Vercel lists DeepSeek V4.1 Flash 1017 at 1,000+ tok/s, but provider speed numbers don't add up
gajesh · x · 2026-11-08
Vercel announced DeepSeek V4.1 Flash 1017 Ultrafast on AI Gateway, claiming 1,000+ tokens per second via deepseek/deepseek-v4.1-flash-1017-ultrafast, aimed at coding, reasoning, and agents.
Developer Gajesh immediately flagged inconsistent provider math: the dashboard shows Baseten at 984 tps (0.5s latency, $0.60/M in, $2.40/M out) versus Fireworks at 742 tps (0.9s latency, 1M context, 93% cache hit, $9.60/M out) — with the slower-on-paper Fireworks option seemingly marketed as faster. A caution flag for anyone picking providers by headline throughput.
More from Infra
- Astrophysicist claims SpaceX's $2T space-compute plan fails on cooling physics — JOBhakdi · 2026-11-08
- Liquid Cooling to Hit 59% of AI Chips by 2027, a 10x Jump in Four Years — Beth_Kindig · 2026-11-08
- Farms use 2,500x more water than data centers, one chart shows — davidpattersonx · 2026-11-08
- SoftBank's default insurance cost hits 4.66%, SB Energy's $50B IPO delayed — ccerrato147 · 2026-11-08
- Debt now covers 32% of Big Tech's AI capex as neocloud financing strains show — ccerrato147 · 2026-11-08
- Multi-GPU water-cooling parts guide: radiators first, Alphacool/Singularity/Aquacomputer top picks — net_termina · 2026-11-08