- Production. Meta will begin manufacturing its custom Iris (MTIA 400) AI chip at TSMC’s 2-nanometre node in September 2026 — one of the first commercial AI chips to reach that process node.
- Scale. Meta is targeting 7 gigawatts of AI compute capacity by end of 2026, with a goal to double that to 14 GW in 2027; Iris is central to reaching those targets without buying all capacity from Nvidia.
- Cost pressure. Meta’s $125–145 billion 2026 capex budget is almost entirely directed at AI infrastructure; the in-house chip programme is designed to reduce exposure to GPU price inflation in a constrained supply market.
Meta confirmed on July 9, 2026, through reporting by TechCrunch and The Next Web, that at least one chip in its Meta Training and Inference Accelerator programme had cleared its testing phase — in approximately six weeks — and that mass production at TSMC would begin in September. The chip, designated MTIA 400 and informally called Iris, is built on TSMC’s 2-nanometre process and co-designed with Broadcom under a partnership that runs through 2029.
The development places Meta alongside OpenAI, which unveiled its own custom inference chip Jalapeño co-designed with Broadcom, as the second major AI lab to bring a proprietary silicon programme close to production scale. Google’s TPU and Amazon’s Trainium chips are the more mature precedents, though analysts note those programmes took several years to meaningfully reduce external GPU dependency.
Architecture and supply chain
The Iris chip uses a modular chiplet design that allows Meta to swap components without a full redesign cycle — an approach intended to match a roughly six-month shipping cadence, faster than the annual cycles typical in the semiconductor industry. TSMC handles fabrication; Samsung supplies RAM; Sandisk provides storage; Sumitomo Electric supplies fibre-optic interconnects. The Broadcom co-design partnership, which extends through 2029, covers both the chip architecture and the custom networking fabric Meta requires to link chips across its data centres.
Meta has been producing AI chips since 2023 under the MTIA programme, beginning with the MTIA 300, which handles ranking and recommendation inference across Meta’s platforms. The 400, along with the 450 and 500 variants introduced in March 2026, targets more computationally demanding generative image and video inference tasks — the workloads most exposed to GPU supply tightness and price volatility.
Reducing Nvidia dependence
Meta remains one of Nvidia’s largest customers globally and is simultaneously pursuing multibillion-dollar agreements with AMD for Instinct GPUs and with Amazon for Graviton5 CPUs — a hedging strategy designed to avoid single-supplier concentration risk at a moment when advanced GPU allocations remain constrained. The Iris programme is not intended to replace Nvidia outright; analysts covering Meta’s infrastructure describe the goal as absorbing incremental compute growth in-house, allowing total GPU procurement to grow more slowly than overall capacity targets while the company’s AI workloads expand.
The company’s 2026 capital expenditure guidance of $125–145 billion is directed almost entirely at data centres, GPUs, and custom silicon. Meta has also explored renting excess compute capacity to third-party customers — a move that would allow the infrastructure investment to generate revenue while capacity builds ahead of peak internal demand. Mark Zuckerberg has publicly cited targets “measured in gigawatts” for future compute deployment, framing Meta’s AI infrastructure as a long-duration capital project rather than a procurement cycle tied to specific model launches.
Production timeline
The September production start for MTIA 400 puts mass deployment across Meta’s data centres on a timeline of late 2026 to early 2027. The 450 and 500 variants, aimed at the most demanding generative workloads, are scheduled for broader deployment through 2027. The six-week testing window for the 400 chip — described by sources as unusually fast — suggests the Broadcom co-design methodology has matured since the first MTIA generation, reducing the integration and validation cycle that typically consumes the most time between tape-out and production release.