Meta's Iris AI Chip to Enter Production in September, Targeting 14 GW Computing Power

Introduction

Meta plans to begin manufacturing a new custom artificial intelligence chip in September 2026, according to reports, a move that will accelerate one of the largest computing infrastructure expansions in the tech industry.

The chip, codenamed Iris, is part of Meta's in-house Meta Training and Inference Accelerator (MTIA) project. According to an internal memo reviewed by Reuters, Iris has completed a six-week testing cycle with no major issues identified and is now entering the production phase.

The chip is only part of a broader plan. Meta expects to deploy approximately 7 gigawatts of computing infrastructure in 2026 and plans to double total capacity to 14 gigawatts by 2027. Its 2026 capital expenditure guidance has also been raised to between $125 billion and $145 billion, reflecting higher component prices and additional data center investments.

Source Note: This is an original, directly publishable Chinese article based on a Reuters report, republished by The Economic Times, and cross-verified with Meta's official infrastructure announcements. It is not a line-by-line rewrite of copyrighted sources. The source article contained one relevant primary image, which has been retained above. Advertisements, app promotion banners, WhatsApp graphics, and unrelated recommended images have been excluded. Reuters later corrected its report, stating that Meta expects to add 2.5 GW by the end of 2026, not 5.5 GW.

Meta Plans to Start Iris Production in September

According to an internal Meta memo reviewed by Reuters, production of Iris is expected to begin in September 2026.

Testing lasted approximately six weeks and found no major issues. This marks a significant milestone for Meta, whose custom chip program has faced delays and technical challenges since the company began developing its proprietary AI accelerators five years ago.

Meta declined to comment on the reported production plans, so the September timeline should be understood as an internal plan rather than a publicly committed release date.

Iris is designed specifically for Meta's own data center needs. It aims to support AI systems behind products like Facebook and Instagram, including recommendation, ranking, advertising, and generative AI workloads.

The chip is not expected to replace all external accelerators purchased by Meta. Instead, it will become part of a hybrid infrastructure portfolio that includes:

• Meta-designed MTIA accelerators
• NVIDIA GPUs
• AMD Instinct GPUs
• Custom and partner-developed CPUs
• Specialized networking, memory, and storage systems

This combination allows Meta to assign different workloads to hardware that offers the best mix of performance, availability, energy efficiency, and total cost.

What is the Iris AI Chip?

Iris is the codename reported by Reuters for a certain generation of chips within Meta's MTIA project.

MTIA stands for Meta Training and Inference Accelerator. It is a family of AI chips built by Meta for specific purposes, designed to improve the efficiency of workloads running in the company's data centers.

Earlier generations of MTIA have already been deployed by Meta for ranking, recommendation, and ad inference tasks.

The company says these chips offer significant energy efficiency advantages over general-purpose vendor chips when deeply optimized for Meta's own models and software stack.

The Iris project reported on is an extension of this strategy.

Key Reported Facts

Detailed technical specifications such as process node, memory capacity, interconnect bandwidth, power consumption, and peak compute performance were not disclosed in the original report. These details should not be inferred until announced by Meta or its partners.

Four Generations of MTIA Roadmap

The Reuters report states that Iris is part of a four-generation MTIA program designed by Meta.

The company plans to introduce a new chip approximately every six months through 2027. This pace is much faster than the typical one-year or multi-year product cycles common in semiconductor development.

Meta has separately confirmed that it is developing and deploying four entirely new generations of MTIA chips within two years. The roadmap aims to support:

• Ranking and recommendation systems
• Advertising models
• Large-scale inference
• Generative AI workloads
• Future training workloads

A six-month cycle does not necessarily mean a completely new architecture each time. Subsequent versions may feature targeted improvements in compute units, memory systems, network interconnects, packaging, workload support, or software compatibility.

The key point is that Meta wants to turn custom chip development into a continuous internal capability rather than an occasional hardware experiment.

Broadcom Supports Design

Meta is collaborating with Broadcom to accelerate the development of custom AI chips.

