Key Takeaways

  • Meta will start producing new AI chips in September.
  • The chips are designed to reduce reliance on external GPU suppliers.
  • Production involves partnerships with Broadcom and TSMC.
  • Meta plans significant capital investment in AI infrastructure.

Production Timeline

Meta is gearing up to launch production of its latest AI chips in September, as reported in an internal memo. This move comes as the company seeks to cut down on GPU expenses amid a significant shortage of components.

Chip Development

According to the memo, at least one of the new chips successfully completed its testing phase in about six weeks. Meta is collaborating with Broadcom on the chip design, while Taiwan Semiconductor Manufacturing Company (TSMC) will handle the manufacturing. Additionally, the company is sourcing RAM from Samsung, storage solutions from Sandisk, and fiber-optic equipment from Sumitomo Electric.

Modular Approach

In March, Meta outlined its four new chips developed under the Meta Training and Inference Accelerator (MTIA) program. Some of these chips are already being deployed, with others expected to follow this year or next. The design approach is modular, allowing for adjustments as AI technology continues to evolve.

Cost Savings and AI Applications

The MTIA chips are anticipated to help Meta save on purchasing GPUs from manufacturers like Nvidia and AMD, although the company still expects to incur significant costs with these suppliers. The chips will be utilized for training models related to ranking and recommendation algorithms, as well as for broader AI workloads and inference tasks.

Investment in AI Infrastructure

Meta has been heavily investing in compute capacity to support its AI initiatives. In April, the company projected capital expenditures between $125 billion and $145 billion for the year, with a considerable portion allocated to AI development. The firm has been securing data center and power agreements globally, committing tens of billions to ensure sufficient computing resources for its new Muse Spark AI models. This year, Meta aims to deploy 7 gigawatts of compute power, with plans to double that in the following year.

Industry Competition

Meta is not alone in its efforts to reduce dependency on Nvidia. OpenAI recently introduced an inference processor in collaboration with Broadcom, while Anthropic is reportedly exploring its own chip development with Samsung. Both Amazon and Google have also been developing proprietary chips for AI training and inference, alongside numerous startups entering the market to meet the growing demand.

Company Response

Meta has chosen not to comment on the details surrounding its chip production plans.