Meta unveils four-chip MTIA inference roadmap, but is not dropping Nvidia and AMD

Meta has revealed its development plan for its in-house MTIA AI accelerators: over the next two years, the company wants to roll out four chip generations for ranking, recommendations and generative AI. It will not be able to abandon Nvidia and AMD entirely, though. Meta treats its own silicon as a complement to purchases of third-party hardware, not an instant replacement.

Why Meta wants its own chips

The main reason is simple: the load on Meta's infrastructure is growing too fast, and general-purpose GPUs keep getting more expensive. The company is expanding data-center capacity for Facebook, Instagram and its own generative services, so it is trying to gain more control over the cost of inference, the stage where a model is already trained and answers user requests.

This kind of work does not always call for the most general-purpose accelerator. Sometimes it pays to build a chip for specific internal use cases and squeeze out the best efficiency per watt and per dollar.

Meta says it already uses hundreds of thousands of MTIA chips for inference in recommendations, organic content and ads. In other words, this is not a lab experiment but a built-in part of its production infrastructure. Against that backdrop, the company is speeding up development of new generations.

Another number is telling: in 2026 Meta expects capital expenditures of $115–135 billion, and a noticeable share of that goes into AI infrastructure. In-house accelerators should help it not only scale faster but also reduce its dependence on the prices and supply of outside vendors.

What the roadmap looks like

At the center of the new roadmap are four MTIA chips that Meta plans to deploy much faster than is typical in the industry. While many AI chips are refreshed every one to two years, Meta wants to move at a pace of roughly six months. A modular approach makes this possible: new generations are designed to slot into already prepared racks and network infrastructure without rebuilding the site.

• MTIA 300 is already in production and is built for ranking and recommendation tasks.
• MTIA 400 will be the next step and, according to the company, is being readied for deployment in Meta's data centers.
• MTIA 450 is being designed primarily for generative AI inference.
• MTIA 500 will continue the same line and is meant to strengthen the infrastructure in 2027.

Starting with MTIA 400, Meta designs not just the chip itself but a complete system around it: several server racks, networking and liquid cooling. This matters because in large AI clusters, the performance ceiling has long been set not only by compute but also by data delivery, power consumption and heat removal. The better the company optimizes the whole stack, the less it overpays for generality it does not always need.

Why this is not a replacement for Nvidia

Despite the loud announcement, Meta does not explicitly pitch MTIA as an alternative that will push Nvidia or AMD out of its data centers. Instead, the company talks about a "portfolio" approach: different types of workloads require different chips, and there is simply no single universal solution for everything.

In the coming years, outside accelerators will remain critically important, especially where the heaviest scenarios of training large models are involved, or where quick access to a ready-made hardware and software ecosystem is needed. Meta's bet is different: move to its own chips the workloads where it can get the biggest gain in price and efficiency.

At the same time, the company is trying not to build a closed ecosystem. MTIA is built from the start around familiar industry tools: PyTorch, vLLM, Triton and Open Compute Project standards. This makes deployment in existing data centers easier. Part of the development is done together with Broadcom, and manufacturing is entrusted to TSMC, so even "in-house" silicon has a long external chain of partners.

What this means

Meta shows that the AI hardware race is shifting from simply buying the most powerful GPUs to a more nuanced model: large platforms are assembling a mixed fleet of third-party and in-house accelerators. For the market, it is a signal that the main shortage in the coming years is not only chips themselves but the ability to quickly tailor infrastructure to a specific type of AI workload.

*Meta has been designated an extremist organization and is banned in Russia.