Meta's official announcement states that Broadcom will contribute in the following areas:

• Chip design
• Advanced packaging
• Network interconnects
• Multi-generational accelerator development

This collaboration is based on Broadcom's XPU custom AI accelerator platform. Meta retains control over workload requirements and system-level design goals, while Broadcom provides semiconductor and infrastructure expertise.

This model is common among hyperscale tech companies. The cloud or platform company defines the workloads and expected performance metrics, while an established semiconductor partner helps translate these requirements into manufacturable chips.

This arrangement reduces development risk without forcing Meta to rely entirely on standard off-the-shelf accelerators.

TSMC to Manufacture the Chip

The Reuters report says TSMC will manufacture the Iris chip.

TSMC is the world's largest dedicated semiconductor foundry, producing advanced chips for many major tech companies. Partnering with TSMC gives Meta access to leading manufacturing processes, packaging technologies, and large-scale production expertise.

However, using a leading foundry does not eliminate supply chain constraints.

Competition for advanced manufacturing capacity is intense, and AI accelerators also depend on scarce components such as:

• High-bandwidth memory
• Advanced packaging
• Substrates
• Optical networking equipment
• Storage
• Power supply systems

Therefore, Meta needs both custom chip design capabilities and long-term component procurement to achieve its infrastructure goals.

Iris Will Complement NVIDIA and AMD GPUs

Meta remains a major purchaser of external GPUs.

The custom chip program is not intended to immediately replace NVIDIA or AMD. The purpose of Iris is to augment the large GPU clusters used for AI training and inference.

This distinction is important because different workloads tend to favor different types of hardware.

External GPUs Still Important, Mainly For:

• Cutting-edge model training
• Rapid deployment of new model architectures
• Broad software compatibility
• Workloads dependent on established CUDA or ROCm ecosystems
• Tasks that evolve faster than custom chip development cycles

Custom MTIA Chips Can Be Used For:

• Stable, high-volume internal workloads
• Recommendation and ranking systems
• Large-scale repetitive inference
• Workloads that can be co-designed with software
• Reducing per-inference cost
• Improving power and rack efficiency

Meta has also signed a long-term agreement with AMD for up to 6 GW of AMD Instinct GPU infrastructure.

This official partnership confirms that Meta's strategy is diversification, rather than completely abandoning third-party accelerators.

The company wants to maintain access to leading external hardware while gaining more control over its infrastructure stack.

Why Meta Wants Greater Control Over AI Chips

The logic behind custom chips is both technical and financial.

1. Lower Infrastructure Costs

Meta operates AI systems at hyperscale. Even minor efficiency gains can reduce power, cooling, hardware, and operational costs for millions of daily workloads.

Custom chips can remove features Meta doesn’t need and allocate more die area to the operations heavily used by its own models.

2. Better Hardware-Software Codesign

Meta can jointly optimize its models, compilers, kernels, data formats, and accelerator architecture.

This codesign may yield better results than adapting every internal workload to hardware built for a broad external market.

3. Reduced Reliance on a Few Suppliers

NVIDIA remains the primary supplier for advanced AI GPUs, while AMD is expanding its influence. Relying entirely on external suppliers could expose Meta to:

• Supply shortages
• Long lead times
• High prices
• Product roadmap delays
• Integration efforts with new GPU generations

Custom chips provide Meta with an alternative source of capacity and greater control over deployment timing.

4. Faster Deployment for Meta-Specific Workloads

A memo quoted by Reuters noted that adopting the latest GPUs at a company the size of Meta requires immense effort and consumes the company’s time.

A stable internal hardware and software platform can reduce some of this transition burden, especially for workloads Meta runs continuously for years.

Meta's Goal: 7 GW of Compute Capacity by 2026

The scale of the infrastructure plan is as important as the chips themselves.

According to an internal memo, Meta plans to deploy 7 GW of compute infrastructure by 2026. It is reported to have already added 1 GW in the first half of the year and expects to add another 2.5 GW by year-end as part of a broader capacity build-out.

