Meta is stepping up its semiconductor ambitions in a major way. According to a report from Reuters, the social media giant plans to begin manufacturing its next-generation custom AI chip — codenamed “Iris” — as early as September 2026. This strategic move is designed to significantly reduce Meta’s heavy dependence on dominant GPU makers Nvidia and AMD for AI computing power.
The news comes at a pivotal time as the global AI arms race intensifies, with every major tech company scrambling to secure its own chip supply amid soaring demand and persistent GPU shortages.
What Is the Iris Chip?
Iris is the latest addition to Meta’s growing family of Meta Training and Inference Accelerators (MTIA). These custom silicon chips are purpose-built for Meta’s unique AI workloads — everything from ranking and recommendation algorithms to generative AI inference powering features across Facebook, Instagram, WhatsApp, and Threads.
Unlike off-the-shelf GPUs that are designed for general-purpose computing, MTIA chips are finely tuned to Meta’s specific software stack, delivering far greater compute efficiency and cost-effectiveness at Meta’s unprecedented scale.
Why Custom Silicon Matters
Meta already operates hundreds of thousands of MTIA chips across its global data centers. By bringing chip design in-house, the company achieves several critical advantages:
- Higher compute efficiency — custom architecture precisely matched to Meta’s workloads eliminates the overhead of general-purpose designs
- Dramatic cost savings — reduced procurement of premium-priced Nvidia H100/B200 and AMD MI300X GPUs
- Blistering iteration speed — new chip generations every six months, far outpacing the industry standard of 1-2 years
- Full-stack optimization — native support for PyTorch, vLLM, Triton inference server, and Open Compute Project standards
The MTIA Family Roadmap
Meta’s custom silicon portfolio now spans multiple generations with specific workload targets:
- MTIA 300 — Already in production, focused on ranking and recommendations training workloads
- MTIA 400, 450 & 500 — Designed for GenAI inference and capable of handling all workload types, expected through 2027
- Iris (new) — Next-generation chip entering manufacturing in September 2026, further expanding Meta’s in-house compute capacity
Meta has committed to a remarkable cadence of releasing new custom chips every six months. This “rapid, iterative development” approach, as the company describes it, allows Meta to quickly adapt to the fast-evolving AI landscape while keeping costs under control.
Notably, Meta’s strategy flips the traditional chip design approach on its head. While most chip makers optimize first for large-scale GenAI pre-training (the most demanding workload), Meta’s MTIA chips are optimized first for inference — the process of running already-trained AI models to serve predictions and responses to billions of users. This “inference-first” approach keeps the chips perfectly tuned to Meta’s dominant use case.
Joining the Custom Chip Club
Meta is hardly alone in this strategy. The company joins an elite group of tech giants developing their own AI accelerators:
- Google — Tensor Processing Units (TPUs), now in their sixth generation
- Amazon — Trainium for training and Inferentia for inference
- Microsoft — Maia AI Accelerator, announced in late 2023
- Apple — Neural Engine integrated into A-series and M-series chips
Each of these companies has recognized that reliance on merchant silicon suppliers creates both supply chain risk and cost inefficiency at AI scale. According to Reuters, internal Meta documents show the company aims to double its computing capacity through Iris and other custom chips, dramatically scaling AI infrastructure without proportionally increasing expenses.
The Bigger Picture
The timing is no coincidence. Nvidia’s near-monopoly on AI training GPUs has created a market where demand far outstrips supply, with lead times stretching months and prices remaining astronomical. AMD’s MI300X has emerged as a credible alternative, but both companies’ products are general-purpose accelerators that carry overhead Meta doesn’t need. Meanwhile, geopolitical tensions around semiconductor supply chains — particularly the ongoing US-China tech rivalry — add another layer of urgency for American tech companies to secure their own silicon production.
Meta’s Iris chip represents a significant step toward vertical integration in AI infrastructure, a trend that’s reshaping the entire technology landscape. As custom silicon becomes a competitive necessity rather than a luxury, the companies that can design, manufacture, and deploy their own AI chips at scale will hold a decisive advantage in the race toward artificial general intelligence.
Frequently Asked Questions
When will Meta’s Iris chip go into production?
Manufacturing is expected to begin in September 2026, according to Reuters. Deployment in Meta’s data centers will follow after the production ramp-up phase.
Will Meta sell Iris chips commercially?
No. Like Google’s TPUs and Amazon’s Trainium chips, Meta’s MTIA family — including Iris — is designed exclusively for internal use. They will power AI workloads across Meta’s family of apps and services, not be sold to third parties.
Will Meta stop buying Nvidia GPUs altogether?
Not immediately. Meta takes what it calls a “portfolio approach” — sourcing silicon from multiple vendors while keeping MTIA at the center of its AI infrastructure. Nvidia and AMD will continue to supply Meta for certain workloads, but custom chips will handle a growing share over time as the MTIA roadmap matures.