Custom silicon is no longer a side project at Meta
Meta is taking the next decisive step in its custom silicon strategy. According to an internal memo reviewed by Reuters, the company will begin manufacturing its “Iris” AI data center chip in September 2026 — the latest generation in its Meta Training and Inference Accelerator (MTIA) program. The chip cleared its bug-testing phase in just six weeks with no significant issues, clearing the path from the lab to the foundry floor.
Iris isn’t happening in a vacuum. It’s one of four planned chip generations, with Meta targeting an aggressive cadence of a new chip roughly every six months through 2027 — far faster than the industry’s typical annual release cycle.
The MTIA roadmap: from recommendation engines to Llama
Meta’s custom silicon journey started modestly. The MTIA 100 and MTIA 200 (detailed at ISCA ‘23 and ‘25) were optimized for ranking and recommendation workloads — the dominant AI task before generative AI exploded. Then Llama happened, and Meta pivoted hard. The subsequent MTIA 300 through 500 chips were re-architected for GenAI inference at planetary scale.
Iris represents the next leap: a chip designed from the ground up to handle both training and inference for Meta’s AI models — including Llama 4 Scout, Llama 4 Maverick, and the Muse family. The design partnership with Broadcom and fabrication at TSMC mirrors the playbook used by other hyperscalers building custom AI accelerators.
Crucially, Iris won’t replace GPUs. The memo explicitly states that Meta plans to use Iris alongside the enormous volumes of Nvidia and AMD GPUs already deployed. It’s a hybrid strategy: off-the-shelf for flexibility, custom silicon for cost efficiency at scale.
The scale: 14 gigawatts by 2027
The chip news comes wrapped in infrastructure numbers that illustrate just how big Meta’s AI ambitions have become. The memo reveals:
- 7 gigawatts of computing capacity in 2026
- 14 gigawatts projected for 2027 — doubling in a single year
- Up to $145 billion in AI infrastructure spending this year alone
- Multi-year supply agreements locked in with Samsung (memory), SanDisk (flash storage), and Sumitomo Electric (fiber optics)
To put 14 GW in perspective: that’s roughly the output of 14 large nuclear reactors, or enough to power about 10 million homes. All dedicated to running AI workloads.
Why Iris matters for the Llama ecosystem
For the open-source AI community that has built around Llama, Iris signals that Meta is serious about the long-term economics of serving these models. The MTIA program isn’t just about cutting a check to Nvidia — it’s about building infrastructure that makes serving Llama at global scale economically sustainable.
The memo candidly acknowledges that “adopting the latest GPUs at Meta’s scale has been a heavy lift, and it has cost us time.” Custom silicon lets Meta tune the hardware to its specific workload patterns — different batch sizes, attention mechanisms, and MoE routing strategies that generic GPUs handle with overhead.
Morgan Stanley analysts have described the broader supply-demand dynamic in AI chips as “chipflation” — a recognition that AI infrastructure demand has grown large enough to distort the entire semiconductor supply chain. Iris is Meta’s hedge against that reality.
What’s next
The September production start puts Iris-derived inference capacity online sometime in early 2027, assuming typical semiconductor manufacturing and deployment lead times. Combined with Meta’s extended partnership with Broadcom through 2029 and a multi-year deal with AMD for up to six gigawatts of Instinct GPUs, the company is building a diversified compute fabric that should keep Llama and Muse models running at competitive cost-per-token for years to come.
For developers running Llama models — whether self-hosted or through Meta’s upcoming API — the Iris program represents the hardware foundation that makes those models economically viable at scale. Custom silicon isn’t glamorous, but it’s the difference between an open-source AI strategy that works and one that’s just marketing.
Sources: Reuters, Open Data Science, Yahoo Finance