(Gigawatts are a unit of power measurement, not compute performance.)

Using gigawatts to describe infrastructure gives an intuitive sense of the scale of power required to operate data centers, cooling systems, networking equipment, and accelerators.

According to Reuters, 1 GW is roughly enough to power 8 million homes. While this comparison is an approximation and depends on local energy consumption levels, it illustrates the enormous power demands of modern AI infrastructure.

2027 Target: 14 GW

Meta reportedly plans to double its total compute capacity again by 2027, reaching 14 GW.

This target indicates the company’s confidence that demand for AI compute will continue to grow across the following areas:

• Advertising and recommendation systems
• Meta AI assistant
• Generative media
• Software development models
• Wearable and mixed reality products
• Personal superintelligence research
• Internal model training and evaluation

Achieving 14 GW requires far more than just chips. Meta will need ample:

• Data center space
• Grid access
• Power generation capacity
• Cooling systems
• Network capacity
• Memory
• Storage
• Skilled construction and operations teams

The company has already announced initiatives in nuclear energy, renewable energy, long-duration energy storage, and grid support, aimed at strengthening the power supply for future data centers.

Capital Expenditure Could Reach $145 Billion

Meta’s official Q1 2026 earnings report raised its full-year capital expenditure forecast to between $125 billion and $145 billion.

This forecast includes principal payments on finance leases. Meta stated that the spending increase reflects higher component prices and increased data center costs to support future capacity.

This official figure is more precise than simply describing it as "all-in on one chip" or "only for AI accelerators." The capital budget covers the broader infrastructure stack, including:

• Data center construction
• Servers and accelerators
• Networking equipment
• Power systems
• Storage
• Memory
• Leased infrastructure
• Capacity reserves for future years

The upper end of this forecast range highlights why reducing hardware costs and improving utilization have become strategic priorities.

Long-Term Supply Agreements

According to internal memos reviewed by Reuters, Meta has signed multi-year supply agreements with several hardware vendors.

Reported agreements include:

• Samsung Electronics for memory
• SanDisk for flash storage
• Sumitomo Electric for fiber optic equipment

Not all of these companies have publicly confirmed the details of the agreements, so they should be described as "reported agreements" rather than "officially announced."

Long-term procurement is becoming increasingly important as AI infrastructure demand stresses the supply chain for:

• Memory chips
• High-bandwidth memory
• Flash storage
• Networking components
• Fiber optics
• Advanced packaging

Without memory, storage, networking, power, or rack capacity in the surrounding systems, custom accelerators cannot be deployed at scale.

What Iris Means for NVIDIA and AMD

Iris does not mean Meta will stop buying hardware from NVIDIA and AMD.

Instead, it signals that hyperscale buyers want bargaining power and workload flexibility.

The most likely future is a heterogeneous infrastructure stack:

• NVIDIA GPUs for frontier workloads
• AMD GPUs for scaling and diversification
• MTIA chips for specific Meta workloads
• Custom CPUs and networking systems
• Software to schedule workloads across different hardware

For NVIDIA and AMD, this presents both risks and opportunities.

Custom chips may take over some high-throughput inference tasks.

At the same time, Meta’s overall compute demand is growing so rapidly that even as its in-house chip capabilities increase, its procurement of external accelerators may still continue to grow.

The overall market can expand while a single supplier’s share declines.

Unknown Information

The report answered several questions about timelines and infrastructure, but key details remain undisclosed.

Meta has not confirmed:

• The final production month (September)
• The expected volume of Iris chips
• The chip’s manufacturing process
• The memory configuration
• Power consumption levels
• Peak training or inference performance
• The first workloads to run
• The planned deployment date in a production data center
• Expected cost savings compared to NVIDIA or AMD hardware

Until these details are released, claims that Iris outperforms specific external GPUs are speculative.

The most plausible conclusion is more conservative: Meta believes its own chips will improve cost efficiency, control, and deployment efficiency for specific internal AI workloads.

Frequently Asked Questions

What is Meta's Iris AI chip?

