Meta begins production of custom AI chips to cut Nvidia dependency

The September production date is the first hard timeline Meta has committed to publicly, which shifts this from a roadmap aspiration to an operational procurement decision with real consequences for Nvidia's forward order book.

This is largely disconnected from recent activity in our archive, as we have no prior coverage of Meta's silicon program or the broader custom-chip race among hyperscalers. The story belongs to a longer arc that includes Google's TPU buildout and Amazon's Trainium investments, both of which established that in-house inference silicon can hit acceptable efficiency thresholds but has historically struggled to match Nvidia on flexibility for novel training workloads. Meta's bet is that its workload mix, heavily weighted toward recommendation and inference at scale, is stable enough to justify the specialization tradeoff. Whether that assumption holds as model architectures continue shifting is the open question the production announcement does not answer.

Watch whether Meta discloses performance-per-watt comparisons against H100 or B200 class hardware within two quarters of the September production start. Silence on benchmarks after chips are in production would suggest the efficiency case is weaker than the supply-independence rationale.