Meta will begin deploying its own artificial intelligence chip in its data centers in the first half of next year. It is called MTIA 450, internally known as Arke, and according to Meta, it can perform more work per watt of electricity and per dollar spent than the chips Nvidia currently offers. The company has also described for the first time its next generation, code-named Astrid, and acknowledged that it canceled one previously planned chip.
First Chips from Taiwan
Meta announced back in 2023 that it would design its own chips. It develops the designs together with Broadcom, while the chips are manufactured by Taiwan’s TSMC. In September, the first twelve Arke chips arrived at Meta from the factory, and engineers immediately began testing them.
The results differed from computer simulations by only two to three percent. That is very good news for a brand-new chip. On the very first day, technicians ran not only Meta’s own models on it, but also models from the Chinese companies DeepSeek and Alibaba.
Yee Jiun Song, head of the custom chip program and vice president of engineering, said of the new generations that each successive one carries slightly greater technological risk while also offering better performance. By this, he means a better ratio of performance to both energy consumption and cost.
What Arke Is Designed For
Arke is not a chip intended to compete with Nvidia’s most powerful processors in training large models. Training is the expensive and demanding phase during which a model is still learning. Arke is responsible for what comes afterward: routine operation. When a user enters something into a chat or content is being generated for them in an app, the completed model must respond. This process is called inference, and Meta performs it billions of times a day.
All generations of the MTIA series are built around fast, high-bandwidth memory capable of supplying data to the chip quickly enough to prevent idle time. None of them targets the fastest forms of inference, where fractions of a second matter. Meta is focusing on routine, everyday operation at high volume. These chips are intended to become the company’s primary tool for general-purpose inference.
Fourth Generation in Development
The fourth generation is designated MTIA 500 and code-named Astrid. Design work on it is expected to be completed in about a month, and it should arrive in data centers toward the end of 2027. Meta expects to deploy Astrid in far greater numbers than Arke, so this chip is intended to carry the main operational load.
A team from Meta Superintelligence Labs is also helping with optimization. It provides estimates of the requirements of models the company is still developing, allowing the hardware to be adapted to the software before it even exists.
Cancellation of the Olympus Chip
Meta originally worked on a chip code-named Olympus that was intended to handle both training and inference. It was expected to arrive sometime in 2028 or 2029. The company halted the project and went all-in on chips designed exclusively for inference. Cost was the deciding factor.
According to Song, a chip capable of handling both tasks would cost roughly 30 percent more. At ordinary volumes, that could be tolerated, but Meta is building capacity on the scale of gigawatts. When you start building gigawatts and gigawatts of capacity, costs really matter, Song said. At that scale, he considers a 30 percent premium completely unacceptable.
What Meta Expects from Its Own Silicon
A chip designed specifically for the company’s own models and servers does not need any extra capabilities and can therefore be cheaper and more energy-efficient. Meta has committed to deploying its own chips with power consumption exceeding one gigawatt during every twelve-month period, and Song indicated that the pace would subsequently accelerate even further.
The second advantage is control. The company can time its hardware development according to the models it is preparing and does not have to wait for what another company will deliver and when. This also reduces its dependence on a single supplier.
However, this does not push Nvidia out of the market. Meta will continue purchasing its processors because it still needs them to train the latest models and handle the most demanding tasks. The point of this strategy is that the company wants to reserve expensive, general-purpose graphics chips for work where they are truly worthwhile, while moving predictable routine operations to its own hardware.
Source: bloomberg.com
