Meta Platforms will begin manufacturing its own artificial intelligence chip in September, according to an internal memo reviewed by Reuters, as the social media giant races to reduce its dependence on Nvidia and double its total computing capacity to 14 gigawatts by next year.
The move marks a significant escalation in Meta's efforts to control its AI hardware stack. The company has spent billions of dollars purchasing Nvidia's graphics processing units to train and run its Llama family of large language models, but executives have grown increasingly concerned about supply constraints, rising costs, and the strategic vulnerability of relying on a single chip supplier.
The custom chip, details of which have not been previously reported, was designed by Meta's in-house silicon team and will be produced by a contract manufacturer. While the company did not disclose the chip's specifications or the foundry partner, people familiar with the matter said it is optimized for inference workloads — the process of running AI models in production — rather than training, which remains dominated by Nvidia's high-end GPUs.
"We are building a computing platform that gives us the flexibility and efficiency to support AI at the scale we need," a Meta spokesperson said in a statement. The company declined to comment on specific production timelines or technical details.
The September production target comes as Meta faces mounting pressure to justify its AI spending. The company has raised its capital expenditure forecast for 2026 to as much as $72 billion, a figure that has made some investors uneasy. Shares of Meta fell sharply in May after the company signaled that AI infrastructure costs would continue to climb, though the stock has since recovered as analysts concluded that the investments would yield long-term returns.
Meta is not alone in pursuing custom silicon. Google has been designing its Tensor Processing Units for nearly a decade, Amazon has developed its Trainium and Inferentia chips, and Microsoft recently unveiled its Maia AI accelerator. The trend reflects a broader realization among technology giants that off-the-shelf hardware, while powerful, cannot match the efficiency of chips purpose-built for specific AI workloads.
For Nvidia, the growing list of customers developing their own chips represents a long-term competitive threat, though analysts note that the company's CUDA software ecosystem and dominant market position make an immediate displacement unlikely. Nvidia's revenue from cloud and AI customers continues to grow even as those same customers invest in alternative hardware.
Meta's chip ambitions also tie into its broader infrastructure strategy. The company announced this week that it will build a C$13 billion data center in Alberta, Canada — its first in the country — and has been expanding facilities in the United States, Europe, and Asia. At 14 gigawatts, Meta's total computing capacity would rival the electricity output of a small country, underscoring the staggering scale of resources being devoted to the AI race.
Image source: i.ibb.co