Iris is the reported codename for a custom data center AI accelerator under Meta's MTIA program. The chip is designed for Meta's internal AI workloads and is expected to complement (not immediately replace) GPUs from NVIDIA and AMD.

When will the Iris chip enter mass production?

Internal memos reviewed by Reuters show Meta plans to begin mass production of the chip in September 2026. Meta declined to comment, so this date is an internal target, not a confirmed public release.

Who designs and manufactures Iris?

Meta designs the accelerator for its workloads, with Broadcom providing design, packaging, and networking support. Reuters reports that TSMC will manufacture the chip.

What does MTIA stand for?

MTIA stands for Meta Training and Inference Accelerator. It is Meta’s family of custom AI chips used for recommendations, ranking, advertising, inference, and gradually expanding to generative AI workloads.

Will Meta stop buying GPUs from NVIDIA and AMD?

No. Iris is designed to work alongside external GPUs to add custom compute capacity. Meta has announced long-term agreements to deploy up to 6 GW of AMD Instinct GPU infrastructure and continues to use NVIDIA hardware extensively.

How much compute capacity does Meta plan to deploy?

Reuters reports that Meta aims for 7 GW of compute infrastructure in 2026 and 14 GW in 2027. These numbers refer to the power capacity of the overall compute infrastructure, not individual chip performance.

How much will Meta spend on AI infrastructure in 2026?

Meta’s official 2026 capital expenditure guidance is $125 billion to $145 billion, inclusive of principal payments on finance leases. This budget covers data centers, components, servers, networking, and other infrastructure, not just the Iris chip.

Why do tech companies build their own AI chips?

Building custom chips can reduce costs, improve energy efficiency, better align with internal workloads, and reduce reliance on a small number of external suppliers. These chips typically work alongside general-purpose GPUs rather than fully replacing them.

Related Tools

Meta Training and Inference Accelerator: Meta's official introduction to its next-generation MTIA accelerator.

• Broadcom Custom AI Accelerator: Broadcom's custom XPU technology for large-scale AI infrastructure.
• TSMC: The semiconductor foundry reportedly manufacturing Meta's Iris chip.
• NVIDIA Data Center: NVIDIA's portfolio of GPUs and systems for AI training and inference.
• AMD Instinct: AMD's data center accelerator platform for large-scale AI deployment.
• Open Compute Project: An open hardware community where hyperscalers share data center, rack, network, and infrastructure design solutions.

Related Links

• Original Report from The Economic Times: A Reuters report republished by The Economic Times.
• Original Report from Reuters: Reuters' story based on an internal Meta memo.
• Meta and Broadcom's Custom Silicon Partnership: Meta's official announcement covering four generations of MTIA architecture and Broadcom's role.
• Meta Q1 2026 Earnings Report: Official financial results including capital expenditure forecasts of $125 billion to $145 billion.
• Meta and AMD Infrastructure Agreement: Meta's official agreement to build up to 6 GW of AMD Instinct GPU infrastructure.
• Meta Infrastructure Evolution: Meta's official overview of data center architecture and MTIA deployment.
• Meta and Arm Data Center Chip Collaboration: Meta's official announcement on designing custom CPUs for AI data centers.
• Meta Nuclear Energy Projects: Details on Meta's energy agreements to support future AI infrastructure.

Summary

According to reports, Meta plans to begin manufacturing its Iris AI accelerator in September 2026, following a six-week testing period. The chip is part of a four-generation MTIA roadmap supported by Broadcom and manufactured by TSMC.

Iris will not replace Meta's existing NVIDIA and AMD GPU clusters but will instead add a customized, workload-optimized layer aimed at reducing costs, improving efficiency, and giving Meta more control over its AI infrastructure.

The chip arrives as Meta pursues even larger expansion: targeting 7 GW of computing power by 2026 and 14 GW by 2027, with capital expenditures of up to $145 billion this year.

The significance of Iris as a single chip pales in comparison to what it represents: evidence of Meta's intent to take greater control over the hardware, power, and supply chain behind AI.