Meta unveils four-generation MTIA custom chip roadmap for AI inference
The MTIA 300 chip is already in production for recommendation systems; three further generations through 2027 are aimed chiefly at generative-AI inference, built on PyTorch and OCP tooling.
- Compute & infrastructure
- Minor
Meta set out a roadmap for four generations of its in-house MTIA (Meta Training and Inference Accelerator) chips, to be developed and deployed within roughly two years — a markedly faster cadence than the industry’s typical multi-year chip cycle. The MTIA 300 generation, designed for training and running the ranking and recommendation systems behind Meta’s ad and feed products, is already in production. The following three generations, numbered 400, 450 and 500, are described as capable of handling a wider range of workloads but are aimed primarily at running generative-AI inference through 2027.
Meta said it intended to release new chip generations roughly every six months or faster by reusing modular design components across generations, rather than redesigning silicon from scratch each cycle. The company also emphasised compatibility with widely used open tooling — PyTorch, vLLM, Triton and the Open Compute Project — as a way of lowering the cost of adopting its custom chips internally rather than requiring a bespoke software stack.
The announcement positioned custom silicon as one part of a broader infrastructure strategy that continues to mix Meta’s own chips with GPUs bought from external suppliers, reflecting a wider trend among the largest AI developers — also visible at Google, Amazon and Microsoft — toward building proprietary inference hardware to reduce dependence on, and cost relative to, general-purpose GPUs from Nvidia.