Meta Is Building Its Own AI Chips to Make AI Cheaper

Meta has quietly finished the first version of its own AI chip. It's called MTIA 450, code-named Arke, and Meta plans to put it inside its data centers in the first half of 2027. The company got its first 12 chips from manufacturer TSMC on Sept. 1 and says early tests matched its expectations within 2% to 3%. Arke is built for one job: "inference," which is the moment a trained AI model actually answers a question, writes a caption, or recommends a video.

Why should you care? Running AI for billions of people is expensive. Every time you ask Meta's AI something, it costs the company money and electricity. Meta serves those features for free, so cheaper chips mean it can afford to keep AI free — and add more of it — without passing costs to you. Meta has committed to installing more than 1 gigawatt of its own chips within a year, roughly the electricity demand of a large city. Lower energy use also means less strain on power grids near its data centers.

The catch is that Arke can't do everything. It won't replace the Nvidia and AMD chips Meta still buys, because those can both teach AI models and respond fast enough for demanding tasks. Arke only handles the routine, high-volume work. Meta even canceled a bigger chip called Olympus, which would have done both jobs, because a dual-purpose design cost about 30% more. A follow-up chip, Astrid, is due by the end of 2027.

For you, nothing changes today — this is a 2027 story. But the whole industry is racing the same direction: Google, Amazon and Microsoft are all designing their own chips to escape Nvidia's prices. If it works, the cost of answering an AI question keeps falling, and the free AI features you already use get faster, smarter and more plentiful. If it doesn't, Meta's savings stay on its balance sheet instead of reaching your phone.