Meta's Iris chip marks its first step into custom silicon for AI inference

Meta is embarking on a major vertical integration play, planning to manufacture its own AI processor called Iris beginning this September. According to internal communications reviewed by Reuters, the chip will handle a portion of the inference workloads that currently depend on third-party GPUs, giving the social media giant greater autonomy over its infrastructure scaling.

The Iris chip strategy#

Iris represents an expansion of Meta's existing Training and Inference Accelerators (MTIA) program, a push to move specific AI inference tasks onto purpose-built silicon. The new processor targets the heavy computational work underpinning content ranking, personalized recommendations, and generative AI services across Facebook, Instagram, WhatsApp, and Meta's other platforms.

Meta's roadmap reflects an ambitious cadence: the company plans to release updated variants roughly every six months through 2027. The MTIA 300 is already running in production for ranking and recommendation inference; upcoming 450 and 500 versions will focus on image and video generation tasks. This pace stands in sharp contrast to how traditional chip manufacturers move, positioning Meta as willing to iterate rapidly rather than wait for off-the-shelf solutions.

Backing the full supply chain#

Custom chip design alone isn't enough to sustain Meta's AI ambitions. Broadcom is supporting the Iris design work, while Taiwan Semiconductor Manufacturing Company (TSMC) will handle production. But the infrastructure puzzle extends far beyond processors: Meta also needs reliable access to high-bandwidth memory, flash storage, and networking equipment—commodities now in fierce competition across the industry.

To secure these critical inputs, Meta has inked long-term contracts with Samsung Electronics for memory, SanDisk for storage, and Sumitomo Electric for fiber-optic networking gear. This multi-layered supply strategy reflects lessons learned from the industry-wide chip shortages of recent years and mirrors similar efforts by Google (which continues scaling its TPU program) and Amazon (which has built Trainium and Inferentia chips).

The scale of ambition#

Meta's infrastructure buildout targets staggering proportions: the company plans to deploy roughly 7 gigawatts of computing capacity in 2026, then double that figure to 14 gigawatts by 2027. At that level, Meta's AI infrastructure alone would exceed the electricity consumption of many entire nations. The company has earmarked between $125 billion and $145 billion in capital expenditures for this year, placing it among the most aggressive infrastructure investors in corporate history.