Meta's Next AI Chip 'Iris' Enters Production in September — 14 GW Compute Target for 2027
Meta Platforms will put its next-generation AI accelerator, internally code-named Iris, into manufacturing in September 2026, according to an internal memo reviewed by Reuters. The chip is the latest entry in Meta’s Meta Training and Inference Accelerator program, now in its fourth generation.
The production timeline arrives alongside an aggressive capacity plan: Meta will deploy 7 gigawatts of AI compute infrastructure in 2026, with a target to double that to 14 GW in 2027. That scale would place Meta’s infrastructure footprint in the range of the largest hyperscalers and ahead of most frontier AI labs.
What Iris Is
MTIA is a four-generation chip program designed to replace third-party AI compute for Meta’s internal workloads. The first two generations handled ranking and recommendation — the systems that determine what appears in Facebook and Instagram feeds. The MTIA 400, which shipped into deployment earlier this year, stepped up to generative AI inference: image generation, video synthesis, and text response workloads.
Iris is the next step. No specifications have been disclosed publicly, but the timing — production in September 2026, deployed against a 14 GW 2027 target — suggests it will form a significant part of Meta’s internal compute fleet alongside continued Nvidia GPU purchases.
The internal memo acknowledged directly that GPU adoption “has been a heavy lift and has cost us time.” That phrasing is notable from a company with Nvidia as a major supplier. It is an admission that external hardware at scale creates procurement, integration, and deployment challenges that custom silicon is meant to solve.
Why 14 GW Is a Large Number
Meta’s 7 GW 2026 target was already substantial. For comparison, the Colossus 1 campus at Memphis — currently the largest single AI compute cluster — runs at roughly 1 GW. Meta’s full-year target is therefore equivalent to deploying seven Colossus-scale campuses in a single year across its data centre footprint.
Doubling to 14 GW in 2027 signals Meta’s estimate of where compute demand is heading. Unlike API-gated AI labs that can throttle capacity to match API demand, Meta’s compute requirement is driven by serving billions of daily active users across Facebook, Instagram, WhatsApp, and Threads — a relatively predictable but large workload that scales with user growth and the AI feature surface expanding inside those apps.
The 14 GW figure also reflects Meta’s bet on on-device and in-network AI for AR glasses and the smart devices roadmap. Physical AI at consumer scale requires inference capacity that differs from cloud-served chatbot workloads; both require chips.
The Custom Silicon Thesis
Meta’s current data centre fleet runs a combination of Nvidia H100/H200 GPUs, the deployed MTIA 400, and now the incoming Iris. The MTIA 300 handles recommendation; the MTIA 400 handles generative inference; Iris is positioned to carry the 2027 load at a scale where buying 14 GW of Nvidia hardware would cost tens of billions more than manufacturing and deploying custom silicon.
That is the same economic rationale that drove Google’s TPU program, Amazon’s Trainium/Inferentia, and Microsoft’s Maia chips. The capital cost of custom silicon design is high; the per-FLOP cost at scale is substantially lower than commercial GPUs.
Meta has held off on selling AI compute externally. Unlike Google (which leases TPU time to Anthropic) or Microsoft (which runs Azure), Meta’s compute infrastructure has remained internal. The scale of the 14 GW 2027 target and the development of a next-generation chip suggest that posture is unlikely to change near-term — Meta is building for its own workloads, not as a compute landlord.
Production Risk
September is three months away. Moving a new chip design from internal testing to production manufacturing involves TSMC schedules, packaging logistics, and integration into existing rack infrastructure. The internal memo’s candid acknowledgment that prior GPU adoption “cost time” suggests Meta’s leadership is aware of the execution risk in this timeline.
The chip is one component of a broader infrastructure push. Meta’s Louisiana campus has grown to a reported 5 GW of planned capacity; Texas, Ohio, and other US states each have announced builds. The Iris production timeline needs to align with those facility completion schedules for the 14 GW target to become deployable compute rather than installed-but-idle hardware.