Meta said in a March 11, 2026 post on the AI at Meta Blog that its Meta Training and Inference Accelerator, or MTIA, program has moved beyond the first public generations and is now advancing through MTIA 300, 400, 450, and 500. According to the company, those four generations are either already deployed or scheduled for deployment in 2026 and 2027, with the goal of serving AI features to billions of users at lower cost.

Meta described MTIA as a family of homegrown AI chips built in close partnership with Broadcom. The company said it has already deployed hundreds of thousands of MTIA chips in production, onboarded internal production models, and tested the platform with large language models such as Llama. The broader message of the post is that Meta no longer wants to wait for a single long hardware cycle. Instead, it is trying to iterate on chiplets, networking, memory, and software fast enough to match changing AI workloads.

In Meta's breakdown, MTIA 300 is already in production for ranking and recommendation training. MTIA 400 extends that foundation toward more general GenAI support and uses a 72-accelerator scale-up domain that Meta said is competitive with leading commercial products. MTIA 450 then shifts more directly toward GenAI inference. Meta said it doubles HBM bandwidth relative to MTIA 400, raises MX4 FLOPS by 75%, and delivers 6x the MX4 FLOPS of FP16/BF16. The company said MTIA 450 is scheduled for mass deployment in early 2027.

MTIA 500 continues that inference-first push. Meta said it adds another 50% increase in HBM bandwidth over MTIA 450, as well as up to 80% more HBM capacity and 43% higher MX4 FLOPS. From MTIA 300 to MTIA 500, Meta said HBM bandwidth rises by 4.5x while compute FLOPS increase by 25x. Those numbers are central to Meta's argument that modern AI models are changing too quickly for a slower two-year hardware cadence built around static workload assumptions.

The software stack is another major part of the strategy. Meta said MTIA is built natively around PyTorch, vLLM, Triton, and Open Compute Project standards so that teams can move models onto the hardware without rewriting them around a proprietary environment. Taken together, the roadmap shows Meta trying to turn MTIA from a cost-saving side project into a long-term AI infrastructure stack optimized first for GenAI inference and then for adjacent workloads.

#hbm

Meta maps MTIA 300-500 roadmap to scale AI services for billions