Meta is spending tens of billions on computing power for AI, but it wants to spend less on Nvidia's expensive chips. The social media giant is building its own hardware instead. That plan now has a firm start date.
Meta currently buys graphics processing units, or GPUs, from Nvidia and AMD. GPUs are specialized chips that excel at the math behind AI. But an unprecedented shortage of these components has driven up costs and made supply unpredictable.
To solve this, Meta began designing its own chips in 2023 under a program called MTIA (Meta Training and Inference Accelerator). In March 2026, the company detailed four new MTIA chips. These chips are built for three jobs: training AI models, running recommendation and ranking algorithms, and doing inference - the process of using a trained model to make predictions for its applications.
The company expects capital expenditures - money spent on long-term assets like data centers and equipment - to reach $125 billion to $145 billion in 2026. Much of that goes to AI infrastructure. Meta has also signed multibillion-dollar deals with AMD to purchase its Instinct GPUs and with Amazon to use the cloud giant's homegrown CPUs for AI-related needs. But the long-term goal is to rely less on outside suppliers.
Get the market news that matters in a five-minute read with Market Briefs, our free daily newsletter
Meta did not design these chips alone. Broadcom, a chip design company, helped create them. Taiwan Semiconductor Manufacturing Company, or TSMC, will manufacture the chips. Meta also lined up suppliers for other components: Samsung provides RAM, Sandisk provides storage, and Sumitomo Electric provides fiber-optic equipment.
One of the new chips passed its testing phase in just six weeks - a fast turnaround in the chip world. Meta already has a deal with ARM, signed last year, to secure compute power for its recommendation systems. The company explained its approach in a statement from March: "Each MTIA generation builds on the last, using modular chiplets, incorporating the latest AI workload insights and hardware technologies, and deploying on a shorter cadence."
In 2026, Meta intends to bring 7 gigawatts of compute capacity online. A gigawatt is a measure of power - enough to run hundreds of thousands of homes. Next year, it plans to double that. That is a huge jump in computing muscle, all to train and run AI across Facebook, Instagram, WhatsApp, and other apps.
Meta is not alone in its effort to reduce the flow of capital toward Nvidia. OpenAI recently revealed it is building an inference processor with Broadcom. Anthropic is reportedly exploring the possibility of making its own chips in partnership with Samsung. Amazon and Google each design their own processors for AI training and inference.
These moves show a growing trend: big AI companies want their own silicon. Nvidia's GPUs are powerful but expensive and hard to get. By owning the chip design, companies can tailor hardware to their specific needs and adapt faster as AI models evolve.
Reuters reported this story citing an internal memo at Meta. The memo confirmed the September 2026 production start and detailed the company's spending plans.
Meta will use its MTIA chips for both training and running AI models across all its apps. The company has already spent tens of billions of dollars on computing capacity for AI. The new chips are designed to reduce that spending over time - and to give Meta more control over its own technology. For now, the race to build custom silicon is on, and Meta has a clear target.
Join Market Briefs, our free daily newsletter, for a quick daily rundown of the markets
