Reuters just reported Meta has started testing its first homemade chip built to train artificial intelligence systems. Two people who work at the company leaked this news, saying Meta had already sent out small batches of early test chips. The Facebook parent company might make more of these chips if the first ones work right. Meta recently showed off a system using NVIDIA Blackwell GB200 chips, but company leaders apparently want their chips instead.
Many big artificial intelligence companies want to stop completely depending on NVIDIA for chips. Last month, news came out that OpenAI finished designing its chips, possibly working with Broadcom and TSMC to make them. One person who talked to Reuters said Meta asked TSMC, a chip factory in Taiwan, to produce its test chips. Tom's Hardware wrote that Meta teamed up with Broadcom to create its first AI training chip.
Meta has been working on chips called Meta Training and Inference Accelerator for a couple of years already. Reuters mentioned that the project started shakily, and the company even threw away an earlier chip design when it reached a similar testing stage. Last year, Meta began using an MTIA chip for what experts call inference—that means running AI systems when users interact with them, like deciding what posts show up on Facebook and Instagram feeds.
The company wants to have its custom chips ready for AI training by 2026. Earlier versions of MTIA chips used open-source RISC-V processor cores for running AI systems, but nobody knows if the new chips still use this design. Building their chips helps Meta control costs and avoid waiting for NVIDIA chips that everyone wants right now. Many tech giants see making their chips as important for staying competitive as AI keeps changing how we use computers.
Many big artificial intelligence companies want to stop completely depending on NVIDIA for chips. Last month, news came out that OpenAI finished designing its chips, possibly working with Broadcom and TSMC to make them. One person who talked to Reuters said Meta asked TSMC, a chip factory in Taiwan, to produce its test chips. Tom's Hardware wrote that Meta teamed up with Broadcom to create its first AI training chip.
Meta has been working on chips called Meta Training and Inference Accelerator for a couple of years already. Reuters mentioned that the project started shakily, and the company even threw away an earlier chip design when it reached a similar testing stage. Last year, Meta began using an MTIA chip for what experts call inference—that means running AI systems when users interact with them, like deciding what posts show up on Facebook and Instagram feeds.
The company wants to have its custom chips ready for AI training by 2026. Earlier versions of MTIA chips used open-source RISC-V processor cores for running AI systems, but nobody knows if the new chips still use this design. Building their chips helps Meta control costs and avoid waiting for NVIDIA chips that everyone wants right now. Many tech giants see making their chips as important for staying competitive as AI keeps changing how we use computers.