Massed Compute

The dollar rates below are the marketplace list checked on 2026-10-08: A100 PCIe $1.35/hr, A100 SXM4 $1.38/hr, RTX PRO 6000 Blackwell $2.19/hr, 2× A100 $2.70/hr, H100 $2.73/hr, H200 NVL $3.62/hr. L40S is $0.97/hr (it moved from $0.88 to $0.97 on 2026-09-08). Clef and JEV were measured on 2026-10-06. Kolibri was measured on 2026-10-07. Confirm the live rate before you launch. These three models are separate measurements. Different weights, different engines, different units.

Why someone would use it

• Clef, when the product is a label. One forward returns probabilities for typed questions. There is no paragraph to parse. The weights are 51.19 GiB, so plan on an 80 GB card.
• JEV, when one model has to route a ticket and, on another call, write. System 1 returns a probability for every option. System 2 generates tokens. Vision is on the model when the ticket includes an image. The weights are 52.17 GiB.
• Kolibri, when you need a large writer serving many completions at once. About 78B parameters, about 3.5B active, in FP8. One Blackwell held the working set. A single waiting user should read the concurrency-1 rate, not the concurrency-32 headline.

What to launch

Use Clef when you need probabilities

Cloudflare/clef at 2f3de3dd85f379784083b0814d997ab627200f0c is a BF16 multimodal decision model. One forward returns probabilities for typed questions. It does not generate text. The file on disk is 51.19 GiB. A load that succeeded held 51.195 GiB.

We launched an A6000. The load ran out of memory. That card reports 47.40 GiB usable, and PyTorch died at 47.08 GiB allocated. On the 2026-10-08 list that SKU is $0.57/hr. The L40S is the same 48 GB class and was not launched. The least expensive launched card that held Clef is the A100 SXM4.

What we timed is single-stream systemone(), 20 warmups and 200 calls, text first and then the same questions plus one 512×512 image. Decisions/s is 1000 divided by p50. The clock includes building the request. Video inputs and batches above 1 were not run. A wiring check on the sample ticket came back technical, urgency near today, outage likely. That checks the pipe. It is not an accuracy test.

On text, Blackwell is 1.6× the A100 rate. Text value is 9779 vs 9634 decisions per dollar, too close to name a winner. On the image workload Blackwell is 1.7× (4.128 vs 2.461 decisions/s) and higher value (6786 vs 6420 decisions per dollar).

An A100 PCIe at $1.35/hr might fit and was out of stock on the capture day. It still has no capacity on 2026-10-08. H100 ($2.73/hr, 80 GB) was not launched.

Screenshots and the raw capture: Clef on gpu-benchmark.

Use JEV when the model must label or write

autotrust/JEV-27B-VL at f34b598d4ef4bcefd337bee8d8e7ddd3b7733ccc is BF16 Qwen3.8-27B with vision plus a System 1 decision adapter. Safetensors are 52.17 GiB, so a single 48 GB card cannot hold the weights. The model card asks for one GPU with 80 GB or more.

System 1 (POST /v1/decide) returns a probability for every option in one forward. System 2 is the base model on the usual generation endpoint. We ran the author’s server with prefix caching off and a cap of 8 sequences. The text decision is a 3-option ticket route. The image decision is one yes/no question on a generated 448×448 frame. The clock is client wall time per request on the same machine, including JSON and the image upload. Image p50 can understate a feed of new frames, because the multimodal processor cache stayed on. A wiring check returned billing on the text ticket and true on the generated frame. That checks the pipe. It is not an accuracy test.

System 1 text, the headline:

Blackwell is the System 1 speed card and the System 1 value card. At concurrency 8 it is 2.1× the A100 and 2.0× the H100. Image decisions follow the same order: Blackwell 23.276/s at concurrency 8, 1.9× the A100. Single-stream System 1 on the H100 is slower than the A100 (164.4 ms vs 156.4 ms on text, 179.5 ms vs 164.6 ms on the image).

System 2, random 128 input tokens and 128 output tokens. The number is output token throughput at concurrency 8:

The three cards are within 7% of each other on that output rate. The H100 is the throughput high. The A100 is the best output tokens per dollar. The A100 PCIe row has no capacity on 2026-10-08. The A100 SXM4 and DGX A100 at $1.38/hr are in stock and were not the cards we measured.

Two-card A6000 listings at a lower list price were not launched: 2× A6000 at $1.14/hr, low-RAM at $1.10/hr, spot at $1.00/hr. On 2026-10-08 all three have capacity. The author’s setup is one GPU, and the System 1 adapter was not checked under tensor parallel.

Screenshots and the raw capture: JEV-27B-VL on gpu-benchmark.

Use Kolibri when the model has to write

Aleph-Alpha/Kolibri-1 at 35bc4d3be745502227a67247de77d70e691614ee is the FP8 pack, about 78B total parameters and about 3.5B active. Measured safetensors are 73.4 GiB. This pass is that FP8 repo. The BF16 sibling was not timed.

The headline is output token throughput at concurrency 32, from vllm bench serve on random prompts of 128 input and 128 output tokens. Prefix caching is off. The clock includes HTTP. One person waiting sees a much lower rate: concurrency 1 is 139.9 tok/s on Blackwell, 154.7 on the H200, and 93.7 on 2× A100.

Attention and KV dtype are not the same on every card. Blackwell used Triton and fp8 KV. The H200 used FlashAttention and fp8 KV. 2× A100 used Triton, bfloat16 KV, and tensor parallel 2, with custom all-reduce disabled after the peer-to-peer check failed. Read the three rows as separate setups.

The least expensive launched card that held the pack is the Blackwell at $2.19/hr. While the server was up it was using 88.44 GiB. A single 80 GB card was not launched. The page records that a single 80 GB card cannot hold this working set. 2× A100 finished at 667.1 output tok/s and a median time to first token of 1576.6 ms. Each c32 row completed 160 prompts and failed 0.

48 GB cards, including the L40S, were not launched. L40S had no capacity on the capture day. It has capacity on 2026-10-08 and is still 48 GB. H100 NVL ($3.11/hr) and 2× Blackwell ($4.38/hr) were in the inventory and were not launched. 2× A100 has no capacity on 2026-10-08.

Screenshots and the raw capture: Kolibri-1 on gpu-benchmark.

Cards that are in stock and still the wrong floor

On the 2026-10-08 list, RTX PRO 4500 Blackwell is $0.92/hr with 32 GB, and it has capacity. A6000 low-RAM is $0.55/hr and A6000 spot is $0.50/hr. Both are 48 GB and have capacity. Clef already failed on the A6000, which is $0.57/hr on that same list. A30 ($0.35/hr, 24 GB) and A5000 ($0.44/hr, 24 GB) have no capacity. None of those VRAM sizes held these weights.

Launch a GPU after you check the live rate. The dollars above are the 2026-10-08 marketplace list, not a contract.