Meta to start producing its own AI chips in September 2026 to reduce reliance on Nvidia

Meta Platforms (designated an extremist organization; its activities are banned in Russia) expects to begin producing its own AI chip as early as September 2026, Reuters reports, citing an internal company memo.

Why Meta wants its own chips

Like most major tech players, Meta has for years relied almost entirely on Nvidia GPUs to train and run its AI models. An in-house chip should reduce that dependence and give Meta more control over the cost and availability of compute, which remains the biggest bottleneck for the entire AI industry. Meta unveiled the first chips in its MTIA (Meta Training and Inference Accelerator) line back in 2023, but used them mainly for inference rather than for training truly large models.

• Production of the in-house chip is scheduled to start in September 2026
• The goal is to reduce dependence on Nvidia chips
• The company intends to scale compute capacity to 14 GW by 2027
• The source of the information is an internal Meta memo obtained by Reuters

How much capacity does Meta need?

The 14 gigawatts of compute capacity Meta expects to have next year is an order of magnitude more than most existing data centers consume. For comparison, a single gigawatt can power hundreds of thousands of homes. The scale reflects Meta's continued investment in training ever-larger Llama models and in infrastructure for its own AI products inside Facebook (the social network is banned in Russia) and Instagram (the service is banned in Russia). The growing power consumption of AI companies' data centers is now visible beyond the tech industry: building new capacity requires not just chips but also long-term electricity supply contracts, including deals with nuclear and gas power plants.

Why this is a strategic issue for the whole industry

Demand for Nvidia chips outstrips supply so heavily that the largest buyers have spent years waiting in line for shipments and have had to accept whatever prices Nvidia sets. In effect, the company has become the main bottleneck for the entire AI industry. Big customers building their own chips doesn't push Nvidia out of the market entirely: training the largest models still requires top-tier GPUs, while in-house designs more often handle narrower workloads such as inference. That means the models' actual responses to user queries, which run billions of times a day and cost less on specialized hardware.

What it means

In-house AI chips let big tech companies depend less on Nvidia and its pricing. Google is already on a similar path with its TPU chips, as is Amazon with Trainium, and Meta's entry into this market further intensifies competition for component supplies and data center capacity. For Nvidia, it's a signal that even its biggest customers want to diversify their compute suppliers. For the market as a whole, it shows that the race for custom AI silicon is becoming the norm for every industry giant, not the exception.