DeepSeek V4.1 Flash earns $38/GPU-hr vs Kimi's $8.34 — a 4.56x gap reshaping inference
bookwormengr · x · 2026-10-26
SemiAnalysis numbers spark an inference-economics comparison: Kimi 3.1 Pro on AMD MI355X generates $8.34 per GPU-hour vs $38 for DeepSeek V4.1 Flash on GB200 — a 4.56x gap. SemiAnalysis also notes AMD's vLLM serving of open-source Kimi K3 (2.8T params, 288GB/chip) can yield $3/hr profit per chip, while B200 can't fit the model in one server and must parallelize over 400Gbit/s links, 18x slower than NVLink-domain bandwidth.
The poster argues DeepSeek V4.1 Flash's architecture (Engram, Causal Encoder-Decoder/Self Decoder, KV Re-use) is too compelling to ignore: players in the US (SpaceX, Muse) and China (Xiaomi, GLM, Kimi) will have to adopt it, though Helios-rack deployments would improve Kimi K3.1's per-GPU revenue without closing the gap.
More from Infra
- ThousandML Open-Sources CQ 4-bit QLoRA: 17.4% Lower Val Loss Than Unsloth NF4 — textclf · 2026-10-26
- Why OpenAI and Anthropic Cut LLM Prices by Up to 80% in Months — Azer_pubes33 · 2026-10-26
- GPU rental rates split by chip generation: B300 +8.6% in a week, A100 only +2% — AccBalanced · 2026-10-26
- Fireworks serves 40T tokens a day, OpenRouter shows just 40-80B — why OpenRouter isn't the whole inference market — AccBalanced · 2026-10-26
- BeNice! KV-sharing for looped Transformers cuts GPU memory 66%, TTFT latency 83.6% — realDanFu · 2026-10-26
- The Economist: AI compute is the new oil — and exchanges like Compute Exchange are building the market for it — AccBalanced · 2026-10-26