Token arbitrage and the neocloud squeeze

Sources: @deedydas on inference-provider economics; The Information, A Neocloud's Gathering Storm and Anthropic's Margin is OpenAI's Opportunity; Arena Agent leaderboard · raw: raw/twitter/feed/2026-10-25-morning-ranked.json, raw/rss/2026-10-24-the-information-*.md Topic: compute infrastructure, inference pricing

TL;DR

Three views of the same market on one day. Token resellers make money: renting B300s on demand at about $8/GPU-hour, open models (GLM 5.3, Kimi K3, DeepSeek 4.1) can earn $10-20/GPU-hour at ~100 tok/s, and DeepInfer reports $38/GPU-hour off-peak. Margin depends on utilization and stack quality; gateways (OpenRouter, Vercel) aggregate demand, Chinese labs now ask large providers for revenue share for day-0 access, and dozens of providers launch weekly. Deedy's prediction: commodity service pushes compute prices up and token prices down. Compute owners with debt suffer: CoreWeave stock nearly halved in 12 days on rate fears and debt load; Meta's new share class to raise AI capital pulled the Nasdaq down 2.6% and CoreWeave 10%, Nebius 7.2%, IREN 8.3%. SpaceX now sells excess compute, a well-funded competitor. Model vendors compete on price: Rippling moved from 57% Anthropic / 16% OpenAI of coding work in May to 52% OpenAI / 18% Anthropic, citing cost.

Relation to prior wiki pages

• Extends the 10-24 "memory sets the rent" item (H200 rent +160% vs H100 +40%): token prices fall while scarce-memory compute gets dearer, so the margin sits with whoever has cheap, memory-rich capacity and a good serving stack.
• Ties to research: every efficiency paper that raises tokens per GPU-hour (MOSAIC's 4x cache cut, LIRQuant's 3.5x memory saving, speculative decoding) moves a provider along deedy's throughput curve. The arbitrage is a stack-quality business.
• Kimi K3.1 Pro enters Arena's agent leaderboard at #11, top open model, $1.03 median cost per task (13% below K3), just off the Pareto frontier.