Meta’s Iris AI Chip Enters Production in September as $145B Build Kicks Off

Meta will begin manufacturing its in-house Iris AI chip in September, according to an internal company memo reviewed by Reuters. The chip, designed by Broadcom and fabricated by Taiwan Semiconductor Manufacturing Co., enters production roughly four months after Meta publicly introduced it under its MTIA program. The memo lays out a $145 billion infrastructure build that doubles Meta’s compute capacity to 14 gigawatts by 2027.

The September ramp is the latest signal that hyperscalers are no longer content to wait in line for Nvidia’s flagship GPUs. Meta, like Google and Amazon before it, is moving key AI workloads onto custom silicon built on its own schedule. Iris adds a layer of internal capacity to Meta’s existing Nvidia and AMD GPU orders. The September production start is also the first concrete schedule Meta has published for the chip since introducing it in March under the MTIA program. Meta declined to comment on the leaked memo.

What the Internal Memo Reveals

Meta’s chip team cleared Iris through bug testing in roughly six weeks without surfacing any significant problems, according to the memo covered in Meta’s September Iris production plans. Broadcom is serving as the chip’s design partner, while TSMC has been tapped to handle high-volume fabrication. The two-company handoff has been Meta’s default playbook since the first MTIA generation, and the agreement was extended earlier this year to cover multiple MTIA generations through 2029.

The memo lays out the engineering rationale. Adopting the latest GPUs at Meta’s scale “has been a heavy lift, and it has cost us time,” the document said, referring to the engineering lift required to integrate each new Nvidia or AMD generation. The custom silicon line is meant to take that load off Meta’s GPU integration teams. Iris is the chip Meta is shipping to address that bottleneck. The September start date lines up with the memo’s broader two-step infrastructure plan.

The memo frames the ramp inside a wider GPU purchase plan that includes millions of Nvidia chips and up to 6 gigawatts of AMD Instinct accelerators. Meta is supplementing those deals, not canceling them. Iris will absorb workloads Meta decides to route internally rather than running them on merchant silicon.

Where Iris Fits Inside Meta’s MTIA Lineup

Iris is one of four planned chip generations under Meta’s Meta Training and Inference Accelerator, or MTIA, program. The full lineup was first unveiled in March 2026, with the MTIA 300 already in production for ranking and recommendation workloads across Facebook and Instagram. Iris is the next to enter production, as covered in Meta’s unveiling of four MTIA chips in March. The MTIA 450 and 500 follow in 2027, with the technical specifications detailed in Meta’s published MTIA technical roadmap. Newer parts in the family are expected among the first custom AI chips built on a 2-nanometer process.

Iris occupies the same kind of role Google reserved for its Tensor Processing Units and Amazon for its Trainium line, that of bulk internal workloads that do not need the bleeding edge of Nvidia’s general-purpose GPUs. Meta’s MTIA chips are not sold to outside customers. They are built solely for the company’s own apps and generative AI products.

The Suppliers Behind the September Ramp

Iris is the chip, but the September ramp depends on a wider supply chain Meta has spent the past several months quietly locking down. The memo confirmed long-term supply contracts across three hardware categories: memory chips from Samsung Electronics, flash storage from Sandisk, and fiber-optic equipment from Sumitomo Electric. None of the three hardware suppliers would comment to the press. Sandisk explicitly declined, while Samsung and Sumitomo Electric had not responded by press time.

• Samsung Electronics: long-term memory chip supply
• Sandisk: long-term flash storage supply
• Sumitomo Electric: long-term fiber-optic equipment supply
• Broadcom: chip design partner, agreement extended through 2029
• TSMC: high-volume fabrication

The hush is itself part of the story. Memory and fiber-optic capacity have become some of the most contested inputs in the AI buildout, with major hyperscalers signing multi-year deals to lock in supply before competitors can. Meta’s contracts are designed to keep its compute buildout on schedule regardless of what rivals do. The deals cover the memory, storage, and fiber-optic cabling that connect Meta’s data centers. Meta itself declined to comment on the leaked memo.

