Meta is starting production of its own custom AI chips in September, moving ahead of schedule in a plan to reduce how much it pays Nvidia and AMD for processing power.
These chips, built under a program called MTIA, are not general-purpose. They are purpose-built to run one category of task: deciding what content and ads billions of people see every day across Facebook, Instagram, and WhatsApp. Every time someone opens their feed, AI processes thousands of signals in milliseconds to rank what appears. That process, called inference, runs billions of times a day, and it is expensive.
Meta already has hundreds of thousands of an earlier chip generation deployed. The September production run launches the next wave, designed to handle more complex AI models at lower cost per operation. Purpose-built chips can run 30 to 50 percent cheaper per operation than buying Nvidia's hardware at standard commercial rates, and Meta claims its new MTIA 400 chip delivers performance competitive with leading commercial products alongside those cost savings.
The chip is co-developed with Broadcom, the semiconductor design firm, and manufactured by Taiwan's TSMC. Meta has locked in a partnership with Broadcom that runs through 2029, covering several future chip generations. It is releasing new chip versions roughly every six months, about three times faster than the typical industry pace of one to two years.
The financial pressure behind all this is significant. Meta spent around $72 billion on infrastructure in 2025. For 2026, that number rises to between $125 billion and $145 billion, nearly double, and more than what the company spent in 2024 and 2025 combined. When your supply costs are rising that fast, building your own hardware is a logical response.
Meta is not alone in doing this. Google has had its own custom chips since 2015. Amazon builds its own. OpenAI is working with Broadcom on a chip of its own. The pattern is clear: any company that runs AI at massive scale is trying to reduce how much of its bill goes to Nvidia.
For the time being, Nvidia is not losing the market. Its share of the AI chip market was around 86 percent in 2025. But for the first time, custom chips from cloud providers are growing almost three times faster than Nvidia's GPU shipments in 2026. The pressure is building.
What does this mean in practice for non-tech businesses? Two things. First, Meta's ad-targeting systems will become more capable over time as cheaper infrastructure lets the company run more complex models. Advertisers on Meta's platforms will likely see continued improvements in targeting precision, which tends to push ad prices higher as results improve. Second, for businesses that rely on organic reach on Meta's platforms, better AI means more competition for attention, as the ranking system becomes more sophisticated at deciding what gets seen and what does not.
Meta is essentially building the engine that drives its advertising business in-house, at scale, faster than anyone expected. The companies most affected are the ones whose revenues depend on what Meta's algorithm decides to show next.