20 Petaflops Sounds Insane. But How Many AI Tokens Can DGX Station Actually Generate?

It depends almost entirely on the model, the prompt length and how many requests run at once. Published independent tests of the NVIDIA DGX Station GB300 range from about 35 output tokens per second (a very large model, one user) to over 22,000 (a small model, 128 concurrent requests). For gpt-oss-120b, a popular mid-size open-weight model, StorageReview measured 9,280 output tokens per second across 128 concurrent short requests. Run around the clock, that is about 33 million output tokens an hour. At a realistic 30% average utilisation it is about 7.3 billion output tokens a month. The 20 petaflops rating barely enters the calculation, because generating tokens is limited by memory bandwidth long before compute runs out.

What throughput has actually been measured?

Source: StorageReview reviews of the MSI XpertStation WS300 (24 Aug 2026) and ASUS ExpertCenter Pro ET900N G3 (Sep 2026), using vLLM; Petronella Tech's published benchmarks (Sep–Oct 2026, setup not disclosed). No MLPerf result has been published for the DGX Station. We have not tested the system ourselves.

Two patterns stand out. Smaller and sparser models produce far more tokens. And total throughput rises steeply with concurrency, while each user's own speed stays at or below the single-stream rate: 9,280 tokens per second shared across 128 gpt-oss-120b requests is about 72 per request.

Why doesn't 20 petaflops translate into millions of tokens?

A common rule of thumb is that generating one token costs about two floating-point operations per active parameter. gpt-oss-120b activates about 5.1 billion parameters per token, so each output token needs roughly 10 billion operations. On that arithmetic alone, 20 petaflops could generate nearly 2 million tokens per second, and even the 6.1 petaflops StorageReview actually measured for NVFP4 maths would allow about 600,000.

The measured 9,280 output tokens per second uses about 95 teraflops for the generation step, under 2% of the measured compute. The limit is elsewhere. Each generation step has to read the model's active weights and every user's KV cache from memory, and the GPU's 7.1 TB/s of HBM3e bandwidth caps how many steps per second are possible. Batching more users into each step raises the total, until the KV cache for all those users fills the 252 GB of HBM3e. Our post 252GB HBM3e vs 496GB LPDDR5X explains this bandwidth limit, and What Can 20 Petaflops Actually Do? covers where the compute does get used: reading long prompts, batch work and fine-tuning.

Petaflops are a ceiling on arithmetic. Tokens per second are set by how fast memory can feed it.

How do you turn tokens per second into monthly capacity?

Multiply by seconds, then apply realistic utilisation. Using StorageReview's gpt-oss-120b figures and a 730-hour month:

• Short prompts (512 in, 512 out), 128 concurrent: 9,280 output tokens per second is about 33.4 million per hour. At 100% utilisation that is about 24.4 billion output tokens a month; at 50%, about 12.2 billion; at 30%, about 7.3 billion.
• Long prompts (8,192 in, 1,024 out), 128 concurrent: 5,023 output tokens per second gives about 13.2 billion output tokens a month at 100% and about 4.0 billion at 30%. Including input, the system processes about 35.6 billion tokens a month at 30% utilisation.
• A single user: 214 to 391 output tokens per second on mid-size models, about 0.8 to 1.4 million an hour. Few businesses keep one session running continuously, which is why shared use matters for the economics.

These are theoretical capacities, and we label them illustrative. Real utilisation depends on working hours, how evenly requests arrive, and whether you schedule batch work overnight. A system used only during office hours at moderate load might average 20–30%; one with overnight batch jobs can run much higher. The 24/7 AI Factory looks at filling the idle hours.

What decides your actual number?

• Model size and type: mixture-of-experts models with few active parameters, such as gpt-oss-120b, run much faster than dense models of similar total size.
• Precision: 4-bit formats halve the bytes read per token compared with 8-bit, roughly doubling the bandwidth ceiling.
• Whether the model fits in HBM3e: models that spill into the slower LPDDR5X tier drop sharply, as the GLM-5.2 result shows. See Can 748GB Really Handle a Trillion-Parameter AI Model?
• Prompt and answer length: long prompts add prefill work and larger KV caches, which reduce the output rate.
• Latency targets: pushing concurrency higher raises total throughput but slows each user. Interactive chat needs a per-user speed people find comfortable.
• Configuration: a high-power display card lowers the GPU's power cap, and the inference engine and its settings matter. Different OEM versions measured within about 2% of each other.

The honest approach is to benchmark your own model, prompt mix and concurrency target. Published figures set expectations; they don't replace a test with your workload.

Why does the token count matter for cost?

Because it sets the cost per token of owning the system. Spread the purchase price, power and support over the tokens you actually generate, and you have a number to compare with cloud API pricing. Because most of the cost is the hardware itself, a system at 30% utilisation produces its tokens at roughly three times the cost per token of the same system at full load. Your Company Uses 1 Billion AI Tokens a Month. Should You Buy a DGX Station? and How Many Tokens Must a DGX Station Generate Before It Pays for Itself? build on these capacity figures. For a smaller team, our post Can One GB10 Support 20 Employees Using AI at the Same Time? covers the entry-level option.

Buying in Singapore

SourceIT lists five DGX Station GB300 systems, all price on request: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. For comparison, StorageReview measured the NVIDIA DGX Spark (S$9,300, listed 5 Nov 2026) at 500 output tokens per second on gpt-oss-120b at 128 concurrent requests, against 9,280 for the GB300. Tell us your model and expected users, and we can help size the right system. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

How many tokens per second can a DGX Station generate? In StorageReview's tests, from about 35 output tokens per second for a very large model with one user to 22,161 for GPT-OSS-20B at 128 concurrent requests. gpt-oss-120b reached 9,280 at 128 concurrent short requests.

How fast is it for a single user? Typically in the low hundreds of tokens per second for mid-size models: StorageReview measured 214 for MiniMax M2.7 and 148 for DeepSeek V4 Flash, and one vendor reports about 391 for gpt-oss-120b on its own unit.

How many tokens a month is that? For gpt-oss-120b with short prompts, about 24.4 billion output tokens at 100% utilisation and about 7.3 billion at 30%. Long prompts reduce the output figure: about 4.0 billion a month at 30% for 8,192-token prompts.

Why don't 20 petaflops give millions of tokens per second? Token generation is limited by memory bandwidth, not compute. At 9,280 tokens per second, gpt-oss-120b's generation step uses under 2% of the compute StorageReview measured.

Is there an official benchmark for the DGX Station? No MLPerf result has been published for it. The figures available come from independent reviewers such as StorageReview and ServeTheHome, and they vary with model, settings and concurrency.