Meta Platforms will begin manufacturing its first in-house AI training and inference accelerator, code-named Iris, in September 2026, marking the company’s most ambitious push yet to break its dependence on Nvidia and AMD for the compute that runs Facebook and Instagram. The chip is the inaugural product of a four-generation roadmap for the Meta Training and Inference Accelerator (MTIA) family, designed in partnership with Broadcom and fabricated by Taiwan Semiconductor Manufacturing Company (TSMC). Internal testing of Iris completed in just six weeks with no major issues, according to an internal memo reviewed by Reuters and reported by CNBC.
The production timeline, disclosed to staff in early July, is part of Meta’s broader plan to scale its data-center compute footprint from seven gigawatts in 2026 to fourteen gigawatts in 2027 — a doubling of capacity aimed squarely at the company’s recommendation, ranking, and generative-AI workloads. Iris will augment, rather than replace, the large fleet of Nvidia and AMD GPUs Meta already operates, with executives framing the in-house silicon as a path to lower inference cost and greater control over its AI infrastructure stack.
Six weeks of testing, four chips through 2027
Meta unveiled Iris under its technical name in March alongside three other AI processors. The company plans to release a new MTIA chip roughly every six months through 2027 — a cadence far faster than the typical industry release cycle of twelve to eighteen months. The accelerated schedule is intended to feed the company an internal supply of inference silicon as the cost of running recommendation models across its 3 billion-plus users continues to climb.
The six-week testing window is itself a meaningful data point. Custom silicon programs at peer companies — including Microsoft’s Maia and Amazon’s Trainium — typically spend twelve to eighteen months in bring-up before tape-out, with multiple re-spins common. Iris’s clean first-pass result, attributed in the internal memo to a tighter co-design relationship with Broadcom, suggests Meta has spent the past two chip generations investing in design methodology rather than just transistor count.
Why in-house, why now
- Cost. Inference at Meta’s scale is now the single largest line item in its AI capex, and executives have signaled that custom silicon can lower per-query cost by reducing reliance on Nvidia’s premium-priced GPUs.
- Independence. Designing its own accelerators gives Meta control over features tailored to ranking and recommendation workloads, which behave differently from the large language model training runs that dominate Nvidia’s customer base.
- Supply. Allocating a portion of capacity to in-house silicon hedges against future GPU shortages, which have periodically throttled Meta’s ability to deploy new models at full scale.
“Adopting the latest GPUs at a firm as large as Meta has been a heavy lift, and it has cost us time,” the internal memo stated, according to CNBC’s report. The acknowledgment underscores how the in-house track is now treated as essential, not experimental.
Broadcom’s three-rival design business
The Iris announcement is the second major signal in two weeks that Broadcom has become the central design house for custom AI accelerators aimed at taking share from Nvidia. The company now counts Google, Meta, Anthropic, and OpenAI as customers for its application-specific integrated circuit (ASIC) design services, a roster that essentially covers three of the largest Nvidia rivals in frontier model training and inference.
OpenAI unveiled its first custom AI inference chip, Jalapeño, built with Broadcom, in late June. Anthropic has been working with Broadcom on its own accelerator program. Google has been shipping its Tensor Processing Units internally for years, with Broadcom involved in the networking and packaging layer. Meta’s Iris is the latest addition to what is effectively a four-horse consortium betting that ASICs, not general-purpose GPUs, will win the next decade of inference economics.
Broadcom’s custom-AI revenue, which management has guided to roughly $20 billion for fiscal 2026, has become the single fastest-growing segment of the company’s semiconductor business. Wall Street analysts now value the AI-accelerator business alone at more than several of Broadcom’s traditional networking acquisitions combined.
What’s next for Meta’s chip roadmap
- September 2026 — Iris enters production at TSMC.
- Early 2027 — second-generation MTIA chip expected, with broader feature coverage for generative-AI inference.
- Late 2027 — fourth MTIA generation targeted, with cumulative compute capacity sufficient to power a significant share of Meta’s recommendation systems.
What it means for Nvidia
Broadcom’s growing custom-chip franchise has been the most consistent proof point that Nvidia’s training-ASIC moat is narrower than the company’s stock multiple suggests. UBS reaffirmed a Buy rating on Broadcom earlier this month, citing rising AI ASIC demand from OpenAI, Anthropic, and Meta. Nvidia, which still controls the bulk of the training compute market, has begun publishing detailed total-cost-of-ownership comparisons to defend its pricing — an acknowledgment that ASIC competition is now a structural, not cyclical, threat.
For Meta, Iris is the first credible signal that the company’s six-year, often-rewritten in-house silicon program is finally shipping at scale. The six-week test cycle and clean manufacturing hand-off to TSMC suggest that the next two MTIA generations will arrive on schedule. If they do, Meta will be the first hyperscaler to operate its recommendation systems at meaningful scale on its own silicon — a milestone that has eluded every other big-tech custom-chip effort to date.
The Iris chip, Meta’s first internally designed accelerator to enter mass production, is the opening move in a four-generation roadmap that will reshape the company’s AI economics through 2027.