Meta's Iris AI Chip Enters Production: The 14-Gigawatt Plan

Meta's Internal Memo Confirms Iris Production in September

Reuters reported on 9 July 2026 that an internal Meta Platforms memo had revealed the company's custom AI chip, code-named Iris, is scheduled to enter mass production in September 2026. The memo, circulated among Meta's infrastructure and engineering leadership, confirmed that validation testing of the Iris silicon cleared in approximately six weeks without a single major design anomaly — an unusually short debug cycle for a novel chip generation. The early test results allowed Meta to proceed with its production schedule without the design revision that typically delays custom silicon programmes by several months.

What the MTIA Programme Is and Where Iris Fits

Iris is part of Meta's Training and Inference Accelerators programme, known internally as MTIA. The programme was established to develop purpose-built Application-Specific Integrated Circuits, or ASICs, for the AI workloads that Meta runs at the largest scale: recommendation ranking for Facebook and Instagram feeds, personalisation for Reels, and inference for Meta AI across WhatsApp, Messenger, and Ray-Ban smart glasses. These workloads run billions of times per day, and the cost of running them on general-purpose GPU clusters has driven Meta to invest in dedicated silicon.

Iris is one of four chip generations that Meta plans to release under MTIA at a cadence of approximately one new design every six months through 2027. That pace mirrors Apple's A-series chip rhythm and reflects a strategic decision to treat semiconductor design as an internal competency rather than a dependency on external vendors such as NVIDIA.

Design and Manufacturing: Broadcom and TSMC

Meta developed Iris in collaboration with Broadcom, which contributed expertise in custom ASIC design and high-bandwidth interface architecture. The chip will be manufactured by TSMC. Meta has not publicly disclosed the specific TSMC process node for Iris, though prior MTIA generations used advanced nodes to balance transistor density, memory bandwidth, and thermal management for inference workloads.

The Broadcom partnership follows a model other hyperscalers have used successfully — Google with its TPUs, AWS with Trainium and Inferentia — to accelerate custom silicon development. Rather than building a full semiconductor design organisation from scratch, Meta leveraged Broadcom's existing expertise to compress the timeline from chip specification to production-ready silicon.

Why an ASIC Rather Than a GPU?

GPU clusters are flexible: they can be reprogrammed for any workload, which is why they dominate AI training. That flexibility comes at cost. A general-purpose GPU consumes more power and die area per inference operation than a purpose-built ASIC handling the same task. For Meta, where recommendation ranking and feed personalisation run billions of inferences per minute across its family of apps, even a modest improvement in inference efficiency per chip translates into hundreds of millions of dollars in annual infrastructure savings.

The 14-Gigawatt Target and What It Requires

The production timeline for Iris maps directly to Meta's datacenter expansion plan. The company added approximately 1 gigawatt of computing infrastructure in the first half of 2026 and plans to add roughly 5.5 gigawatts more by year-end, bringing total capacity to about 7 gigawatts by the close of 2026. The goal for 2027 is to double that figure to 14 gigawatts — an addition of 7 gigawatts in a single calendar year.

For context, one gigawatt of datacenter capacity is roughly equivalent to the power consumption of a large industrial city. At 14 gigawatts, Meta's AI infrastructure would draw electricity at a scale comparable to the national grids of mid-sized countries. Meta's 2026 capital expenditure is guided at between 125 billion and 145 billion US dollars, with the majority directed at AI infrastructure: land, buildings, power, cooling, networking, and chips.

What This Means for Teams Building on Cloud AI in India

For engineering teams outside the hyperscaler tier — including the software product and AI companies building on AWS, Azure, or Google Cloud — Meta's infrastructure expansion has an indirect but meaningful effect. The more inference capacity Meta deploys on proprietary ASICs like Iris, the less pressure it places on NVIDIA GPU supply, and the more competitive pricing pressure builds on inference compute across cloud platforms.

For software teams in India building fintech platforms, enterprise SaaS products, or AI-native applications on top of cloud GPU infrastructure, lower inference costs translate directly into expanded feature economics. AI features that are presently too expensive to run in production at scale become viable as infrastructure supply grows and per-inference costs decline. Meta's MTIA programme and the Iris production ramp are part of the supply-side story that will drive those cost curves downward over the next twelve to eighteen months.

The Bottom Line

An internal Meta Platforms memo reported by Reuters on 9 July 2026 confirmed that the company's custom Iris AI chip will enter mass production in September 2026. Iris was developed in collaboration with Broadcom and will be manufactured by TSMC as an ASIC optimised for high-volume inference workloads including recommendation ranking and feed personalisation. The chip cleared validation testing in six weeks with no major design anomalies. Iris underpins Meta's plan to expand datacenter compute from 7 gigawatts at end-2026 to 14 gigawatts by end-2027, supported by capital expenditure of 125 to 145 billion US dollars in 2026. Meta plans to release four MTIA chip generations at roughly six-month intervals through 2027.