← Lab notes

Five GPU Marketplace Nodes Measured: All Healthy, 30–37% Less per Token on RTX PRO 6000 and H100

Lium is a GPU marketplace where each host sets its own price. We rented five nodes, one GPU each, and ran our standard evaluation: a health check against a healthy card, then Qwen3-8B on vLLM under a latency target. All five ran at their full power limit with normal clocks and memory bandwidth. RTX PRO 6000 and H100 served the same load as the same cards on RunPod Secure Cloud and cost 30–37% less per million output tokens at the hosts' prices; Lium's H200 cost the same per hour as RunPod's and its B200 more. Lium provided the credit; the measurements and the write-up are ours.

Choosing GPUs? Compare GPUs for your traffic · Get your node measured, free

Disclosure: Lium provided the credit for these runs (we used $31.75 of a $75 credit). They saw the results before publication and confirmed the node details; the measurements and this write-up are ours.

On a GPU marketplace, anyone can list a machine and each host sets its own price, so you'd expect surprises. We've found them before: an H100 from a marketplace community host running at 40% of its normal clock, and on a managed cloud, H100 NVL cards set to 310 W. So when Lium asked to be listed on our GPU price page, we asked to measure some of their nodes first.

All five nodes were healthy. RTX PRO 6000 and H100 served the same load as the same cards on RunPod Secure Cloud, at 30–37% less per token.

What we ran

We rented five nodes on November 10, one GPU each, all on Lium's Secure tier, its tier for uninterrupted work. We picked them by a rule written down before renting: for each GPU model, the highest-reliability nodes from two different hosts, with at least 100 GB of free disk and 500 Mbit/s download. Each node got the same two steps we run everywhere:

  1. Health check against a healthy card: power limit, clocks under load, memory bandwidth, BF16 compute, ECC, PCIe.
  2. Capacity: Qwen3-8B BF16 on vLLM 0.33.0, a chat-like workload (median 1,024 input and 256 output tokens), and the highest request rate that holds p99 time to first token ≤ 2 s and p99 time per output token ≤ 50 ms. Quick mode: two-minute windows, two repeats. Then an answer check on 50 GSM8K questions.

Results

Req/s is the highest request rate that held the latency target; $/M is the cost per million output tokens at the node's price.

Node GPU Location $/GPU-h Health Req/s $/M output tokens
6b40f6ba RTX PRO 6000 Almaty, KZ 1.75 pass 3.09 0.39
88e2f6eb RTX PRO 6000 Dublin, IE 1.67 pass 3.09 0.38
df4234ab H100 SXM Helsinki, FI 2.50 pass 8.79 0.19
f092246d H200 Jackson, US 5.30 pass 9.89 0.37
1272828d B200 Kansas City, US 8.88 pass 19.10 0.31
  • All five were healthy. Every card ran at its full default power limit (600 W, 700 W, 1,000 W), clocks held under load, and memory bandwidth and compute matched our healthy-card references (RTX PRO 6000 Blackwell Server Edition: 1.46 TB/s and 406–407 TFLOPS, against 1.46 and 408; H100: 3.05 TB/s and 684 TFLOPS, against 3.04 and 675). No power caps, no throttled clocks. Answers were 92–94% correct on every node, as on our earlier runs.
  • Same throughput as the same cards on RunPod. The RTX PRO 6000 nodes handled 3.09 requests/s and the H100 8.79, the same as our earlier runs of this model on RunPod Secure Cloud (RTX PRO 6000, H100). Our search steps through fixed request rates, so "the same" means within one step; output tokens per second agreed to within 0.1%.
  • Two hosts, one result. The two RTX PRO 6000 nodes, from different hosts in Kazakhstan and Ireland, measured the same.

