Meta Platforms is done waiting in line for someone else's silicon. According to an internal memo reviewed by Reuters and reported on July 9, the company will begin manufacturing its first data-center AI chip in September, an in-house processor code-named "Iris" that sits at the heart of an audacious plan to more than double Meta's computing capacity to 14 gigawatts by 2027.
The timing matters. For a company that has spent more than half a decade struggling to field competitive custom silicon, moving Iris into production is the clearest signal yet that Mark Zuckerberg's chip ambitions have finally cleared the lab. The memo revealed that testing the chip took just six weeks and turned up no major issues, a pace that insiders read as unusually smooth for an effort that has, in the memo's own words, "floundered" since it began.
The chip and the roadmap
Iris is not a one-off. It is one of four generations of processors under Meta's Training and Inference Accelerators program, known internally as MTIA. Meta unveiled the chip under its technical name in March alongside three other AI processors, and expanded its partnership with Broadcom in April to bring the designs to production. Broadcom is the design partner; Taiwan Semiconductor Manufacturing Co., the world's dominant advanced foundry, will manufacture the chips.
The cadence is aggressive by industry standards. Meta plans to launch a new chip roughly every six months through 2027, whereas most firms release AI accelerators at intervals of a year or more. The processors are tailored to Meta's own workloads: the ranking and recommendation systems that decide what surfaces in an Instagram or Facebook feed, plus the generative-AI tasks powering chatbots and content tools.
Crucially, Iris is meant to augment rather than replace the graphics processing units Meta buys in enormous quantities from Nvidia and Advanced Micro Devices. The memo was candid about why an alternative is needed: adopting the newest GPUs at Meta's scale "has been a heavy lift, and it has cost us time." Custom silicon, the company believes, will lower its massive computing costs and reduce its dependence on outside suppliers.
A $145 billion year
The chip is only one line item in a staggering infrastructure budget. Meta expects to spend as much as $145 billion on AI infrastructure this year alone, a substantial slice of Big Tech's projected $700 billion-plus outlay on the technology in 2026. This year the company plans to deploy seven gigawatts of computing infrastructure, then double that figure in 2027 to reach the 14-gigawatt target.
To feed that expansion, Meta has locked in long-term supply agreements across a strained components market, including deals with Samsung Electronics for memory chips, Sandisk for flash storage and Sumitomo Electric for fiber-optic equipment. Those arrangements have become critical amid a memory-chip shortage severe enough that Morgan Stanley analysts have flagged "chipflation" as a genuine macroeconomic concern, with prices rising fast enough to push firms such as Apple to raise their own.
The compute buildout has a physical face, too: the multi-gigawatt data centers Zuckerberg has been touting for a year. Prometheus, a roughly one-gigawatt supercluster in New Albany, Ohio, is slated to come online in 2026, while Hyperion, a Louisiana campus reportedly costing around $10 billion, is designed to scale toward five gigawatts. "We're building multiple more titan clusters as well," Zuckerberg wrote in announcing the projects. "Just one of these covers a significant part of the footprint of Manhattan." Meta declined to comment on the chip memo itself.
Why It Matters
Meta is running the vertical-integration playbook that Google, Amazon and Microsoft have already validated. Google has shipped TPUs for years, Amazon has its Trainium and Inferentia lines, and Microsoft has Maia. What each hyperscaler has concluded is that leaning entirely on Nvidia is a strategic vulnerability, not a convenience: Nvidia's margins are effectively a tax on every AI ambition, and its supply is finite.
Building Iris does not mean Meta stops buying from Nvidia overnight, and it won't. But every generation of MTIA silicon that successfully carries production traffic chips away at Meta's marginal dependence on external suppliers and gives it leverage in the next negotiation. The bigger prize is cost per unit of compute: at 14 gigawatts, even single-digit efficiency gains translate into billions of dollars and, potentially, faster model iteration.
The clearest beneficiaries beyond Meta are its partners. Broadcom, which has positioned itself as the go-to design shop for hyperscaler custom chips, gains another marquee revenue stream to sit alongside its Google work. And TSMC captures value no matter whose name is stamped on the die, because it remains the only foundry that can produce leading-edge chips at scale. The party most exposed is Nvidia, whose largest customers are quietly building the off-ramp.
What to Watch
The September production start is the concrete milestone. If Iris rolls off TSMC's lines on schedule and performs in Meta's data centers, it validates the entire four-generation MTIA roadmap and the tens of billions being spent to execute it. Watch for whether Meta hits its seven-gigawatt 2026 deployment target, whether Prometheus comes online on time, and whether the six-month chip cadence holds. And watch Nvidia's next earnings call for any hint that hyperscaler custom silicon is finally bending the demand curve.
"We're building multiple more titan clusters as well. Just one of these covers a significant part of the footprint of Manhattan."— Mark Zuckerberg, CEO, Meta Platforms