Meta Platforms will begin manufacturing its in-house artificial intelligence chip, code-named Iris, from September this year, according to an internal company memo first reported by Reuters on Wednesday. The production timeline is part of a larger plan under which the Facebook and Instagram parent intends to double its total computing capacity to 14 gigawatts by 2027.
The memo, reviewed by Reuters journalists, also revealed that bug testing on the Iris chip wrapped up in just six weeks without surfacing any major problems. Neither the completion of testing nor the September production date had been disclosed publicly before now. Meta declined to comment on the report.
The disclosure lands at a moment when every major technology company is racing to lock down computing capacity, and when the cost of that race has started showing up in component prices across the electronics industry.
A Six-Week Test Run That Changed the Mood Inside Meta
Iris belongs to the Meta Training and Inference Accelerator family, better known as MTIA, a program of custom data center silicon that Meta designs internally. The company has been working on its own chips for more than half a decade, and progress over most of that period was slow enough that outsiders questioned whether the effort would ever matter at scale.
That perception is now shifting.
According to the memo, validation of Iris moved at an unusual pace for a chip of this complexity. Six weeks of testing produced no significant issues, a result that engineers inside the company reportedly read as a sign the program has matured. In semiconductor development, a clean bring-up cycle of that length is uncommon, since new data center processors routinely go through months of debugging before they are cleared for volume manufacturing.
Meta first showed Iris under its technical designation in March 2026, when it announced four new processor generations at once. Public documentation from the company describes the lineup as MTIA 300, 400, 450 and 500. The MTIA 300 is already running production workloads for ranking and recommendation models, while the later generations are being built to handle generative AI inference through 2027.
Broadcom Designs, TSMC Manufactures
Meta is not building Iris alone. Broadcom is assisting with the chip's design, continuing a partnership that has run through earlier MTIA generations. Taiwan Semiconductor Manufacturing Co will handle fabrication, placing Meta's silicon on the same production lines that turn out processors for nearly every major chip designer on the planet.
The arrangement mirrors what Google did with its Tensor Processing Units and what Amazon has done with its Trainium line. Hyperscale companies have concluded that designing chips tuned to their own workloads costs less over time than buying general-purpose accelerators for every task.
For Meta, the math is straightforward. The company purchases enormous volumes of graphics processing units from Nvidia and AMD, and those purchases carry premium pricing because demand across the industry far exceeds supply. Every workload Meta can shift onto its own silicon reduces its dependence on outside vendors and trims one of the largest line items in its budget.
The internal memo acknowledged friction with off-the-shelf hardware. Adopting the newest GPUs at Meta's scale, the document stated, "has been a heavy lift, and it has cost us time."
A New Chip Every Six Months
One detail in Meta's roadmap stands apart from industry norms. The company plans to release a new AI processor roughly every six months through 2027, while most chip designers work on cycles of a year or longer.
Meta's engineering blog explains how that cadence is possible. Each MTIA generation reuses modular building blocks from the previous one, so designers iterate on proven components rather than starting fresh. The chips are also built to drop into existing rack infrastructure inside Meta's data centers, which cuts the time between a finished design and hardware running real workloads.
The company says it has already deployed hundreds of thousands of MTIA chips in production and has tested the hardware against large language models, including its own Llama family.
There is a strategic logic behind the sprint. AI models change faster than traditional chip development timelines can accommodate, and a processor designed around today's workloads may be poorly matched to the models in use two years later when it finally ships. Shorter cycles keep the hardware aligned with whatever Meta's researchers are actually running.
Seven Gigawatts This Year, Fourteen the Next
The memo puts hard numbers on Meta's infrastructure ambitions. The company intends to deploy seven gigawatts of computing capacity during 2026 and plans to double that figure to 14 gigawatts in 2027.
For context, a single gigawatt is roughly the output of a large nuclear reactor. Meta is describing a computing footprint that will consume as much electricity as several million homes.
Funding that buildout requires spending on a scale the company has never attempted. Meta raised its 2026 capital expenditure guidance in April to a range of $125 billion to $145 billion, up from an earlier forecast that topped out at $135 billion. The upper end of that range is nearly double the $72.2 billion the company spent in 2025, and it exceeds the combined total of 2024 and 2025.
Chief Financial Officer Susan Li attributed the higher guidance to rising component prices and additional data center costs meant to support capacity in future years. Investors reacted uneasily at the time, sending the stock down more than 6% in after-hours trading, with analysts openly questioning when the spending would translate into measurable returns.
Meta's outlay forms a significant slice of the more than $700 billion that Big Tech collectively expects to pour into AI infrastructure this year. One recent industry estimate puts combined 2026 capital spending by Meta, Alphabet, Amazon and Microsoft at roughly $725 billion, an increase of 77% over 2025.
Supply Deals Signed Before the Shelves Empty
Building at this scale requires more than chips. The memo shows Meta has locked in long-term, multi-year supply agreements covering several categories of hardware that data centers consume in bulk.
Samsung Electronics will supply memory chips under one of those agreements. Sandisk is providing flash storage, while Sumitomo Electric of Japan covers fiber-optic equipment for networking. Sandisk declined to comment on the report, and the other two suppliers did not respond to requests from Reuters.
The timing of these contracts is no accident.
A global memory shortage has been squeezing the electronics industry for months as data center operators buy up available supply. Prices have climbed sharply enough that companies well outside the AI sector are feeling the effect, with Apple among those raising prices on products to absorb higher component costs. Morgan Stanley analysts have started using the term "chipflation" to describe the phenomenon, treating memory price increases as a macroeconomic variable rather than an industry-specific quirk.
By signing multi-year agreements now, Meta guarantees itself allocation of parts that may be difficult to obtain at any price later. During the first quarter of 2026 alone, the company added roughly $107 billion in new contractual commitments, most of them tied to cloud and infrastructure agreements running through 2027.
What Iris Means for Nvidia and AMD
The obvious question is whether Meta's chip program threatens the GPU giants that currently supply it. The short answer, for now, is no.
Iris is designed to augment Meta's Nvidia and AMD fleets rather than replace them. Custom accelerators excel at the specific, well-understood workloads they were designed for, such as serving recommendations to billions of Facebook and Instagram users or running inference on deployed models. Frontier model training, the most demanding category of AI computation, still runs overwhelmingly on Nvidia hardware across the industry.
The longer-term picture is less comfortable for merchant chip vendors. Mark Zuckerberg told investors in April that Meta is rolling out more than one gigawatt of custom silicon developed with Broadcom, alongside substantial AMD deployments that complement its newest Nvidia systems. Each gigawatt that shifts to in-house chips is revenue that never reaches an outside supplier, and Meta is one of Nvidia's largest customers.
Broadcom emerges as a quiet winner in this story. The company has positioned itself as the design partner of choice for hyperscalers building custom accelerators, with Google and Meta among its marquee clients, and every new MTIA generation extends that franchise.
If September's production run goes smoothly, Meta will have proven that a social media company can compete in one of the hardest engineering disciplines on earth, on a release schedule the traditional chip industry considers close to impossible.
Comments 0
Join the discussion and share your perspective.
Sign in to post a comment and reply to other readers.
No comments yet
Be the first to share your perspective on this article.