Meta's self-developed AI chips to be deployed in the first half of next year, aiming to reduce costs and dependence on Nvidia

Meta is accelerating the deployment of its self-developed AI chips to reduce its reliance on NVIDIA and cut data center operating costs.

The company stated that the third generation in-house chip, MTIA 450 (codenamed Arke), will begin to be used in data centers in the first half of next year. The fourth generation, MTIA 500 (codenamed Astrid), will have its design work completed in about a month, with expectations to enter data centers by the end of 2027, at a much larger scale.

This chip roadmap marks substantial progress for Meta in reducing the cost of AI infrastructure.

Meta Vice President of Engineering Yee Jiun Song said that, compared with products currently shipped by NVIDIA, Meta's self-developed chips are more efficient when running AI models, "precisely because we handle so much of the engineering internally."

Within the next 12 months, Meta’s commitment to deploying self-developed chips already exceeds 1 gigawatt equivalent in power consumption. Yee Jiun Song said that afterwards, further accelerated expansion is expected—provided the AI market or demand does not collapse.

At the time of writing, on Tuesday Meta was up 0.18% intraday and NVIDIA was up 0.74%.

Meta collaborates with Broadcom for chip design and works with TSMC for manufacturing, following an overall direction of reducing dependence on NVIDIA’s industry dominance.

Meta first announced its self-developed chip plan in 2023. The current third-generation product received its first batch of 12 samples from TSMC on September 1, with measured performance differing from simulation results by less than 2% to 3%. On the first day of testing, the engineering team successfully ran Meta’s own model on the new chip.

Yee Jiun Song noted that the test results indicate no significant defects in the chip design phase, though several months of debugging and optimization are still needed, and TSMC's production yield is also steadily climbing.

He emphasized that this chip series mainly relies on high bandwidth memory and is oriented toward general inference scenarios, not the ultra-fast inference market with extremely high response requirements. "This is our main chip for general inference," he said.

Meta previously planned to develop a chip codenamed Olympus, intended to support both AI model training and inference, originally scheduled for release between 2028 and 2029.

But according to Yee Jiun Song, that project has now been canceled, with the company focusing instead on inference chips, mainly for cost reasons.

"Once you start to scale up to several gigawatts of capacity, cost becomes absolutely critical," he said. He pointed out that if a chip supporting both training and inference costs about 30% more, this would be "completely unacceptable" at large deployment scales.

Meta Superintelligence Labs is helping fine-tune its chips to further improve inference performance by providing projected demand information for future AI models.

After the Astrid development is complete, Meta says that future chip research will shift toward improving computational speed and throughput—the number of AI tasks processed per unit time. The introduction of optical fiber technology could further enhance chip performance as well.

"We have a very solid roadmap," Yee Jiun Song said. "In the next few years, we expect to continue producing chips that can compete with supplier products."

This position reflects Meta’s intention to gradually take control of core AI infrastructure technology through in-house silicon, while also cementing a long-term advantage in cost and energy efficiency.