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
New processor will be running in data centers next year, with another model to follow by end of 2027
Context & Ripple Effects
Meta’s chip effort has progressed from a 2023 plan for MTIA to train and run AI models, through 2024 production of earlier accelerators, to the March 2026 roadmap for four MTIA generations. MTIA 300 was already in production for content ranking, making the 450 and 500 deployment targets a concrete next step in a workload-specific silicon program.
The schedule gives Meta’s data-center organization a defined 2027 handoff: MTIA 450 in the first half, followed by MTIA 500 by year-end. It extends the strategy outlined when Meta planned to pair its own second-generation chips with commercially available GPUs.
First-order effects
- Meta can plan 2027 data-center capacity and software deployment around MTIA 450 in H1 and a broader MTIA 500 rollout by year-end.
- The MTIA program moves beyond the MTIA 300 content-ranking deployment toward additional in-house AI compute generations.
Second-order effects
- Meta’s infrastructure teams must allocate workloads across successive MTIA generations and commercial GPUs, making accelerator-to-workload fit a central procurement and deployment decision.
- Suppliers of commercial AI accelerators face a customer whose internal silicon roadmap is becoming more explicit for selected data-center workloads.
Third-order effects
- If Meta continues moving recurring AI workloads onto MTIA, large platform operators will increasingly treat custom accelerators as a complement to merchant GPUs rather than a one-off hardware project.
- The competitive advantage in AI infrastructure shifts toward tightly coordinating models, ranking workloads, software, and bespoke chips across data centers.
The trend: Meta’s 2027 schedule is one data point in the shift toward heterogeneous AI compute, in which hyperscalers combine proprietary accelerators with commercial GPUs for different workloads.