Meta plans to begin manufacturing its custom data-center AI chip, codenamed Iris, in September as part of a four-generation MTIA roadmap. The company aims to scale computing infrastructure to 14 gigawatts by 2027.

Building its own chips gives Meta more control over AI performance and cost. Custom silicon lets the company optimize for its specific workloads instead of buying off-the-shelf processors from Nvidia or AMD.

Why Custom Chips Matter

The race for AI dominance hinges on semiconductor power. Companies that design their own chips save money, reduce latency, and lock in competitive advantages. Google, Amazon, and Apple have all gone this route. Meta’s move signals the company is serious about long-term AI dominance.

Iris will power Meta’s internal systems first. The company runs Facebook, Instagram, WhatsApp, and Threads on enormous computing clusters. Dedicated chips tailored to Meta’s needs could save billions in hardware costs over time.

The MTIA roadmap spans multiple generations, suggesting Meta expects demand to grow. Four-generation timelines are rare. Only companies betting on sustained hyperscale investment make that kind of long-term commitment.

The 14-Gigawatt Target

Fourteen gigawatts is staggering. It represents roughly 1 percent of global electricity consumption. Meta is betting that AI training and inference will consume massive power for at least the next five years.

The target also reflects the scale of AI infrastructure. Major tech companies are spending $100 billion-plus annually on data centers and chips. Building custom silicon is a way to wring efficiency from that staggering investment.

Iris production in September means Meta expects to start deployment by fall 2026, with scaling into 2027 and beyond. The company is racing against competitors to own its infrastructure end to end.

Every major tech company is now building custom chips. It’s no longer optional.