Meta to Put Its Own AI Chip into Production in September, Aiming to Double Computing Capacity
Meta plans to put its own artificial intelligence chip into production in September and is aiming to roughly double the computing capacity across its data centers, according to Reuters reports on Thursday.
Background
- Chip Details: The chip belongs to Meta’s in-house silicon line, the Meta Training and Inference Accelerator (MTIA).
- Manufacturing: Produced by TSMC and co-developed with Broadcom.
- Partnership: Extends through 2029, covering several generations of custom silicon.
Objectives
- Reducing Reliance: Part of a wider effort to reduce Meta’s reliance on Nvidia.
- Capacity Target: Doubling computing capacity in data centers.
- Timeline: September production, with chips deployed over 2026 and 2027.
Strategic Logic
- Cost Savings: Every workload shifted to in-house chips reduces Meta’s dependence on Nvidia margins.
- Workloads: MTIA has mainly handled inference so far, but a training-capable chip would be a significant advancement.
Spending Plan
- Capital Expenditure (CapEx): Projected $125bn to $145bn in 2026, focusing on data centers, GPUs, and custom silicon.
- Mark Zuckerberg’s Vision: Ultimate targets measured in gigawatts.
- Spare Compute: Meta is exploring renting out excess compute power to external customers.
- Diversification: Hedging bets with Amazon’s Graviton5 chips and AMD accelerators.
Conclusion
Custom silicon, especially designed for Meta’s specific models, offers cost savings through reduced power draw and unit costs. However, it is unlikely to replace Nvidia entirely in the near future, as in-house chips typically compete within specific niches.