Full Breakdown
Meta Begins Mass Production of Its “Iris” AI Chip
7/13/2026, 11:33:35 AM
Launch of the Iris Accelerator
Meta Platforms will start mass-producing its fourth-generation AI accelerator, code-named Iris, in September 2026. The chip is being co-designed with Broadcom and fabricated by Taiwan Semiconductor Manufacturing Co. (TSMC) after a six-week bug-testing phase that uncovered no major issues. Iris is the first of four planned generations under Meta’s Meta Training and Inference Accelerator (MTIA) program, with a new version slated for release roughly every six months through 2027. The company says the chip will complement, not replace, the Nvidia and AMD GPUs it currently purchases for ranking, recommendation and generative-AI workloads across Facebook and Instagram.
Context and Strategic Rationale
Meta’s internal memo shows the firm targeting 7 gigawatts (GW) of AI-computing capacity in 2026 and 14 GW by 2027—enough electricity to power roughly 800,000 homes per GW. Achieving this scale with Nvidia GPUs alone would require substantially larger cash outlays. The memo also notes a $125-$145 billion capital-expenditure plan for AI infrastructure this year, nearly double the $72 billion spent in 2025. Industry peers such as Google (TPUs), Amazon (Trainium/Inferentia) and Microsoft (Maia) have pursued similar in-house silicon strategies, positioning Meta’s move as part of a broader “chip-independence” trend among hyperscalers.
Principal Actors
- Mark Zuckerberg, CEO of Meta, who highlighted the quarter’s strong performance and the launch of a model from the company’s Super Intelligence Lab.
- Mike Gualtieri, vice-president and principal analyst at Forrester, commenting on the competitive need for proprietary chips.
- Broadcom (design partner) and TSMC (manufacturer).
- Samsung Electronics, SanDisk, and Sumitoro Electric, suppliers of memory, flash storage and fiber-optic equipment under long-term agreements.
- Justin Post, analyst at BofA Securities, and Cody Acree, analyst at Benchmark Research, who evaluated the cost implications.
Key Numbers
- 7 GW of AI compute deployed in 2026; 14 GW projected for 2027.
- 1 GW added in H1 2026, 2.5 GW expected by year-end.
- Capital spend: $125-$145 billion this year; $22 billion per GW of capacity (vs. earlier $45 billion estimates).
- Production cadence: a new chip roughly every six months through 2027.
Why It Matters
By shifting a portion of its AI workload to custom ASICs, Meta aims to lower per-gigawatt costs, reduce reliance on Nvidia’s high-priced GPUs, and improve margins on ad-ranking and recommendation models. Faster, cheaper compute could enable more granular real-time personalization of feeds and ads, potentially increasing revenue per user. The scale of the build-out also underscores the physical limits of AI expansion, as power and cooling requirements become a strategic constraint.
Official Statements & Responses
Meta declined to comment on the memo. Samsung, SanDisk and Sumitomo Electric did not respond to interview requests. In earnings remarks, Zuckerberg said the quarter “was an important milestone for us” and highlighted the launch of a new Super Intelligence Lab model.
Criticism & Opposition
Analysts note that Iris is positioned as a supplement to existing GPUs, so Nvidia’s data-center revenue is unlikely to be displaced in the near term. Privacy advocates raise concerns that greater compute capacity may enable more extensive behavioral profiling, yet Meta has not disclosed new data-use policies linked to Iris.
Conflicting Reports & Gaps
Cost-per-gigawatt estimates differ: BofA’s $22 billion figure contrasts with earlier internal projections of $45 billion per GW. The memo provides no timeline for when Iris-enabled improvements will be visible to end users, and the impact on feed algorithms remains speculative.
Verbatim Quotes
- “You can't become an AI titan if you are dependent on another company for chips,” — Mike Gualtieri, Vice President and Principal Analyst, Forrester
- “If this data is close to reality, it means Meta has successfully driven the build-out cost per megawatt (MW) of AI computing capacity well below widely held market estimates.” — Justin Post, Analyst, BofA Securities
- “Benchmark Research analyst Cody Acree commented on this: "While Nvidia's share within Meta may decline as Meta continues to expand its AI build-out, overall AI capital expenditure from major cloud service providers is still expected to more than double.” — Cody Acree, Analyst, Benchmark Research
- “Cramer stated: "We just need one heavyweight company to say its AI business is now making money, and then you can forget about holding commodity semiconductor stocks.” — Jim Cramer, Financial Program Host
- “This quarter was an important milestone for us. Our family of apps continues to show strong growth momentum, and we also launched the first model from our Super Intelligence Lab.” — Mark Zuckerberg, CEO, Meta
What’s Next
The memo confirms that mass production of Iris begins in September 2026, with subsequent chip generations slated for release every six months through 2027. Meta also plans to continue expanding its AI-compute footprint to meet the 14 GW target, while exploring the possibility of leasing excess capacity to external customers.
