Radium racked its own GPUs to undercut OpenAI and Anthropic pricing — small devs still won't switch
No_Raspberry7273 · reddit · 2026-09-10
Radium, a small North American team, bought and racked GPUs to serve tuned open-weight models behind OpenAI- and Anthropic-compatible endpoints, arguing that classification, extraction and formatting tasks shouldn't pay frontier-model prices. The pitch lands with enterprises — several run production workloads with them — but solo devs create API keys and never return. The author rules out trust, switching cost (just a baseurl change) and docs, and asks the community: what's the actual bar for moving workloads to a smaller inference provider, and where has "OpenAI-compatible" broken on you?
More from Venture
- VC Mike Dudas debunks the 'VCs run companies' myth: 'they don't listen to me' — whatsallthiss · 2026-09-10
- No, a $2,000 foldable iPhone won't scare off buyers — demand psychology explained — Linahuaa · 2026-09-10
- Nvidia NVL72 rack shipments seen up 50% in 2027, output forecast to top $710B — Beth_Kindig · 2026-09-10
- Meta's real AI play: becoming the OS for the entire ad funnel for SMBs, analyst argues — rohanpaul_ai · 2026-09-10
- September's model boom: Gemini 3.8 Flash, MuSpark 1.3, and why model choice now matters more — The AI Daily Brief · 2026-09-10
- Jeff Dean, Sanjay Ghemawat, Quoc Le and Oriol Vinyals launch AI science startup Discovery Loop — JeffDean · 2026-09-10