•
•
•
•
•
•
•
Meta has started mass production of its custom-designed "Iris" AI chip, part of an aggressive push to build out in-house computing capacity for its AI systems rather than relying solely on Nvidia's GPUs, according to reports citing Reuters and confirmed by multiple industry outlets this month.
Iris is the latest generation of Meta's MTIA (Meta Training and Inference Accelerator) custom silicon program, co-designed with chipmaker Broadcom. Reports indicate the chip entered production in September 2026 after passing internal testing, with Meta aiming to use it across its data centers to support the enormous computing demands of training and running its AI models, including its Llama model family and recommendation systems that power Facebook, Instagram, and WhatsApp.
Meta reportedly plans to use Iris and other infrastructure investments to roughly double its total AI computing capacity to around 14 gigawatts by 2027, up from an estimated 7 gigawatts currently — a scale comparable to the power draw of a mid-sized country's electricity grid.
Meta, like other major AI companies, has relied heavily on Nvidia GPUs to train and run its large AI models. Nvidia chips remain in high demand industry-wide, with long lead times and enormous costs. Building custom silicon like Iris allows Meta to:
Broadcom, which also designs custom AI chips for other major technology companies, has become a central player in this shift toward hyperscalers building bespoke silicon rather than buying only off-the-shelf GPUs.
Meta first began developing custom AI accelerators through its MTIA program several years ago, initially focused on inference workloads (running trained models) for its recommendation systems. Iris represents a more ambitious generation of the program, reportedly intended to handle a broader range of AI workloads, including training, as Meta scales its AI ambitions across its core apps, AI assistant products, and its Reality Labs division.
The move mirrors similar custom-silicon strategies at Google (TPUs), Amazon (Trainium/Inferentia), and Microsoft, all of which have pursued in-house chip development to manage the cost and supply constraints of relying entirely on Nvidia.
Meta's stated target of roughly 14 gigawatts of AI compute capacity by 2027 will require continued data center construction, power procurement, and further chip production scaling. Industry watchers will be looking for confirmation of Iris chip performance benchmarks once deployed at scale, as well as whether Meta expands its custom silicon strategy to additional chip generations or workload types in the coming product cycles.
What is Meta's "Iris" chip?
Iris is Meta's newest custom AI accelerator chip, part of its MTIA chip family, designed in partnership with Broadcom to support training and running Meta's AI models.
Why is Meta building its own AI chips instead of just buying Nvidia GPUs?
Custom chips can reduce Meta's reliance on a single supplier, potentially lower costs for its specific AI workloads, and give the company more control over its infrastructure roadmap.
How much AI computing capacity is Meta targeting?
Reports indicate Meta aims to roughly double its AI compute capacity to about 14 gigawatts by 2027.
Does this mean Meta will stop using Nvidia chips?
No. Custom chips like Iris are expected to supplement, not fully replace, Meta's continued use of Nvidia GPUs across its AI infrastructure.
Meta's move to mass-produce its in-house Iris AI chip marks a significant step in the broader industry trend of major AI companies developing custom silicon to manage the soaring cost and scale of AI infrastructure. With a stated goal of roughly doubling compute capacity by 2027, the initiative reflects just how capital- and energy-intensive the AI race has become for the industry's largest players.
