Iris and Beyond: Inside Meta's Custom AI Chip Strategy

Meta is accelerating its push to reduce reliance on external chip suppliers with a new in-house AI accelerator codenamed "Iris," part of an aggressive four-generation roadmap that spans just two years. Here's what's known, what's planned, and what it all means.

What Is Meta's Iris Chip?

"Iris" is the internal code name for the MTIA 400, the third generation of Meta's Meta Training and Inference Accelerator (MTIA) family. It is a custom data center chip designed primarily for inference — running already-trained AI models — rather than training them from scratch . The chip is one of four successive MTIA generations (MTIA 300, 400, 450, and 500) that Meta is deploying to handle ranking, recommendations, generative AI, and general AI workloads at massive scale .

When Will Iris Go Into Production?

Meta plans to start manufacturing the Iris chip in September 2026, according to an internal memo reviewed by Reuters. The memo states that bug-testing has been completed, clearing the path for production .

How Fast Does Meta Plan to Release New Chips?

Meta is pursuing a roughly six-month release cadence for its MTIA chip generations — significantly faster than the typical industry pace of a year or more . The full roadmap across 2026 and 2027:

• MTIA 300 — Already deployed and operational, powering ranking and recommendation systems for Facebook and Instagram .
• MTIA 400 ("Iris") — Slated for production in September 2026, designed for gen AI inference .
• MTIA 450 ("Arke") — Scheduled for mass deployment in early 2027, engineered for generative AI inference .
• MTIA 500 ("Astrid") — Scheduled for mass deployment in 2027 .

Meta says it has "accelerated MTIA development" across these four generations in roughly two years .

What Are Meta's Computing Capacity Targets for 2026 and 2027?

• 2026 target (internal memo): Deploy 7 gigawatts (GW) of computing infrastructure this year .
• 2027 target: Reach 14 GW of overall computing power by next year .
• Morgan Stanley model: Meta is expected to add roughly 2 GW of owned, operated IT capacity in 2026 and 3.5 GW in 2027, starting from a baseline of about 3 GW at the end of 2025 .
• Earlier public target: Meta previously aimed for more than 10 GW of total capacity by the end of 2026 .

The 14 GW 2027 figure from the internal memo is the most current and aggressive target .

Who Are Meta's Partners in Designing and Manufacturing Iris?

• Broadcom — Co-developer and design partner for Meta's MTIA chips. An expanded agreement announced April 14, 2026 extends the partnership through 2029, covering chip design, advanced packaging, and multiple chip generations. The initial commitment covers compute capacity exceeding 1 GW, with multi-GW scale-up planned .
• TSMC (Taiwan Semiconductor Manufacturing Company) — The manufacturer that will fabricate the Iris chip .

How Does Iris Relate to Meta's Existing GPU Purchases From Nvidia and AMD?

Iris is designed for inference workloads, not training. Meta continues to buy Nvidia GPUs (e.g., H100/H200 families and next-gen Blackwell) and AMD GPUs for AI training workloads, where raw parallel compute is critical . The in-house MTIA chips are meant to supplement rather than immediately replace those GPU purchases — handling the massive and growing inference load (ranking, recommendations, serving AI models) at lower cost and higher efficiency. Meta's strategy is to diversify its silicon base, not cut off GPU vendors entirely.

What Is Meta's New Cloud Computing Initiative 'Meta Compute'?

"Meta Compute" is a two-part initiative:

1. Internal infrastructure division — Announced January 12, 2026, Meta Compute is a new division co-led by Santosh Janardhan (Meta's head of global infrastructure) and Daniel Gross. It manages Meta's global data center network and builds gigawatt-plus-scale AI clusters .
2. Commercial cloud business — First reported by Bloomberg on July 1, 2026, Meta is developing plans to sell excess AI computing capacity to external customers, offering both hosted AI model access and raw GPU compute. This would put Meta in direct competition with AWS, Microsoft Azure, Google Cloud, and "neocloud" providers like CoreWeave .

The plans are still in development and the strategy is not yet finalized .

How Does This Cloud Move Affect Meta's Stock and Competitors Like CoreWeave?

• Meta stock: The news was received positively. Meta shares rose nearly 9% on July 1, 2026, the day the cloud plans were confirmed . The move signals to investors that Meta has a path to monetize its enormous AI infrastructure spend rather than letting it sit idle .
• CoreWeave and neocloud competitors: The cloud move is a direct threat to companies like CoreWeave that sell raw AI compute capacity. Bloomberg reported that Meta may "copy CoreWeave's business model" . However, some analysts (including SemiAnalysis) argue that fears of Meta crashing the neocloud market are "erroneous" — the sheer scale of Meta's own internal demand means excess capacity for sale may be limited, and the market for AI compute is growing so fast that new entrants can coexist .

What Are Meta's Projected Capital Expenditure Figures for 2026?

Meta's 2026 capital expenditure guidance has been updated:

Period	Capex Range
Q1 2026 earnings guidance	$115 billion – $135 billion
July 1, 2026 updated guidance	$125 billion – $145 billion

The increase from $115–135B to $125–145B reflects "higher component pricing this year and, to a lesser extent, additional data center costs to support future growth" . For comparison, Meta spent approximately $72.2 billion on capex in 2025 . Research firm SemiAnalysis projects that Meta's 2027 capex will be "shockingly large" — even higher than 2026 — as contracted data center capacity continues to come online .