The debate centres on whether DeepInfer's reported throughput depends on a tightly controlled, homogeneous GPU fleet rather than general infrastructure design. Engineers planning large-scale inference deployments on GB200 clusters can weigh whether NVIDIA Dynamo plus 3FS achieves comparable results outside DeepSeek's own setup.
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.
The implication is that inference efficiency, not GPU count, is becoming the competitive lever for API revenue among Chinese labs. Builders pricing their own token workloads can use the 155T tokens-per-day figure as a benchmark for what a small rack footprint can sustain. Compared with US labs running far larger fleets, the capital-per-revenue ratio here is unusually lean.
有人拿 Nvidia 2025 年出货的 27.3K 个 GB200 机架跟 DeepSeek 定价算了笔账，结论是 30 万张卡就能赚 1000 亿美元，算力可能严重过剩。
DeepSeek 年化收入据说破 10 亿美元了，而且 3000 块 GPU 就能扛住全部流量，10 个月回本，这账算得挺有意思。
A 7.8x per-node throughput jump means serving large mixture-of-experts models could get much cheaper per token once Rubin systems ship. Teams running inference infrastructure should track vLLM's Rubin maturity before planning their next GPU refresh, especially against GB200-based deployments. There is no ASEAN-specific angle here; the concrete comparison is Rubin versus GB200 in vLLM.
Weight resharding has been a hidden tax on RL-based post-training pipelines, and folding it into a single collective removes traffic duplication that direct sends and gather-then-broadcast both suffer. Teams running large MoE models on multi-node GPU clusters can adopt it to reclaim a meaningful share of step time. The DeepSeek-V3 numbers also give practitioners a concrete baseline for benchmarking their own weight-sync setups against NeMo-RL.
LTX 出了新视频模型，6.8 秒就能出 10 秒 720P 视频，还免费给小团队用，想玩视频生成可以试试。
