Code named Iris, the chip is the first part from Meta's MTIA (Meta Training and Inference Accelerator) program to reach manufacturing after five years, putting the company alongside Google, Amazon, and Microsoft as a custom silicon operator.
Meta is set to begin manufacturing its in-house AI chip "Iris" in September at Taiwan Semiconductor Manufacturing Co. (TSMC), according to an internal memo reviewed by Reuters, accelerating a five-year effort to bring more of its AI compute in-house and formally placing Meta alongside Google, Amazon, and Microsoft as the fourth hyperscaler with a custom AI accelerator in volume production.
Iris is the third generation of Meta's MTIA, or Meta Training and Inference Accelerators, a program the company launched more than five years ago to reduce its dependence on Nvidia and AMD's graphics processing units (GPUs). The memo, first reported by Reuters and confirmed by CNBC, TechCrunch, and The Standard HK, says chip testing wrapped up in six weeks without major problems. Earlier MTIA generations had a bumpy public track record, with delays and retooling at prior stages, so a clean six-week validation window is the kind of internal signal that suggests the program has steadied.
The chip is built primarily for inference, the work of running the recommendation and ranking models that decide which Facebook post a user sees next and which Instagram reel to surface, rather than the largest training workloads that still run on Nvidia. The memo explicitly frames Iris as a complement to Meta's existing GPU fleet, not a replacement. Meta has been a top-three buyer of Nvidia hardware for years, and even after Iris reaches mass production the company plans to keep buying Nvidia and AMD accelerators in parallel.
The timing is the story. AI infrastructure has become one of the largest line items in Big Tech capital budgets, and Meta's 2026 capex run-rate is measured in tens of billions tied to AI compute. The shift across the sector is from GPU-only buying toward a mix of merchant and custom silicon, and Meta now sits in the second camp. Earlier this year Meta acknowledged in an earnings call that adopting the latest GPUs had been "a heavy lift" and had "cost us time", the company's own framing for why an in-house path matters now. Emarketer analyst Jacob Bourne has put the tension in plainer terms for investors: Meta needs AI to grow ad revenue fast, but custom-silicon savings arrive years after the upfront spend, which is why the September production start matters as a capex signal rather than a near-term margin one.
Meta is also the last of the four hyperscalers to put a custom AI accelerator into volume production. Google has shipped its TPU (Tensor Processing Unit) family for nearly a decade. Amazon has both Trainium for training and Inferentia for inference. Microsoft has been deploying its Maia chip. Iris closes the structural gap that left Meta buying from the outside. Google has years of compiler, model, and workload tuning built on TPU — a depth of optimization that custom silicon adopters typically build over time rather than at launch.
The production start is the milestone the MTIA program has been building toward since 2021, but the memo does not disclose Iris's process node, performance benchmarks against Nvidia's current generation, or yield assumptions. The memo language leaves each of those as a question for later disclosure. Whether Iris lowers Meta's cost-per-inference at scale and survives contact with Nvidia's next GPU generation is what the next year of disclosures will reveal.