Reuters reported, citing an internal memo, that Meta plans to start manufacturing the latest version of its AI-specific chip in September to lower the cost of GPUs amid an unprecedented parts shortage.
At least one chip passed the testing phase in about six weeks, according to the memo. Meta is collaborating with Broadcom on chip design, but will use Taiwan Semiconductor Manufacturing Company (TSMC) for manufacturing. According to the report, it also buys RAM from Samsung, storage from SanDisk, and fiber optic equipment from Sumitomo Electric.
In March, Meta detailed four new chips developed under the Meta Training and Inference Accelerator (MTIA) program. Some of them are currently in place or will be in place this year or next. The company is taking a modular approach to the design of these chips, anticipating that needs will change as AI rapidly evolves by the time the chips reach mass production.
“Each MTIA generation builds on the latest generation by using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying them in shorter cadences,” the company wrote at the time.
These chips are expected to help the company save on the cost of purchasing GPUs from chipmakers like Nvidia and AMD, but the company still expects to spend significant money on these providers as well, Reuters reports. Meta plans to use MTIA chips to train models for ranking and recommendation algorithms, broader AI workloads, and inference targeting applications. The social media company has been producing its own AI chips since 2023.
Meta has spent significant amounts of money securing sufficient computing power to power its various AI efforts. The company said in April that it expects capital spending to be between $125 billion and $145 billion this year, much of which will go toward AI efforts.
The company has data center and power contracts around the world and is spending tens of billions of dollars to secure the computing power to train and deploy its new Muse Spark series of AI models. Reuters, which cited the memo, said it plans to deploy 7 gigawatts of computing power this year and double that next year.
The company also signed a multibillion-dollar deal with AMD last year for Instinct GPUs, a multibillion-dollar deal with Amazon to use the cloud giant’s in-house CPUs for AI-related needs, and a deal with ARM to secure compute for its recommendation systems.
Meta isn’t the only company trying to stem the flow of capital to Nvidia. Last month, OpenAI announced an inference processor it was developing with Broadcom, and Anthropic is said to be considering developing its own chip with Samsung. Amazon and Google have both developed their own chips for AI training and inference, and a number of startups are building in this space to meet surging demand.
Mehta declined to comment.
If you buy through links in our articles, we may earn a small commission. This does not affect editorial independence.