Beyond the new supply deals, Meta has also deepened its existing relationships with Broadcom and TSMC. Broadcom’s chip design agreement was extended earlier this year to cover multiple MTIA generations through 2029. TSMC remains the fabrication partner for the Iris ramp.

The combined chain gives Meta control over every layer of the Iris ramp, from chip design through fabrication and into the components that fill out the data center. The custom silicon line is one piece of a wider vertical integration strategy. The September production start is the first concrete payoff of those contracts. How those costs compare to merchant GPU alternatives will become clearer once the chips reach data center production in late 2026.

$145 Billion and 14 Gigawatts

The Iris ramp sits inside one of the largest single-year infrastructure plans in corporate history. The company is guiding 2026 capital expenditure to between $125 billion and $145 billion, with nearly all of the increase directed toward data centers, GPUs, and custom silicon. The 14-gigawatt target exceeds the total electricity consumption of many small countries, underscoring the energy scale of training and running large AI models.

By the end of 2026, Meta expects 7 gigawatts of computing capacity to come online, doubling to 14 gigawatts by 2027. The scale puts Meta in the same league as Microsoft, Google, and Amazon on infrastructure spend. Most of the spending still flows to third-party GPUs from Nvidia and AMD. The custom silicon slice is meant to absorb some of the inference workload growth that would otherwise fall on merchant GPUs.

• 7 gigawatts of compute capacity online in 2026
• 14 gigawatts of compute capacity planned by 2027
• Up to 6 gigawatts of AMD Instinct GPUs under multiyear deal
• A new chip every six months through 2027

What Changes for Nvidia and AMD

Iris is not a direct replacement for Nvidia’s data center GPUs. Meta continues to expand its GPU purchases in parallel with the custom silicon rollout. Earlier this year, Meta inked deals for millions of Nvidia chips and a multiyear agreement with AMD to deploy up to 6 gigawatts of AMD Instinct accelerators. The Nvidia purchases are some of the largest single-vendor GPU commitments in the industry. Those deals continue alongside Iris as part of Meta’s broader GPU diversification strategy.

The competitive read shows up in the cadence. Meta’s six-month chip release cycle is shorter than Nvidia’s annual product timeline. That gives Meta room to optimize specific parts of its stack faster than the merchant GPU roadmap can move. Nvidia keeps the most demanding training workloads, while Meta’s MTIA chips take the inference and recommendation jobs that drive daily traffic on Facebook and Instagram. The same pattern is playing out at Anthropic, where the company is in early talks with Samsung over a custom chip on the 2-nanometer process, as detailed in Anthropic’s parallel 2nm chip talks with Samsung.

The custom silicon line runs in parallel with the Nvidia and AMD purchases. Iris absorbs workloads Meta decides to route internally. Nvidia still sits at the top of the stack for the largest model training.

Why Meta Is Iterating Every Six Months

The cadence is the most unusual piece of Meta’s strategy. Most chip companies operate on an annual or two-year release schedule. Meta’s VP of engineering, Yee Jiun Song, made the case directly when the MTIA family was first unveiled in March. Song said the cadence was set to keep pace with AI model evolution rather than traditional chip industry timelines. The modular chiplet design lets Meta swap in newer process nodes and HBM stacks on a faster clock. The same playbook is showing up at DeepSeek, which is designing its own inference processor to cut reliance on Nvidia and Huawei hardware, per DeepSeek’s custom inference silicon effort.

It’s unusual for any silicon company or team to be releasing a new chip every six months. It’s a very quick cadence.

Song’s explanation was that AI workloads are evolving faster than traditional chip development cycles. Meta deliberately takes an iterative approach, with each MTIA generation building on the last using modular chiplets, incorporating the latest AI workload insights and hardware technologies. Song said Meta expects its chips to have a “standard five-plus years of useful lifetime,” a longer horizon than the release cadence implies. Iris is the first MTIA generation that will ship on a production schedule tight enough to test that strategy in earnest.