DeepSeek V4.1 Flash tops OpenRouter usage, but demand looks small
V4.1 Flash has led OpenRouter since Oct 25 at roughly 30% of token traffic, yet one GB300 rack could serve a day of its traffic; API is $0.15/$0.60 per 1M tokens off-peak.
DeepSeek V4.1 Flash just topped OpenRouter's usage leaderboard, and the demand behind it is smaller than it looks. Here's what you need to know. OpenRouter's rankings show DeepSeek at #1 and #2 by tokens processed. Per the ranking trackers, V4.1 Flash has led since at least Oct 25. Chon Tang says DeepSeek is roughly 30% of all OpenRouter token traffic. Users call the 1017 build the best open-source model right now, and say it is close to the frontier. Then dax (@thdxr) put the demand in context on Oct 28. OpenRouter is the most successful model router. Almost all of its DeepSeek traffic from the day before could be served on a single GB300 rack. That rack is worth $6-7M. Even if real demand were 10x higher, he says, it is still under $100M of compute. Chon Tang puts it at 5-10 GB300 NVL72 racks, or 50-100 racks at 10x growth next year. He asks how that squares with NVIDIA shipping 90k racks next year. The post passed 125K views. People read the numbers differently. One reply says people just aren't buying open-model tokens. Tenobrus estimates open-weight models are under 10% of the overall market, and basically flat over the last year. Another says OpenRouter is not the hyperscaler or enterprise picture, where Anthropic, OpenAI and Google dominate. The model itself is cheap. The deepseek-flash API lists $0.15 per 1M input tokens and $0.60 per 1M output tokens off-peak, double at peak, with cache hits at $0.003. DeepSeek says the Oct 17 build is stronger on benchmarks, faster, and shorter in its outputs than the previous version. Cursor added it earlier today. SGLang and vLLM now claim parity with DeepSeek's DeepInfer engine. Key numbers: - #1 and #2 on OpenRouter usage - About 30% of OpenRouter token traffic - About 1 GB300 rack covers a day of its DeepSeek traffic - $6-7M per rack, under $100M even at 10x - Open-weight models under 10% of the market, per Tenobrus So DeepSeek leads the usage charts, but the compute behind that lead is small.
This post quotes another post: x.com/thdxr/status/2115165365630026178