA
NVDAFiscal Q2 2027Revenue rose 106% to $96.2B, 92.5% of it from Data Center, “driven by the ramp of our Blackwell Ultra infrastructure”. Gains on equity stakes of $7.8B lifted net income to $59.7B.
Show as a table
| Line | Value |
|---|---|
| Revenue | $96.2B |
| Gross margin | 75.0% |
| Operating margin | 66.2% |
| Net income | $59.7B |
| Hyperscale | $48.7B |
| AI clouds & enterprise | $40.3B |
| Data Center | $89.0B |
| Edge Computing | $7.2B |
| Gross profit | $72.1B |
| Cost of revenue | $24.1B |
| Other income | $7.8B |
| Operating income | $63.7B |
| Operating expenses | $8.4B |
| Tax | $11.8B |
| R&D | $7.1B |
| SG&A | $1.4B |
Reader comments
0 comments