Meta Begins Production of Custom AI Chips in September, Targeting GPU Cost Reduction
Key Takeaways
- ▸Meta begins manufacturing custom MTIA AI chips in September, with one design passing testing in six weeks
- ▸The company plans to use custom chips for training and inference workloads to reduce GPU spending from NVIDIA and AMD
- ▸Meta's modular chiplet design allows rapid iteration as AI workloads evolve, with plans to deploy 7 gigawatts of compute in 2026
Summary
Meta is set to begin production of its latest artificial intelligence chips in September as part of its broader strategy to reduce dependence on expensive GPUs from NVIDIA and AMD. The company's Meta Training and Inference Accelerator (MTIA) program has developed four new chips that have reportedly passed testing in approximately six weeks. Working with Broadcom on design and TSMC for manufacturing, Meta will source RAM from Samsung, storage from SanDisk, and fiber-optic equipment from Sumitomo Electric.
The modular chiplet approach allows Meta to adapt the chips as AI workloads evolve, with each MTIA generation building incrementally on the last. The chips are intended for training models for ranking and recommendation algorithms, broader AI workloads, and inference for Meta's applications. This move is part of Meta's massive investment in compute capacity, with the company projecting capital expenditures between $125 billion and $145 billion this year, much of it directed toward AI infrastructure and power deals worldwide.
Meta is not alone in this effort—other major tech companies including OpenAI, Anthropic, Amazon, and Google are also developing proprietary AI chips to manage the soaring costs of compute. The trend signals a broader industry pivot away from exclusive reliance on commodity GPU suppliers, reshaping the competitive landscape for semiconductor companies.
- Major tech companies including OpenAI, Anthropic, Amazon, and Google are similarly developing proprietary chips to reduce dependency on GPU makers
Editorial Opinion
Meta's shift toward custom AI chips represents a critical inflection point in the AI infrastructure wars. While the company will remain a major buyer from NVIDIA and AMD, developing specialized hardware signals that hyperscalers can no longer tolerate the cost and dependency constraints of the commodity GPU market. The modular architecture is particularly smart—it hedges against rapid changes in AI workloads while keeping manufacturing timelines competitive. If Meta's chips deliver expected cost savings at scale, we should expect this pattern to accelerate across the industry, potentially reshaping the semiconductor landscape for the next decade.
