Meta to Begin 'Iris' Chip Production Amid 14GW Infrastructure Expansion
Meta plans to begin production of its in-house 'Iris' AI chip in September, co-designed with Broadcom and manufactured by TSMC, to supplement existing GPU deployments. The company also intends to scale its infrastructure to 14 gigawatts by 2027 and explore a cloud model by leasing excess computing capacity.
Key Takeaways
- Production of the 'Iris' AI chip starts in September to supplement existing Nvidia and AMD GPU deployments.
- Meta plans to scale data center capacity from 7 gigawatts in 2026 to 14 gigawatts by 2027.
- Annual capital expenditure for AI infrastructure is projected to reach between $125 billion and $145 billion.
- Internal roadmap targets a six-month release cycle for new AI chips through 2027 to match the pace of model evolution.
- Long-term supply agreements secured with Samsung, SanDisk, and Sumitomo Electric for memory, storage, and fiber-optic components.
Why It Matters
Meta's move into custom silicon and third-party compute leasing marks a fundamental shift from a pure social media service to a vertically integrated infra-hyperscaler. By producing in-house chips like Iris, Meta can bypass 'chipflation' and vendor supply bottlenecks, directly impacting the unit economics of content recommendation and generative AI delivery. This infrastructure scale allows the company to potentially undercut established cloud giants by leasing excess capacity, turning internal capital costs into a B2B revenue stream. Watch for the 2027 rollout of the MTIA 500 chip as a signal that Meta has transitioned from augmenting third-party silicon to achieving internal hardware parity for generative AI inference.
Additional Context
The rollout of the Iris chip coincides with Meta's formation of 'Meta Compute,' a top-level initiative led by infrastructure head Santosh Janardhan and Superintelligence Labs director Daniel Gross. Per Bloomberg (July 2026), this program is exploring a dual-layer cloud model: an API-based service for hosted Llama models—similar to Amazon Bedrock—and a 'neocloud' offering providing direct access to raw GPU clusters. Analysts at Deutsche Bank suggested in July 2026 that Meta could undercut existing hyperscalers by as much as 30% by monetizing idle capacity between its own massive training runs. Meta's 2026 capital expenditure guidance of $125 billion to $145 billion represents one of the largest single-year infrastructure outlays in corporate history, nearly doubling its 2025 spend. According to Reuters (July 2026), internal memos indicate that the Iris chip recently cleared bug testing in just six weeks, an unusually fast turnaround for silicon of this complexity. This speed is essential for Meta's plan to deploy chips on an unprecedented six-month cadence, ensuring that hardware remains optimized for rapidly changing model architectures like the newly unveiled 'Muse Spark.' Beyond silicon, Meta is securing the physical foundation of this expansion through significant regional investments. Per regional reporting in July 2026, the company recently announced its largest data center outside the U.S., a $13 billion, 1-gigawatt facility in Alberta, Canada. By pre-loading capacity now, Meta aims to accommodate anticipated 2028 demand while using temporary surplus to compete with CoreWeave and Lambda in the high-growth market for bare-metal AI compute.
Read full article at msn.com
Enjoy our coverage?
Add StreamingMeme as a preferred source on Google to see more of our streaming news at the top of your Search results.
Add as preferred source