Disclaimer
This article is based on an internal Meta memo obtained and reported by Reuters in July 2026, along with the MTIA roadmap released in March 2026.
The internal memo included production plans for Meta’s proprietary AI chip, Iris, as well as plans to expand its AI computing infrastructure and secure long-term supplies of key components. Meta declined to comment on the report.
SanDisk also declined to comment on the reported supply agreement, while Samsung Electronics and Sumitomo Electric did not respond to Reuters’ requests for confirmation. Some details may change following future official announcements from Meta and the companies involved.
What This Article Covers
Key Takeaways
Meta reportedly plans to begin producing its proprietary AI chip, Iris, in September 2026.
According to the internal Meta memo obtained by Reuters, Iris is a data center AI chip developed as part of Meta’s MTIA, or Meta Training and Inference Accelerator, project.
Meta is designing Iris around the specific requirements of its own services. Broadcom is supporting the chip’s design, while TSMC is expected to handle manufacturing.
Although Meta is leading the chip’s requirements and design direction, it is not performing every stage of the design and semiconductor manufacturing process on its own.
The following points have been reported through the internal memo and Meta’s previous announcements:
- Meta plans to begin producing Iris in September 2026.
- Broadcom is supporting the chip’s design, while TSMC will handle semiconductor manufacturing.
- Meta is pursuing a roadmap to release new generations of MTIA chips approximately every six months through 2027.
- The company plans to secure approximately 7 GW of computing infrastructure in 2026 and expand that capacity to 14 GW in 2027.
- Meta intends to use its proprietary AI chips alongside NVIDIA and AMD GPUs rather than replacing those GPUs entirely.
- The company is expected to spend up to $145 billion on AI infrastructure in 2026.
- Meta has reportedly secured long-term supply agreements with Samsung Electronics, SanDisk, Sumitomo Electric, and other companies for memory, storage, and optical communications equipment.
The report means more than simply that another AI semiconductor is entering production.
It is important because it shows that Meta is expanding its control not only over AI models, but also over the broader infrastructure required to operate AI services, including chips, data centers, memory, storage, and networks.
Iris Is Not a Newly Introduced Chip
Iris was not first revealed in July 2026.
In March 2026, Meta introduced a four-generation proprietary AI chip roadmap covering the MTIA 300, 400, 450, and 500.
MTIA 300 is already being used for content ranking and recommendation systems within Meta’s services. Later generations are being developed to support a broader range of workloads, including generative AI and inference.
Iris was also introduced at the time as the codename for one of these processors.
The key development in the latest report, therefore, is not the existence of a new chip. It is that Iris has moved through testing and is expected to enter production in September 2026.
According to the internal memo, Iris was tested for approximately six weeks, and no major problems were found. However, Meta has not officially confirmed the chip’s detailed performance or the scale of its deployment across the company’s data centers.
Why Does This Matter?
Until now, most AI companies have built their computing infrastructure primarily around GPUs supplied by NVIDIA and AMD.
GPUs are well suited to training large AI models and processing a wide variety of workloads. However, as a company’s AI services expand, securing enough of the latest GPUs and deploying them across data centers requires significant time and money.
Meta’s internal memo reportedly noted that introducing the latest GPUs at the scale required by a company like Meta involved considerable work and caused delays.
One way to reduce these difficulties is to develop custom AI chips designed around the specific needs of a company’s services.
A general-purpose GPU is designed to process many different kinds of workloads. A custom chip, by contrast, can focus on the functions needed for repetitive tasks such as recommendation systems and AI inference.
By removing unnecessary functions and designing the chip around its own software and data center architecture, a company may be able to reduce electricity consumption and operating costs.
However, Meta’s decision to develop its own AI chips does not mean that it intends to stop using NVIDIA and AMD GPUs.
MTIA chips, including Iris, are intended to complement the large numbers of GPUs Meta purchases, not replace them completely.
GPUs can continue to handle large-scale AI model training and workloads that require greater flexibility, while Meta’s proprietary chips can process repetitive tasks specific to its own services.
Does Meta Build Its Proprietary AI Chips Entirely on Its Own?
The term “proprietary AI chip” can create the impression that Meta handles every stage of semiconductor design and production by itself.
