Meta to launch AI chip in September amid GPU shortage
Meta will start making its next-gen AI chip in September, teaming with Broadcom and TSMC to cut GPU costs and boost its recommendation system performance.
Key Takeaways
Meta is preparing to begin production of its newest AI‑specific chip in September, aiming to lower its reliance on external GPUs.
The company is working with Broadcom on chip design and will have TSMC manufacture the silicon. It also plans to source RAM from Samsung, storage from Sandisk, and fiber‑optic equipment from Sumitomo Electric.
The chips belong to the Meta Training and Inference Accelerator (MTIA) program, which uses a modular chiplet approach. Four generations were outlined in March, and some are already in deployment.
These MTIA chips will power ranking and recommendation algorithms, broader AI workloads, and inference for Meta’s apps, helping the firm save on Nvidia and AMD GPU purchases while still buying from them.
Meta has been investing heavily in compute capacity, planning to deploy 7 GW of compute this year and double it next year, and has signed deals with ARM, AMD, and Amazon for additional resources.
The internal memo cites that at least one chip cleared testing in roughly six weeks, and the modular design is intended to adapt as AI workloads evolve, using chiplets that can be updated more quickly.
Potential Impact Areas
- Enterprises can expect more cost‑effective AI inference, lowering deployment expenses.
- Start‑ups may gain access to custom silicon, fostering innovation in niche AI applications.
- Developers could benefit from tighter integration of AI models with Meta’s platforms, enabling richer features.
- Industry competition may accelerate as more firms design proprietary chips, potentially easing the GPU supply crunch.
Our Insight
Meta’s push to produce its own AI chip reflects a broader trend of large platforms seeking to reduce dependence on external GPU vendors.
By partnering with Broadcom and TSMC, the company can tailor hardware to its specific ranking and recommendation workloads, potentially achieving better performance per watt.
The modular chiplet strategy allows rapid iteration as AI models evolve, giving Meta flexibility to upgrade without full redesigns.
However, the effort requires significant capital investment and engineering expertise, and the article notes that Meta will still purchase GPUs from Nvidia and AMD, limiting immediate cost savings.
For the industry, widespread custom silicon could increase competition, spur innovation in chip architecture, and possibly ease supply constraints, but it also raises barriers for smaller players lacking similar resources.
Overall, the initiative underscores the strategic importance of hardware control in the AI race, while highlighting the trade‑off between investment intensity and tangible economic benefit.
External Credit
Original source: techcrunch.com
Full credit goes to the original publisher. We link to this content for informational and commentary purposes only.