Meta Wants AI to Cost Less: Its Custom Chips Could Reshape Data Centers

Meta is set to roll out its next-generation AI accelerators for use in data centers by 2027. Meta believes that the accelerators will help it reduce power consumption and expenses as the need for AI continues to rise.

The development of the MTIA series has been ongoing since 2023. While the company is currently testing the MTIA 450, which is known as Arke, the next one, MTIA 500, or Astrid, is nearing completion. It is expected to be implemented in the company's data centers later in 2027.

Meta has a plan, and it is simple. The company would like to increase the amount of control it has over the hardware behind its AI services and reduce reliance on Nvidia. The development of the chips is being facilitated by Broadcom, whereas TSMC will handle production.

Meta Is Optimizing Every Watt and Every Dollar

However, Meta is not just focused on maximizing efficiency. It wants more artificial intelligence computation per watt of energy and dollars spent on hardware.

Such an issue takes even more significance when Meta aims to power AI accelerators with one gigawatt of power. Twelve of those chips have already been delivered to engineers by TSMC for evaluation purposes. The real-world performance of the processor reportedly was only about 2 to 3 percent lower than the simulation results.

Moreover, Meta also designs its own AI accelerators using the models developed by itself. In other words, engineers understand the workload these chips must process. Meta claims that its processors are general-purpose machines for AI inference and not the extremely fast processors required for instantaneous reactions.

Our Opinion

We think that Meta’s strategy may well impact the AI products that you’re using on a daily basis. Reduced costs for infrastructure might provide more opportunities for adding AI functionality without raising prices in the same way. This may force other tech giants to create their own custom hardware.

Only time will tell if Meta is able to translate its engineering edge into savings.

If you found this useful, share it with someone following the AI hardware race and explore our other technology stories for more updates.