DeepSeek‑V4 Inference Cost Showdown: NVIDIA H100 vs Ascend 950PR vs 910C

1. Why DeepSeek‑V4 is demanding

DeepSeek‑V4 has a total parameter count of 1.6 trillion but uses a Mixture‑of‑Experts (MoE) architecture, activating only 49 B parameters per inference. This makes it appear heavyweight yet run efficiently. It also incorporates CSA+HCA hybrid attention, reducing inference FLOPs by 27 % compared with the previous V3.2 generation, cutting KV‑cache memory by 10 %, and using FP4 weights with FP8 computation for lower precision, memory, and compute requirements.

2. The three contenders

NVIDIA H100 – the high‑price powerhouse

FP8 performance: 1979 TFLOPS (the strongest in the set)

Memory: 80 GB HBM3, bandwidth 3.35 TB/s

Price: $30 000–$40 000 per card, hourly rental $3.5

Position: industry leader, ecosystem‑rich, but extremely expensive and hard to procure

Huawei Ascend 910C – the awkward middle

FP8 support: none (explicitly not supported)

FP16 performance: roughly 800 TFLOPS

Memory: 64–96 GB, bandwidth varies

Price: cheap, but fundamentally mismatched with DeepSeek‑V4

Huawei Ascend 950PR – the cost‑effective dark horse

FP8 performance: 1000 TFLOPS (native support)

Memory: 128 GB HiBL 1.0 (largest in the group)

Inter‑connect bandwidth: 2 TB/s (outperforms H100 by two lanes)

Price: $13 700 per card, hourly rental $1.2

Position: domestically produced, optimized for MoE long‑context workloads, price‑cutting

3. Real‑world benchmark: Prefill vs Decode

Prefill (reading the prompt) – compute‑bound

Token throughput:

H100: 70 700 tok/s (FP8 full‑throttle)

950PR: 35 700 tok/s (steady FP8)

910C: 28 600 tok/s (FP16 only)

Cost per million tokens (Prefill):

H100: $1.38

910C: $1.94 (41 % more than H100)

950PR: $0.93 (cheapest)

Decode (generating output) – memory‑bandwidth‑bound

Token throughput:

H100: 3 400 tok/s

950PR: 2 000 tok/s

910C: 1 200–3 600 tok/s (highly variable)

Cost per million tokens (Decode):

H100: $28.6

910C: $30.9 (more expensive)

950PR: $16.7 (42 % cheaper than H100)

Conclusion from the two phases: the 950PR beats both competitors, being ~30 % cheaper in Prefill and ~50 % cheaper in Decode.

4. Scaling to a 10 k QPS inference cluster

H100: requires 3 000 cards, total cost $105 million, annual fee $92 million

910C: requires 4 500 cards, total cost $112 million, annual fee $79 million

950PR: requires 3 500 cards, total cost $48 million, annual fee $37 million

Thus, the 950PR’s procurement cost is roughly half of the H100 and even less than half of the 910C, freeing a substantial budget for other operational expenses.

With DeepSeek‑V4’s API pricing of $12 per million input tokens and $24 per million output tokens, hardware costs on a 950PR deployment account for only 5 %–12 % of the total price, leaving the majority as profit margin.

5. Bottom line

Ascend 950PR: price ≈ 1/3 of H100, performance ≈ 50 %–60 % of H100, 128 GB memory, native FP8, large inter‑connect – the clear cost‑performance champion.

NVIDIA H100: unmatched raw performance but 1.5 ×–1.7 × the cost of 950PR, subject to export controls and supply risk; suited for well‑funded users who must stay in the CUDA ecosystem.

Ascend 910C: lacks FP8, slower, and ends up the most expensive per token; best relegated to training or non‑V4 workloads.

Overall, for DeepSeek‑V4 inference the Ascend 950PR provides the best value, delivering sufficient performance at a fraction of the cost of the high‑end H100.