Price is where the difference is

With the same throughput, cost per token follows the hourly price. On a marketplace each host sets its own, and we picked nodes by reliability, not price, so the table shows the whole range of Lium's Secure asks for single-GPU rentals on November 10 next to RunPod's list prices that day (unchanged since our earlier tests):

GPU Lium asks (nodes) We measured RunPod Secure RunPod Community
RTX PRO 6000 $1.50–1.75 (4) $1.67, $1.75 $2.49 $1.69
H100 SXM $2.50 (3) $2.50 $3.99 $2.69
H200 SXM $3.85–5.30 (8) $5.30 $5.29 $3.59
B200 $6.10–8.88 (3) $8.88 $7.99 $5.98
  • RTX PRO 6000 and H100: 30–37% less per token than RunPod Secure. $0.38–0.39 against $0.56 per million output tokens, and $0.19 against $0.31. Lium's Secure prices for these cards were close to RunPod's Community Cloud, whose hosts we haven't measured on this model.
  • H200 and B200: the nodes we measured were the most expensive listed. Our rule took the most reliable node, which here was also the priciest: the H200 cost the same per hour as RunPod Secure, the B200 11% more. Cheaper H200 and B200 nodes were listed (the median H200 ask was $4.20); we didn't measure them.
  • An 8B model doesn't need an H200 or a B200. On this workload they cost $0.37 and $0.31 per million output tokens, against $0.19 on the H100. We include them because they were healthy and fast, and because larger models are what they're for.

What this means if you rent on a marketplace

  • Check each node before you rely on it. These five were fine; that's a sample, not a guarantee. A two-minute check catches the expensive surprises: pip install tokenwatt-check && tokenwatt-check (what it checks).
  • Compare per token, not per hour. A healthy card at a lower hourly price is simply cheaper per token; a capped or throttled one can erase the discount.
  • Pick the tier for the job. Lium's Spot tier is cheaper, but the host can reclaim the node at any time (renters aren't charged while it's down). All five of these were Secure.

Limitations

  • Five nodes on one day, from four hosts (the H100 and the B200 had the same host). A marketplace changes by the hour; this doesn't represent every host.
  • Only the RTX PRO 6000 had two nodes from different hosts. Both qualifying H100 nodes belonged to one host, one B200 node met our rule, and the three H200 candidates for a second host were already rented or offline when their turn came.
  • One H100 run failed on our side (a file copy) and we reran it on the same node.
  • Quick mode, one model, one GPU per node. It can't separate differences of a few percent.
  • Our earlier RunPod measurements used vLLM 0.31.0; these used 0.33.0.
  • Prices are list prices on November 10, 2026.

The data

Per-node reports with the raw JSON: RTX PRO 6000, Almaty · RTX PRO 6000, Dublin · H100, Helsinki · H200, Jackson · B200, Kansas City. Method: how we measure. Prices across providers: GPU prices.

Keep reading

6 min read

H100 vs H200 for LLM inference: 18% more capacity, 33% higher price — and where the H200 pulls ahead

We ran the same model, engine and latency target on an H100 SXM and an H200 on RunPod. At a 50 ms-per-token SLO the H200 served 18% more requests for 33% more money, so each token cost 11% more. At a fixed load it generated tokens 36% faster, and with a 30 ms target its lead grew to 29%, close to breaking even on price. Measured numbers, cost per million tokens and the break-even price.

H100H200LLM inference
5 min read

MI350X vs H100: AMD's own software stack more than doubled the MI350X; tuning the H100 changed nothing

We served Qwen3-8B and Qwen3-Coder-30B on one AMD Instinct MI350X and one H100 SXM under the same latency target. On vLLM's default ROCm build the MI350X served 36% fewer tokens than the H100; on AMD's own image with AITER kernels it served 52% more on Qwen3-8B and 31% more on Qwen3-Coder-30B. vLLM's recommended H100 settings changed nothing. Per million tokens, the cheapest result came from a $1.50-an-hour spot H100.

AMD MI350XH100ROCm
3 min read

What a prompt cache is worth: the same H100 served 82% more requests at a 90% cache hit rate

We served Qwen3-30B-A3B-Instruct-2507 FP8 on our production H100 with vLLM and repeated the same latency-target search with 0%, 50%, 75% and 90% of every prompt drawn from shared prefixes. vLLM's own counters showed cache hit rates of 0%, 49%, 74% and 89%. Capacity rose from 11.5 to 21.0 requests per second, and revenue per GPU-hour from $10.37 to $18.84 at flat prices. With cached input billed at a quarter of the price, the node still earns more and the customer pays far less.

prefix cachingprompt cachevLLM