Meta is preparing four generations of MTIA chips to reduce its dependence on Nvidia and AMD

Meta is dramatically accelerating its own chip program: the company has already launched MTIA 300 into production and plans to release three more generations of accelerators by the end of 2027. The goal is straightforward — to reduce dependence solely on Nvidia and AMD in the race for compute power for recommendations, advertising, and generative AI.

Why Meta needs its own chips

Meta's infrastructure load is growing across multiple directions. The company needs capacity for both traditional content ranking systems on Facebook and Instagram, and for generative AI that answers queries, creates images, and supports other products in its ecosystem. Buying all of this only from external suppliers is too expensive and risky: demand for accelerators is high, delivery times are long, and margins heavily depend on prices for third-party hardware.

That's why Meta is betting on its own MTIA line — Meta Training and Inference Accelerator. It hasn't abandoned its partners and has already signed multi-billion-dollar deals with Nvidia and AMD, but in parallel it's trying to handle some tasks with its own silicon. The logic is that internal chips don't need to be universal like commercial GPUs.

They can be more precisely tailored to Meta's specific scenarios and thereby reduce inference costs.

MTIA Roadmap

Meta now has a fairly clear plan for several generations ahead. The company says it can release a new chip roughly every six months — notably faster than the industry's usual cycle, where one to two years often pass between generations. A modular architecture helps: new accelerators can be embedded into already-prepared racks and the network infrastructure of data centers.

• MTIA 300 is already running in production and is used for training ranking and recommendation models.
• MTIA 400 has passed laboratory tests and is preparing for deployment in data centers.
• MTIA 450 is being designed primarily for AI inference and should ship in volume in early 2027.
• MTIA 500 is planned for the second half of 2027 as the next step in the same line.

According to Meta, the transition from MTIA 300 to MTIA 500 should deliver approximately 4.5x growth in HBM memory throughput and 25x growth in compute performance. The company is placing special emphasis on inference: this stage, when the model is already responding to the user, is becoming the most expensive and widespread scenario for services with a billion-strong audience.

"Over the last two to three months, the pace of AI development has

accelerated sharply again, and silicon programs must keep up with the evolution of workloads."

Where are the bottlenecks

The problem is that developing your own chips is not just expensive, but very slow and risky. From design to factory production typically takes about two years, and the actual product refinement costs billions of dollars. Such a project only pays off if the company can load the hardware at massive scale and clearly understands what tasks will be relevant by launch time.

Earlier, it was reported that Meta abandoned its most advanced training chip, Olympus, after design problems and shifted focus to a simpler version. To accelerate the program, Meta tried to buy South Korean startup FuriosaAI for 800 million dollars, and after that failed, acquired Rivos along with more than 400 employees. This shows that the shortage in this race is not only in GPUs, but also in engineers who know how to build complex data center silicon.

What this means

Meta is not building a Nvidia replacement as "one big chip," but rather a hybrid model: it will continue to buy some of its infrastructure externally, while gradually shifting some work to its own MTIA. If the company maintains the pace of releases and actually manages to reduce inference costs, it will gain an important advantage in the AI race: greater control over costs, timelines, and the launch of new features for billions of users.