Meta's Iris AI Chip Enters Production in September, Targeting 14 Gigawatts of Compute by Next Year
The AI compute build-out accelerates with hyperscalers pursuing custom silicon, yet faces bottlenecks from power constraints, component shortages, and community opposition to new data centers.
The story
Meta is set to begin manufacturing its in-house AI chip, codenamed "Iris," in September, marking a significant step in the company's strategy to reduce reliance on external GPU suppliers. This move is part of Meta's ambitious plan to nearly double its computing capacity, aiming for a total of 14 gigawatts across its data centers by next year.
The "Iris" chip, a component of Meta's Meta Training and Inference Accelerator (MTIA) project, has undergone six weeks of testing with no major issues, signaling positive momentum for its internal silicon efforts. This push for custom hardware is a direct response to the massive computing costs associated with AI and the desire for greater independence from companies like Nvidia and AMD.
The production start in September, co-developed with Broadcom and manufactured by TSMC, underscores the accelerating trend among major tech firms to tailor silicon for specific AI workloads. Meanwhile, the broader AI infrastructure build-out faces headwinds, with over $130 billion in US data center projects blocked or delayed in the first quarter of 2026 due to local concerns over power and water resources. Additionally, a global shortage of Dynamic Random-Access Memory (DRAM) is expected to persist, as manufacturing shifts heavily towards High-Bandwidth Memory (HBM) essential for high-density AI data centers, with 70% of all memory chips now allocated to this segment.
Silicon
Iris (MTIA)
Maker: Meta
What: In-house AI chip for training and inference, designed to power Facebook and Instagram platforms; starts manufacturing in September.
For Whom: Meta's internal AI infrastructure
DeepSeek AI chip
Maker: DeepSeek
What: Custom AI silicon to power future models, aiming to reduce dependence on Nvidia and Huawei hardware.
For Whom: DeepSeek's own AI models
GAIA
Maker: Samsung
What: NPU-based AI accelerator for AI PCs, built on a 4nm process, optimized for on-device AI tasks.
For Whom: PC manufacturers like Lenovo and HP for AI PCs
The build-out
| Project | Who | Scale | Where |
|---|---|---|---|
| Namsskogan Data Center Lease | Bitzero Holdings Inc. (operator), OneQode (tenant) | 110 megawatts initial capacity, 15-year lease generating $2.6 billion revenue. | Namsskogan, Norway |
| Data Center Capacity Expansion | Meta Platforms | Targeting 14 gigawatts of computing power by next year, doubling current capacity. | Across Meta's data centers globally |
| Wisconsin Data Center Lawsuit | Microsoft (defendant), Wisconsin residents (plaintiffs) | Class-action lawsuit over noise from cooling systems. | Wisconsin, USA |
Supply & policy signals
Global DRAM supply shortage extending due to HBM shift; 70% of all memory chips manufactured globally routed to high-density data centers.
Implication: Increased costs and potential delays for non-AI hardware sectors, and sustained high prices for AI memory.
Over $130 billion in US AI data center projects blocked or delayed in Q1 2026.
Implication: Growing community backlash over power and water consumption is a significant bottleneck for AI infrastructure expansion in the US.
Rising capacity charges across the PJM grid in the US Rust Belt.
Implication: Manufacturers are bearing increased electricity costs due to the surge in AI data center demand.
TSMC's advanced-node and CoWoS packaging capacity remains tight.
Implication: Continued challenges for AI chipmakers to secure sufficient high-end production and packaging, pushing demand to other parts of the semiconductor supply chain.
What we'll be watching
- Meta to begin manufacturing its "Iris" AI chip in September.
- Cerebras Systems to scale its manufacturing capacity by 8 to 10 times this year.
- Huawei's planned entry into the South Korean AI accelerator market in Q4 2026 with its Ascend 950 series.
- Nvidia's next-generation Vera Rubin processors expected to debut in the second half of 2026.
- OpenAI's custom AI chips (developed with Broadcom) are scheduled for installation starting in the second half of 2026.
- TSMC's new fabrication plant in Tainan, Taiwan, slated for volume production in H1 2027.
- TSMC's second fabrication plant in Arizona, USA, reaching volume production in H2 2027.
Reporting + analyst voices: grounded via Google Search at publish time.