CrofAI shutdown probe: Kimi K3 requests secretly routed to cheaper models
richdotca · x · 2026-09-29
CrofAI went from "cheapest inference provider in the world" to shutdown in two weeks after an independent investigation found requests for models like Kimi K3 were routed via OpenRouter to smaller, cheaper models.
The post proposes three questions to ask any hosted inference provider: which model actually answered, who served it, and who can read what you sent. The reblogger's own SayGm gateway discloses its cross-provider routing upfront and runs inside a hardware-attested TEE (Intel TDX) so the handling code can be verified rather than trusted.
More from Infra
- Justine Tunney: from Occupy Wall Street infra to Cosmopolitan Libc and llamafile — bigaiguy · 2026-09-29
- Qualcomm CEO: global token demand to hit 1.27T per 10 seconds by 2030, a 40x jump — rohanpaul_ai · 2026-09-29
- Is one fine-tuned small model per tenant a viable alternative to RAG? — alexrada · 2026-09-29
- xAI Built a 670K-GPU Full Stack in Just 3 Years, From Colossus to Grok 4.7 — FinanceYF5 · 2026-09-29
- The AI compute fight: it's about developer ecosystems, not just chips — Thionne_WTZ · 2026-09-29
- Solo dev weighs $5k-$20k local AI upgrade: dual RTX 3090 rig vs clustered DGX Spark GB10 — ChopSticksPlease · 2026-09-29