Meta is set to start production of its latest AI-focused chips in September, marking a significant step in its strategy to lower reliance on external GPU suppliers and scale its AI operations.
- Modular AI chip design starts production in September
- Chips reduce Meta's dependence on GPUs from Nvidia and AMD
- Part of a $125 billion+ capital expenditure focused on AI
What happened
Meta announced it will begin producing new AI-specific chips in September as part of its Meta Training and Inference Accelerator (MTIA) program. The chips are designed with a modular chiplet architecture to adapt to evolving AI needs. This production will be handled by Taiwan Semiconductor Manufacturing Company (TSMC), while Broadcom co-designed the chips and other components come from Samsung, Sandisk, and Sumitomo Electric.
These new chips follow successful testing phases and aim to serve a variety of AI workloads such as model training for ranking and recommendation algorithms as well as inference tasks. Some versions of these chips are already deployed or will be deployed soon across Meta’s extensive AI infrastructure.
Why it matters
The introduction of Meta’s proprietary AI chips is a strategic move to reduce the company’s dependency on external GPU suppliers, like Nvidia and AMD, especially amidst component shortages and rising costs. By developing custom hardware optimized for its AI workloads, Meta can better control performance, supply, and costs in its vast AI compute operations.
This initiative aligns with Meta’s broader commitment to AI, reflected in massive capital expenditures estimated between $125 billion and $145 billion this year. These investments fund data centers, power deals, and compute capacity important for training and deploying AI models such as its Muse Spark series.
What to watch next
Industry observers should monitor how Meta’s modular chip architecture performs in production environments and whether these chips significantly reduce reliance on Nvidia and AMD GPUs going forward. The pace of future MTIA chip generations, which Meta plans to develop on a faster cadence, will also indicate how the company adapts to AI’s rapid evolution.
Additionally, watching Meta’s compute capacity milestones—targeting 7 gigawatts this year and double next year—as well as impacts on costs and AI performance, will reveal the broader implications for AI hardware development. Competitor moves in building custom AI chips, including those by OpenAI, Anthropic, Amazon, and Google, will further shape the competitive AI chip market landscape.