One H100, 3rd of 23 on Qwen

Twenty-two companies served the same open model, Qwen3.8-27B. I timed them, then added one more contestant: a single H100 I had rented minutes earlier, in Europe, set up with denv. It placed 3rd of 23.

The slowest streamed 16 tokens per second. The fastest streamed 1,279. Same model, same prompts. About an 80× gap depending on who answered. Against a typical provider in this test, the H100 was 2.6–3.9× faster, depending on the work.

This is adapted from the public post. A follow-up on what GPU providers asked for afterward is also on X. A different run, the 50.8 → 213 tuning, is a separate post. Do not merge the two boards.

How it was measured

• Four prompts for everyone: a chat answer, a piece of code, a 350-word rewrite, and a long guide.
• Three runs each. Temperature 0. 600 max tokens. Thinking off.
• One request at a time, streamed to the same laptop in the US.
• Every provider pinned through OpenRouter, fallback off, so the name on the row is who answered.
• The H100 was in Europe. Every request to it crossed the Atlantic. That shows up in time to first token, not in a way I can fully separate from the stack.

The board

Median tokens per second, as published. Ranks 10 through 21 were an ellipsis in the post. They are not filled in here.

Place	Who	Median tok/s
1	Cerebras, custom chip, FP16	1,279
2	ModelRun, FP4	278
3	One H100, FP8, denv’s stack	241
4	DekaLLM	211
5	Modal	206
6	Near AI	171
7	Tenstorrent, custom chip, FP4	149
8	CoreWeave	142
9	Wafer	130
22	DeepInfra	37
23	PrimeIntellect	16

The two ahead of the H100 were not the same kind of machine. Cerebras is a custom chip, not a GPU. ModelRun served a 4-bit checkpoint, half the bits the H100 used. Everyone else on the published list, including names you would expect to be fast, was slower.

Where it lost

Time to first token was 0.61 seconds, against 0.24 to 0.42 seconds for the leaders. Most of that, I think, was the Atlantic.

On code it hit 360 tokens per second, close to the best GPU row I called out (DekaLLM at 376). On plain chat it was 211. The chat number is the one still worth working on.

What this does not prove

One day. One request at a time. One laptop. Under real concurrency the order can change. A dedicated GPU that looks fast when it serves only you will look different when the batch fills up. The tuning post measures a slice of that, on a different day, and it should be read on its own.

Precision was not the same on every row. Comparing FP8 on one H100 with FP4 or with a wafer-scale chip is a comparison of systems, not of “the model.”

denv does not sell these providers’ GPUs. The H100 was rented. The software that placed it is what denv runs. If you have GPUs and want them listed, that conversation started from this test and is written up for providers on list your GPUs.

The video of the runs is in the original post. I am not reposting it here.