Meta could begin production of its Iris AI inference chip in September 2026, according to a report from @DCDnews. The in-house chip targets data center inference workloads, marking a step in Meta's push to reduce reliance on external suppliers like NVIDIA.
Key facts
- Iris production may start September 2026.
- Chip targets inference workloads in Meta data centers.
- Meta spent $37B on capex in 2025.
- Iris is part of Meta's MTIA family.
- No foundry partner or volume disclosed.
Meta could start production of its Iris AI inference chip in September 2026, according to a report from @DCDnews. The chip is designed for inference workloads in Meta's data centers, which handle tasks like running large language models for its platforms.
Iris is part of Meta's broader push to reduce reliance on external chip suppliers, particularly NVIDIA, which dominates the AI accelerator market. Meta previously disclosed its custom silicon efforts, including the Meta Training and Inference Accelerator (MTIA) family, but Iris represents a more targeted inference chip.
The report did not specify a foundry partner or manufacturing volume. Meta has historically worked with TSMC for its custom chips, but no confirmation was provided. The timeline suggests Meta is accelerating its custom silicon roadmap, likely to improve cost efficiency and performance for its massive AI workloads.
Meta's data center spending has surged, with capital expenditures reaching $37 billion in 2025, per its Q4 2025 earnings. Custom silicon could help Meta reduce per-inference costs, which are critical as it scales AI features across Facebook, Instagram, and WhatsApp.
How Iris fits into Meta's chip strategy
Meta's MTIA program already produces training and inference accelerators, but Iris appears to be a specialized inference chip. This mirrors moves by Google (TPU), Amazon (Trainium/Inferentia), and Microsoft (Maia), all of whom have developed custom silicon to optimize AI workloads and reduce NVIDIA dependency.
Iris is expected to be deployed in Meta's data centers for inference tasks, such as serving its Llama family of large language models. The September production start suggests volume deployment could begin later in 2026 or early 2027.
What remains unknown
Key details are missing: the chip's performance metrics (e.g., TOPS, power efficiency), the foundry node (likely 5nm or 3nm), and the total investment. Meta has not commented publicly on the report. The company's next earnings call, expected in late April 2026, may provide more details on its custom silicon roadmap.
Key Takeaways
- Meta could produce Iris AI chip in September 2026 for data center inference, per report.
- Reduces NVIDIA reliance.
What to watch
Watch for Meta's Q1 2026 earnings call in late April for official confirmation or details on Iris. Also track TSMC's capacity allocation for custom chips and any NVIDIA response to hyperscaler custom silicon momentum.
[Updated 10 Jul via gn_dc_power]
Separately, Meta plans to build a C$13 billion (US$9.17 billion) gigawatt-scale data center in Alberta, Canada, its first in the country [per Reuters and Data Center Dynamics]. The facility would support Meta's growing AI inference needs, potentially housing the Iris chips once production begins. The Alberta site underscores Meta's aggressive infrastructure expansion to power its custom silicon strategy.