The actual structure is different.
Meta determines the functions and architecture required for its data centers and services, while Broadcom supports the chip design. TSMC is responsible for manufacturing the finished semiconductor.
Meta’s proprietary AI chip strategy is therefore not a complete internalization of the entire semiconductor process.
It is better understood as a custom semiconductor strategy in which Meta moves beyond purchasing general-purpose GPUs and directly determines the design direction and intended use of chips built for its own services.
In April 2026, Meta extended its partnership with Broadcom through 2029. The two companies plan to jointly develop multiple generations of custom AI processors.
Broadcom’s Ethernet networking technology is also being used to connect Meta’s large-scale AI computing clusters.
AI Competition Is Becoming an Infrastructure Race
The expansion of Meta’s AI computing infrastructure is just as important as Iris itself.
According to the internal memo, Meta plans to secure approximately 7 GW of computing infrastructure in 2026 and expand that capacity to 14 GW in 2027.
No matter how advanced an AI model may be, operating it as a real-world service requires data centers and semiconductors capable of running it.
A data center requires more than AI chips. It also needs large amounts of memory, storage systems, optical communications equipment, networks, cooling systems, and a stable electricity supply.
The internal Meta memo reportedly included long-term supply agreements covering the following areas:
- Samsung Electronics: Memory semiconductors
- SanDisk: Flash storage
- Sumitomo Electric: Optical communications equipment
When Broadcom, which supports the design of Iris, and TSMC, which manufactures the chip, are added to this list, it becomes clear that Meta’s AI infrastructure is not being built by a single company acting alone.
Meta is leading the overall strategy and design while connecting semiconductor, memory, storage, and communications companies into a broader AI infrastructure supply chain.
This shows that competitiveness in the AI era does not end with securing one particular semiconductor.
The ability to secure chips and components reliably over the long term, deploy them quickly in data centers, and operate them as part of real-world services is also becoming a critical competitive advantage.
What to Watch Next
Meta is not the only company developing proprietary AI chips.
Google uses its internally developed TPUs, or Tensor Processing Units, for AI training and inference. Amazon operates Trainium for training and Inferentia for inference, while Microsoft has developed its own AI accelerator, Azure Maia.
This does not mean that NVIDIA and AMD GPUs will soon become unnecessary.
The flexibility and performance of GPUs remain important for large-scale AI models and a wide range of software workloads. Major technology companies continue to purchase large numbers of NVIDIA and AMD GPUs even as they develop their own chips.
The main change is that companies are moving away from assigning every workload to a single type of semiconductor.
They are designing AI infrastructure that combines GPUs, proprietary AI chips, CPUs, and networking equipment according to the requirements of each workload.
The following developments will be worth watching:
- Whether Iris begins production as planned in September 2026
- Which Meta services and AI workloads adopt Iris first
- Whether proprietary AI chips actually reduce electricity consumption and operating costs
- Whether Meta can secure 14 GW of computing infrastructure by 2027
- Whether the long-term supply agreements described in the internal memo are officially confirmed by the companies involved
- How Meta divides workloads between GPUs and MTIA chips
DANA NOTES Commentary
Many people think of AI competition as a race to build more advanced models.
For companies that operate AI services, however, operating costs and the ability to secure infrastructure are becoming just as important as model performance.
As the number of users increases, companies need more semiconductors, memory, storage, networking equipment, and electricity.
Meta’s Iris should therefore not be viewed simply as a new semiconductor designed to deliver higher performance.
It is better understood as part of an infrastructure strategy in which Meta uses custom chips for repetitive workloads, diversifies its reliance on GPUs, and seeks to operate AI services more reliably and at a lower cost.
However, developing proprietary AI chips does not make Meta completely independent from the semiconductor supply chain.
Meta still needs NVIDIA and AMD GPUs. Iris also depends on Broadcom’s design support and TSMC’s manufacturing capabilities. Memory, storage, and optical communications equipment must also be supplied by external companies.
Ultimately, the next stage of AI competition is unlikely to be a race in which every company attempts to build every technology on its own.
Instead, it is likely to become a competition over how efficiently companies can design and connect the entire infrastructure stack—from AI models and chips to data centers, electricity, and component supply chains.
