Do inference providers really make $10–20 per GPU-hour? We measured DeepSeek V4.1 Flash on H200s

A post going around this week argues that running an inference business is easy money: rent a B300 at $8 an hour, serve open models like GLM 5.3, Kimi K3 or DeepSeek V4.1 Flash at around 100 tokens per second per user, and earn $10–20 per GPU-hour. DeepSeek itself reportedly reaches $38 off-peak.

That's a measurable claim, so we measured one version of it.

What we ran. DeepSeek V4.1 Flash (763B total parameters, FP8 weights with FP4 experts), served with vLLM 0.32.0 on a rented node with 4× H200 141 GB (RunPod Secure Cloud, $5.29 per GPU-hour), following vLLM's published recipe for H200: tensor parallel 4, FP8 KV cache, the model's Engram tables in host memory, otherwise defaults. Our usual chat-like workload (median 1,024 input and 256 output tokens, random prompts, so no cache hits), at two latency targets. 90% on our GSM8K check.

What the node can sell

At the 50 ms target, the median user saw 33 ms per token, about 30 tokens per second. The 10 ms target, the "100 tokens per second" experience, is reachable on H200, but only for about one request at a time: the node serves a fifteenth of the requests it handles at 50 ms.

What that's worth

Thirty-eight OpenRouter endpoints (33 providers) sold DeepSeek V4.1 Flash on October 27. Output prices ranged from $0.18 to $2.40 per million tokens, median $0.87. Twelve endpoints, including most of the large providers, charged exactly $0.30 input and $1.20 output.

Revenue per GPU-hour at full load = (input tokens/s × input price + output tokens/s × output price) × 3,600 ÷ 4 GPUs:

Against $5.29 per GPU-hour in rent, only the highest price clears it, and only above 60% average utilization. At the most common price, this setup loses money even when every GPU is busy all the time. A cheaper H200 changes that: at $3.59 (RunPod's community price) the common price breaks even at 81% utilization.

Why we're far from $10–20

Our number is a floor for a careful operator, not a ceiling for the market:

• Hardware. H200 has no FP4 units, so the FP4 experts run through slower paths. On OpenRouter, the FP4 endpoints (Blackwell) charged a median $0.25 per million output tokens against $0.75 for FP8. That gap is a cost advantage, and the post's example is a B300.
• Tuning. We used the published recipe and defaults. Expert parallelism, disaggregated prefill, batch scheduling and speculative-decoding settings can move capacity a lot, and big providers tune all of them.
• Cache hits. Our prompts never repeat. Real chat and agent traffic does, and cached input costs a provider almost no compute while still earning something.
• The 100 tok/s promise. Interactive speed is expensive: on this node it cut sellable output by 19×. Selling it at the same price as 30 tok/s doesn't work.

DeepSeek's own $38 comes from its own stack and traffic, at a scale no new provider starts with.

So, is it easy money?

The post's conclusion holds in one sense: margins exist and depend on utilization and the stack. But on the GPUs most people can rent today, with default settings, the common price for this model doesn't cover the rent. The money is in the gap between a default setup and a tuned one, and between today's GPUs and Blackwell. That gap is exactly what a new provider should measure before it signs a GPU commitment.

Limits

One node, one model, one platform, quick mode (2-minute windows, 2 repeats), default recipe settings, random-token prompts with no cache hits, a context limit of 12,288 tokens, list prices on the day. We tested the September release of V4.1 Flash; the October 17 release has the same architecture. Revenue assumes every token is sold at list price, with no gateway fees, discounts or idle hours.

Raw data: report · revenue calculation and price snapshot · Earlier: break-even math for an inference provider · Size GPUs for your own traffic: GPU cost calculator.