Abstract charcoal banner with a glass orb and an orange rim light, Massed Compute logo.

Arete-27B: DGX A100, Blackwell, or H100?

Rent gpu_1x_DGX_A100 for Nous Research Arete-27B. It is the least expensive card that ran the exact BF16 weights, and it returns the most tokens per dollar.

GPUNVIDIALLMInference

Rent the DGX A100 for Arete-27B, then step up for a named reason

Step up only for a named reason. Blackwell is the shorter first token. gpu_1x_h100 is the higher throughput. A 48 GB card does not hold these weights, and a different A100 or H100 listing is a different machine.

The numbers

Same checkpoint. Same engine. vLLM v0.27.1, random prompts, 128 tokens in, 128 tokens out. The headline is output token throughput at concurrency 32. Exact BF16, NousResearch/Arete-27B, about 51 GB before the cache. Capture 9 November 2026. Live list rates on 12 November 2026 still match the three rows. Method and screenshots: Arete-27B on GitHub.

The DGX A100 left max_num_seqs at the server default. Blackwell and the H100 crashed at that default and those two rows used --max-num-seqs 64. Concurrency 32 is under 64, so each row is still a full 32-way run. The three rows did use different serve flags. The Qwen3.8-27B Blackwell page left the default in place, so a comparison with that page is indicative.

SKU List $/hr Output tok/s (c32) TTFT median tok/s per $ $/1M output tokens
gpu_1x_DGX_A100 $1.38 613.3 961.0 ms 444.4 $0.625
gpu_1x_pro_6000_blackwell $2.19 625.8 659.6 ms 285.8 $0.972
gpu_1x_h100 $2.73 754.1 568.4 ms 276.2 $1.006

Scatter chart of Arete-27B output tokens per second versus list dollars per hour at concurrency 32. DGX A100 is the value point, Blackwell is the first-token point, and the H100 is the speed point.

On 12 November 2026, gpu_1x_DGX_A100 and gpu_1x_h100 had no open capacity. gpu_1x_pro_6000_blackwell did. That was the check. Confirm current GPU pricing and capacity when you launch.

Value

613.3 tok/s at $1.38/hr is 444.4 tok/s per dollar, and $0.625 per million output tokens. Blackwell adds about 2% more tokens and costs about 59% more per hour. The H100 adds about 23% more tokens and costs about 98% more per hour. Both return fewer tokens per dollar.

Arete-27B output tokens per dollar at concurrency 32. 1× DGX A100 is highest at 444.4, RTX PRO 6000 Blackwell is 285.8, and 1× H100 is 276.2.

Choose the DGX A100 when the job is a queue and 961 ms to the first token is acceptable.

First token

Blackwell cuts median time to first token from 961.0 ms to 659.6 ms. That is 301 ms. Throughput stays close: 625.8 tok/s against 613.3. Tokens per dollar fall from 444.4 to 285.8. The hourly is $2.19.

The H100 goes to 568.4 ms. That is another 91 ms, with the throughput jump in the next section. None of the three cards returns the first token in under half a second on this profile.

Median time to first token for Arete-27B at concurrency 32. DGX A100 is 961.0 ms, Blackwell is 659.6 ms, and the H100 is 568.4 ms. All three are above half a second.

Choose gpu_1x_pro_6000_blackwell when someone is waiting on the first word and 961 ms is the problem. At the 12 November check, Blackwell was the measured card with open capacity.

Speed

gpu_1x_h100 delivered 754.1 tok/s. That is about 23% more than the DGX A100 and about 21% more than Blackwell. Median first token is 568.4 ms, the shortest in the set.

The hourly is $2.73. A million output tokens costs $1.006, against $0.625 on the DGX A100. Tokens per dollar are 276.2, the lowest of the three.

Dollars per million Arete-27B output tokens at concurrency 32. DGX A100 is $0.625, Blackwell is $0.972, and the H100 is $1.006.

This figure is gpu_1x_h100 only. On 12 November 2026 that SKU had no open capacity.

What does not fit

The BF16 weights are about 51 GB before the cache. These 48 GB cards were on the list and were left untested: RTX A6000 at $0.57/hr, RTX 6000 Ada at $0.79/hr, L40 at $0.86/hr, L40S at $0.97/hr. The L40S list rate has been $0.97 since 8 September 2026.

Same name, different box

The 613.3 tok/s result is gpu_1x_DGX_A100 only. gpu_1x_A100_SXM4 is also $1.38/hr and was in stock on 12 November 2026. It was not the measured machine. gpu_1x_a100 is $1.35/hr, 80 GB, and was not launched.

The 754.1 tok/s result is gpu_1x_h100 at $2.73/hr only. gpu_1x_H100_SXM5 is $3.14/hr. It was in stock on 12 November 2026. It did not run this bench.

Blackwell in this set is gpu_1x_pro_6000_blackwell at $2.19/hr.

How to launch

  1. Sign in to the Massed Compute marketplace and open Deploy.
  2. Pick gpu_1x_DGX_A100 for tokens per dollar, gpu_1x_pro_6000_blackwell for the first token, or gpu_1x_h100 for 754.1 tok/s. Read the SKU string. Confirm it is in stock.
  3. Choose an Ubuntu image with NVIDIA drivers. Add your SSH key. Launch.

Pay only for the hours you use. Confirm the live rate on GPU pricing at launch.

Launch the SKU that matches the job

DGX A100 for tokens per dollar. Blackwell when the first token is the constraint. The measured H100 when you need 754.1 tok/s. A 48 GB card does not hold these weights.

Frequently Asked Questions

01Which GPU should I rent for Arete-27B?

gpu_1x_DGX_A100 when you want tokens per dollar. gpu_1x_pro_6000_blackwell when 961 ms to the first token is the problem. gpu_1x_h100 when you need 754.1 tok/s.

02Will a 48 GB GPU run this model?

These BF16 weights need about 51 GB before the cache. A6000, RTX 6000 Ada, L40, and L40S were left untested.

03Is every A100 the card in the table?

gpu_1x_DGX_A100 produced 613.3 tok/s. gpu_1x_A100_SXM4 is the same hourly and was not measured. gpu_1x_a100 at $1.35/hr was not launched.

04Is every H100 the card in the table?

gpu_1x_h100 at $2.73/hr produced 754.1 tok/s. gpu_1x_H100_SXM5 at $3.14/hr did not run this bench.

05Did you quantize the model?

The weights stayed BF16. The KV cache dtype on the serve was fp8. This profile is 128 in and 128 out at concurrency 32. It is the throughput headline, and it is not a single-user chat trace.