DeepSeek V4.1F shows 5x-7.4x decode edge on Ascend 950DT vs GB200
The cost crossover at 6.8K tokens/s/chip reframes the Ascend-versus-Blackwell debate from raw benchmark losses to workload-dependent economics, suggesting DeepSeek could run its API on a sub-US$1B Blackwell cluster while pivoting post-training to Ascend. Practitioners tracking inference infrastructure or Huawei's CANN ecosystem get a rare concrete per-chip price and throughput comparison. The proposed 128-chip Ascend supernode topology has no Nvidia equivalent, which matters for anyone planning large-scale decode deployments.
A 30 Sep benchmark put Huawei Ascend 950DT decode throughput at 5K tokens at 100 tps with 131K context, versus 25.5K for Nvidia GB200 running DeepSeek V4.1F with DeepInfer — a 5x gap that widens to 7.4x at 230 tps (20K vs 2.7K). The analysis notes DeepSeek stated three weeks ago in DeepEP-Ascend that EP64-128 configurations remain unoptimised, with Huawei firmware fixes planned to reduce HBM contention. DeepSeek now recommends 64-128 GPU supernodes for decode, a topology no Nvidia setup matches. Assuming US$139K per Blackwell GPU and US$37K per 950DT, Blackwell stays more economical in the current regime but loses to Ascend beyond roughly 6.8K tokens/s/chip, such as for RL rollouts.
some thoughts. the current decode max for v4.1F+GB200 is 25.5K at 100 tps and 131K ctxt. 950DT with CANN, as per the Sep 30 benchmark, does 5K. At 230 tps: 20K vs 2.7K. From 5x to 7.4x. But: DeepInfer Nvidia uses EP64-128/DP64-128. 3 weeks ago, DeepSeek explicitly said in DeepEP-Ascend that they have not optimized EP64-128 (see pic 4, "Larger EP sizes and combine remain under optimization; combine has additional local reduction overhead and HBM contention with URMA. Huawei firmware improvements are also planned to reduce this contention") yet and they're working on it with Huawei. Today they recommend "ordinary 8-GPU machines for prefill, ordinary 64–128-GPU supernodes for decode". There is no 64-128 GPU "Nvidia supernodes". It may be they've solved that part at least internally, and that 100TPS case can improve to ≈8-10K in 128-Chip SuperNodes using DeepInfer with high EP. (I ignore the actual mix for simplicity, directionally it'll be similar). In that case, assuming $10M per NVL72 rack (reasonable current smuggling premium, extrapolated from >$1M per 8-GPU DGX B300 server a few months ago) thus $139K per GPU, and $37K per 950DT (rumored), Blackwells likely remain more economical in the actual V4.1F product regime, but clearly lose in the throughput-maximizing one (eg RL rollouts). The crossover is roughly 6.8K tokens/s/chip. Given the efficiency of Blackwell inference and the competitive advantage of V4.1F being that fast (but still making a ton of dumb turns), it'd make sense for DeepSeek to cover the API with a small Blackwell cluster worth <$1B, ie ≈100 racks (it'll pay for itself in well under a year), and fully pivot to Ascend for post-training inference. They will do training of larger future models on Ascend SuperPoDs/SuperClusters, when available.
