What happened
Meta is set to begin production of its latest in-house AI chips in September. According to an internal memo cited by Reuters and reported by TechCrunch, the move aims to reduce the company's dependency on expensive GPUs.
The chips are part of the Meta Training and Inference Accelerator (MTIA) programme. Meta is designing them with Broadcom and will use TSMC for manufacturing. They are intended for training ranking and recommendation models, as well as broader AI workloads.
How the room's reading it
The consensus is that this isn't just about Meta — it's part of a broader trend. Infra teams see this as another major player trying to escape Nvidia's orbit, following Amazon, Google, and OpenAI's own custom silicon efforts. The move is framed as a long-term strategy to control spiralling compute costs, especially for inference on recommendation and ranking algorithms which are Meta's bread and butter.
However, no one expects Meta to stop buying Nvidia hardware anytime soon. The memo itself confirms massive continued spending. Observers on X note this is about diversification and cost optimisation for specific workloads, not a full-on replacement for general-purpose GPUs.
Sailfish's take
We see this as an internal optimisation play, not a market disruption for external builders — at least not yet. Meta is building these chips to run its own massive-scale recommendation engines more cheaply. They aren't about to start selling MTIA cloud instances to compete with AWS or GCP.
The real signal here is that the hyperscalers have decided the current GPU market is unsustainable for their core workloads. For builders, this means the future of compute won't be a single hardware monoculture. We think this fragmentation is a good thing long-term, but it will demand more sophisticated infrastructure choices. The useful question isn't whether Meta can beat Nvidia; it's what specific workloads run best on which custom silicon.