Meta announces four new in-house AI chips

Look, every major tech company building custom silicon isn't news anymore — Google's had TPUs for a decade. What's different here is the scale: Meta's burning through H100s at a rate that apparently makes it worth shipping their own hardware. That's a serious pain point.

MTIA 300-500 series. The real question nobody's asking is whether these are inference chips or training chips — because that determines whether this eventually matters to us as API consumers or stays internal.

Here's the thing about compute pricing: it doesn't drop because someone announces competing silicon. It drops when that silicon ships at scale and the hyperscalers actually pass savings downstream. Mass deployment by end of 2027 means we're 18+ months out from any real pricing pressure on APIs we use.

Honestly? NVIDIA's stock dipped on the news, but they're not sweating. One customer vertically integrating doesn't kill a monopoly. But four major customers all cutting custom chips simultaneously? That's a different conversation over a 5-year horizon.

Watch Llama API pricing in late 2027. That's the actual signal — not the chip announcement.