Meta is becoming a chipmaker. The company behind Facebook and Instagram is building Meta AI chips in-house, its own custom silicon to run the enormous AI workloads that power ranking, recommendations, and its generative-AI products. The clearest signal yet came this month: Meta plans to put a new data-center chip, code-named Iris, into production in September 2026, part of a broader push to design its own accelerators rather than buy every chip from Nvidia and AMD.
The short version: Meta’s chip effort runs under a program called MTIA, short for Meta Training and Inference Accelerators. It has already deployed one generation, has three more planned over roughly two years on an unusually fast schedule, and is co-developing the designs with Broadcom while Taiwan Semiconductor Manufacturing Company handles the actual fabrication. The goal is to cut compute costs, loosen its dependence on outside chip vendors, and roughly double its computing capacity by 2027. This piece explains what Meta is building, the Iris timeline, the four-chip roadmap, the partners involved, and why Meta is doing it.
What Meta AI chips are: the MTIA program
MTIA is Meta’s family of custom AI chips, designed in-house for its own workloads rather than sold to anyone else. The name spells out the ambition: these are accelerators built for both training AI models and running them in production, which is inference. Historically Meta’s custom silicon leaned toward inference and toward the ranking and recommendation systems that decide what shows up in your feed, but the current roadmap pushes into training and generative-AI workloads too.
The strategic idea is straightforward. Meta runs some of the largest AI systems in the world, and buying every accelerator from third parties is expensive and leaves the company dependent on a small number of suppliers. Designing its own chips, tuned specifically for the workloads Meta actually runs, is a way to control cost and supply. For a company spending on the scale Meta is, even modest efficiency gains translate into very large numbers.
The Iris chip and the September production start
The most concrete piece of news is Iris. It is Meta’s in-house data-center AI chip, and the company plans to begin manufacturing it in September 2026. According to reporting on an internal memo, testing the chip took only about six weeks and turned up no major issues, which is a fast and clean bring-up for a piece of silicon this complex.
Iris is one part of the larger MTIA project rather than a standalone product. Its production start matters because it marks the point where Meta’s custom-silicon ambitions move from deployed-but-limited to a serious, scaling effort aimed at handling more of the company’s AI workload on its own hardware. Meta is tying this to a large capacity expansion, with the aim of significantly increasing the computing power available across its apps.
Four chips in two years
What sets Meta’s approach apart is the pace. The company has detailed four MTIA chip generations arriving over roughly two years. MTIA 300 handles ranking and recommendations training and is already in production. MTIA 400, 450, and 500 are designed to handle all workloads but will primarily support generative-AI inference in the near term and into 2027.
The cadence is the striking part. The industry typically launches a new AI chip every one to two years. Meta says it has built the capacity to release a new generation every six months or less, and it credits modular, reusable designs for that speed. Instead of designing each chip from scratch, Meta reuses building blocks across generations, which shortens the path from one chip to the next. If it holds, that rhythm is aggressive by the standards of the whole industry.
The partners: Broadcom and TSMC
Meta is not doing this entirely alone, and the partners tell you how the work is split. Broadcom is Meta’s co-development and design partner across multiple generations of MTIA chips. Broadcom has deep experience turning a customer’s requirements into a manufacturable custom chip, which is exactly the expertise a company newer to silicon design wants alongside it.
Taiwan Semiconductor Manufacturing Company, or TSMC, handles fabrication, the physical manufacturing of the chips. That is the same foundry that builds leading-edge silicon for most of the industry, including Nvidia, AMD, and Apple. So the shape of Meta’s operation is a common one for a company designing its own chips: Meta and Broadcom design, and TSMC manufactures. Meta owns the design and the workload knowledge; it does not need to own a fabrication plant.
Why Meta is doing it
The motivation comes down to cost and independence. By building its own silicon, Meta is trying to cut spending on compute and loosen its reliance on third-party chip vendors, specifically Nvidia and AMD. Those companies supply the accelerators that most of the AI industry runs on, and demand has made them both essential and expensive. A chip tuned to Meta’s own workloads can, in principle, do the same work more cheaply for the specific things Meta runs.
The scale behind this is worth sitting with. Meta has outlined a plan to bring seven gigawatts of computing capacity online in 2026 and grow to fourteen gigawatts by 2027, with projected AI infrastructure spending for the year reported as high as $145 billion. Custom silicon is how Meta hopes to make spending at that magnitude more sustainable. For a broader sense of how fast the hardware frontier is moving, our coverage of the IBM sub-1nm chip and the NVIDIA RTX Spark superchip shows the pace on the manufacturing and consumer sides of the same race.
The honest caveats
A few things keep this in perspective. First, building your own chips does not mean abandoning Nvidia and AMD. Meta continues to buy their hardware in large quantities, and it recently signed major deals with both. Custom silicon is about reducing dependence and cost at the margin, not replacing external suppliers outright, at least for now.
Second, the six-month cadence is a claim, not yet a proven track record. Shipping new silicon that quickly is genuinely hard, and modular designs help but do not guarantee it. The ambition is real; the execution is what to watch.
Third, these chips are internal. Unlike Nvidia or AMD, Meta is not selling MTIA to anyone else, so this does not change what hardware other companies can buy. It changes Meta’s own cost structure and supply, and it is one more example of a hyperscaler designing its own accelerators, alongside Google, Amazon, and Microsoft. Meta is joining a club rather than inventing the idea.
Frequently Asked Questions
What are Meta’s AI chips called?
They are part of a program called MTIA, short for Meta Training and Inference Accelerators. The specific data-center chip entering production in September 2026 is code-named Iris. MTIA is a family of custom chips Meta designs for its own AI workloads.
When does Meta’s Iris chip go into production?
Meta plans to begin manufacturing the Iris chip in September 2026. Reporting on an internal memo said testing took about six weeks and found no major issues, an unusually fast and clean bring-up for a chip of this complexity.
Is Meta replacing Nvidia and AMD?
No, not outright. Meta is building its own chips to cut costs and reduce its dependence on outside vendors, but it still buys large quantities of Nvidia and AMD hardware and recently signed major deals with both. The custom silicon complements those purchases rather than fully replacing them.
What is MTIA used for?
MTIA chips are built for both training AI models and running them in production. MTIA 300 is used for ranking and recommendations training and is already deployed, while MTIA 400, 450, and 500 are designed to handle all workloads and will primarily support generative-AI inference in the near term.
Who manufactures Meta’s chips?
Meta co-develops the designs with Broadcom, and Taiwan Semiconductor Manufacturing Company (TSMC) handles the fabrication. Meta and Broadcom design the chips; TSMC, the same foundry that builds silicon for Nvidia, AMD, and Apple, manufactures them.
How fast is Meta releasing new chips?
Meta says it can release a new MTIA generation every six months or less, compared with the industry norm of every one to two years. It credits modular, reusable designs that let it carry building blocks from one chip to the next rather than starting each design from scratch.
Can other companies buy Meta’s AI chips?
No. MTIA chips are designed for Meta’s own use and are not sold to other companies. They change Meta’s internal cost structure and supply, not the hardware available to the rest of the market.
Why is Meta investing so heavily in its own silicon?
Cost and independence at very large scale. Meta plans to grow computing capacity from seven gigawatts in 2026 to fourteen by 2027, with AI infrastructure spending reported as high as $145 billion for the year. Custom chips tuned to its own workloads are how it hopes to make spending at that scale more sustainable.
