Xinzhiyuan · WeChatMay 29, 2026, 13:15 UTC+8

Three DeepSeek Models Enter OpenRouter Monthly Top 10 With Over 17 Trillion Tokens

DeepSeek狂吞17万亿Tokens？国产AI算力，最关键一战来了！

DeepSeek placed three models in OpenRouter’s monthly top 10 with more than 17 trillion tokens combined, including V4 Flash at 9.13T tokens; the article says Ascend’s MegaMoE operator raised Prefill throughput by 20% to 30% on DeepSeek V3.1 and Qwen3-235B tests.

Why it matters: HKR-H/K/R all pass: the story has a 17T-token hook plus concrete OpenRouter and MegaMoE Prefill numbers. It stays at 82 because the compute-sovereignty framing is strong, while reproducible test conditions are not disclosed.

Read the original ↗Export Markdown

Take

The useful signal here is not the “domestic compute battle” framing; it is OpenRouter showing agent traffic smashing inference infrastructure. DeepSeek V4 Flash hit 9.13T tokens and ranked No. 1, V3.2 hit 4.07T at No. 8, and V4 Pro hit 3.89T at No. 9. Hermes Agent and OpenClaw posted 10.8T and 6.25T tokens, so tool loops are eating the budget, not chat UX.

The Ascend part needs a colder read. MegaMoE fuses Alltoall dispatch/combine, GMM, and Swiglu into one operator. On Atlas 800 A3, the article claims 20–30% Prefill gains for DeepSeek V3.1 and Qwen3-235B, plus 10%+ Decode gains. That is a real systems hook. But it gives no end-to-end latency, cost per token, concurrency setup, or same-condition H100/H200 comparison. Without those, “general AI infrastructure platform” is still vendor copy.

Source: Xinzhiyuan · WeChat · mp.weixin.qq.com
Published: May 29, 2026, 13:15 UTC+8 · May 29, 2026