alt Hacker NewsI don't see how the cost of inferencing existing models can become prohibitive. Multiple providers are selling DeepSeek 4.1 Flash tokens at just above cost, and seem to be doing so profitably.
What we do potentially lose out on are the next few generations of models in the case of a market correction, as the big labs won't be able to afford training anymore.
Those providers don't have the compute to meaningfully undercut the private labs.
The reason it's so cheap is because they have compute sitting idle. Meanwhile the big players are rate limiting, using quotas, and keep prices high to keep enough compute available to meet demand.