Meta plans to start manufacturing its in-house AI chip, codenamed “Iris,” in September, according to an internal memo seen by Reuters. The goal is to cut GPU costs and loosen its reliance on Nvidia amid a component shortage. It’s part of a fast-moving chip program aiming to double Meta’s computing capacity by 2027.

Key Takeaways

• Meta will start producing its “Iris” AI chip in September
• The chip cleared bug testing in about six weeks with no major issues
• Broadcom co-designs it, while TSMC handles manufacturing
• The goal is cutting reliance on Nvidia and AMD GPUs
• Meta aims to reach 14 gigawatts of compute by 2027

What the Memo Reveals

The report is a Reuters exclusive. Citing an internal memo, Reuters says Meta is on track to start making the latest versions of its AI chip in September, in a bid to lower its GPU costs amid an unprecedented component shortage.

The chip has a name and a lineage. Codenamed Iris, it’s a data-center chip that falls under Meta’s Meta Training and Inference Accelerator, or MTIA, an in-house silicon line the company has been scaling as part of a record spending push.

The testing news is the standout detail. According to the memo, testing the chip took only six weeks and found no major issues, quick progress that signals positive momentum for an effort that had floundered since its launch more than half a decade ago.

The sourcing is worth noting. Meta declined to comment on specifics, and both the September timing and the capacity target come from the memo and anonymous sources rather than any public disclosure, so the details are reported rather than officially confirmed.

The Supply Chain Behind Iris

Building a chip at this scale takes a roster of partners, and Meta has lined up several. Broadcom is serving as the design partner, while Taiwan Semiconductor Manufacturing Company handles fabrication.

The component list runs deeper. Meta is buying RAM from Samsung, storage from Sandisk, and fiber-optic equipment from Sumitomo Electric, according to the report, assembling a supply chain that spans several of the industry’s biggest names.

The Broadcom relationship is long-term. Their partnership now runs through 2029 and covers several generations of custom silicon, and Broadcom has said the newer MTIA parts will be among the first custom AI chips built on a 2-nanometer process, a cutting-edge node.

That mix matters strategically. By sourcing memory, storage, and networking from specialized suppliers while owning the chip design itself, Meta keeps control of the part that matters most to its workloads without having to build everything from scratch.

The Aggressive Chip Cadence

Iris isn’t a one-off. Meta unveiled four new MTIA chips in March, the 300, 400, 450, and 500, some already in deployment and others due this year or next.

The release schedule is unusually fast. Meta plans to ship a new chip roughly every six months through 2027, a far more aggressive pace than the annual or slower cycles common across the chip industry.

The design philosophy is built for change. Meta is taking a modular approach, anticipating that its needs will shift as AI evolves rapidly between now and when the chips reach production, so the architecture is meant to flex rather than lock in.

It’s also a turnaround story. After years as a side project that struggled, Meta’s silicon effort has shifted into a core plank of its infrastructure strategy, and the swift six-week test result is being read as proof the program has finally found its footing.

Why Meta Is Building Its Own Silicon

The motive is money and independence. Custom chips are expected to help Meta save on buying GPUs from Nvidia and AMD, though the company still expects to spend heavily with those providers too.

Iris is a supplement, not a replacement. Meta intends to use the MTIA chips to train the ranking and recommendation algorithms behind Facebook and Instagram, handle broader AI workloads, and run inference for its apps, offloading work rather than fully cutting the cord.

The scale explains the urgency. Meta runs recommendation systems for more than 3 billion daily users, and every workload it can move onto its own chips is one it doesn’t have to buy at Nvidia’s margins.

The company was candid about the pain. Adopting the latest GPUs at Meta’s scale has been a heavy lift and has cost the company time, the memo said, underscoring why owning more of the stack is so appealing.

The Massive Spending Context

The chip push sits inside an enormous budget. Meta said in April it expects capital expenditures between $125 billion and $145 billion this year, much of it flowing toward AI.

The capacity goals are just as large. The memo outlined a two-step expansion, with seven gigawatts of computing capacity coming online in 2026 and growing to 14 gigawatts by 2027, effectively doubling Meta’s compute.

The spending spills well beyond chips. Meta has been striking data-center and power deals around the world, committing tens of billions to secure the capacity it needs to train and deploy AI efforts like its Muse Spark series.

Why It Matters

Meta is now firmly in the custom-silicon race. Google, Amazon, OpenAI, and Anthropic have all pursued their own chips or chip deals, and Meta’s accelerating cadence puts it among the most aggressive of the group.

The bigger theme is dependence. The entire industry is trying to reduce its reliance on Nvidia, and Meta’s ability to move real workloads onto Iris will be a test of whether in-house silicon can meaningfully dent that grip, even as Nvidia’s overall position stays dominant.

Investors are watching closely. Meta shares slipped around 3.5% in premarket trading after the report, a reminder that the market is weighing the payoff of this spending against its staggering cost. If Iris performs and the six-month cadence holds, Meta’s long-struggling chip bet may finally start earning its keep.

Digital Trendings is your trusted source for AI news and updates, stay tuned for more.