Meta says it plans to start deploying MTIA 450, its third-generation in-house AI chip, in data centers during H1 2027, followed by MTIA 500 at the end of 2027
Context & Ripple Effects
Meta’s accelerator effort has progressed from a 2023 plan for chips serving AI training and inference to a second-generation data-center deployment alongside commercially available GPUs in 2024. By March 2026, Meta said its MTIA 300 was already in production for content ranking as part of a four-chip roadmap through 2027.
The new timetable turns that roadmap into an explicit deployment sequence for later-generation silicon. It matters because Meta is committing its data-center architecture and software operations to successive proprietary accelerators rather than treating the program as a one-off chip release.
First-order effects
- Meta schedules MTIA 450 for data-center deployment in the first half of 2027 and MTIA 500 for the end of 2027, giving its infrastructure teams a defined transition path beyond the MTIA 300.
- Meta’s AI systems will need to support successive MTIA generations while retaining the mixed proprietary-and-commercial-GPU approach it described for its earlier deployment.
Second-order effects
- The MTIA roadmap makes workload placement and software compatibility a recurring operating task for Meta: gains from new silicon depend on applications moving cleanly across accelerator generations.
- A sustained internal-chip cadence gives Meta a stronger basis to tailor data-center purchasing and capacity planning around its own hardware roadmap, rather than a single external accelerator configuration.
Third-order effects
- If Meta executes the stated sequence, it reinforces an integrated AI-stack model in which a large platform company designs compute around its own workloads and deploys it over multiple chip generations.
- The broader structural shift is toward heterogeneous AI infrastructure, where proprietary accelerators coexist with commercial GPUs and software portability becomes a central constraint.
The trend: Large AI platforms are extending proprietary-silicon roadmaps from isolated deployments into multi-generation data-center strategies built around heterogeneous compute.