The social-media giant plans to begin producing its “Iris” processor in September as it doubles computing capacity and seeks greater control over the costly infrastructure powering artificial intelligence
Meta is preparing to place a new internally designed artificial-intelligence chip into production in September, marking a significant escalation in the technology industry’s effort to reduce its dependence on Nvidia and other dominant semiconductor suppliers.
The data-centre processor, code-named Iris, forms part of Meta’s expanding family of Training and Inference Accelerators, known as MTIA. The company intends to use the custom silicon to power artificial-intelligence systems across Facebook, Instagram and its wider portfolio of products.
According to an internal company memorandum reviewed by Reuters, Meta plans to expand its total computing infrastructure to seven gigawatts during 2026 and double that figure to 14 gigawatts in 2027. The scale of the programme illustrates how access to processors, electricity and data centres has become as strategically important to major technology companies as the development of the AI models themselves.
Iris was reportedly tested for six weeks without encountering major technical problems, an encouraging result for an internal semiconductor programme that has previously struggled with delays. Meta plans to introduce a new chip approximately every six months through 2027, a substantially faster cycle than the annual or longer schedules commonly followed by chipmakers.
The processor is being tailored specifically to Meta’s workloads rather than designed as a general-purpose competitor to every chip sold by Nvidia or AMD. Broadcom is assisting with its development, while Taiwan Semiconductor Manufacturing Company is expected to handle production.
Meta will continue buying large quantities of graphics processors from Nvidia and AMD. Its custom chips are intended to supplement those systems, particularly for inference—the process through which a trained AI model responds to users, recommends content or generates images, video and text.
That distinction is commercially important. Training sophisticated models requires enormous computing clusters, but the continuing cost of operating those models for billions of users can eventually become even larger. A processor optimised for Meta’s specific applications could reduce energy consumption, improve efficiency and lower the cost of every AI-generated response.
The initiative also reflects a growing concern among the world’s largest technology groups: companies seeking leadership in artificial intelligence remain heavily dependent on a small number of semiconductor manufacturers.
Nvidia’s graphics processors have become central to the AI economy because of their performance and supporting software. That dominance has produced remarkable growth for the chipmaker, but it has also created supply constraints, high prices and strategic vulnerability for its largest customers.
Amazon, Google and Microsoft have already developed specialised processors for their own cloud and AI platforms. Meta’s accelerated programme suggests that custom silicon is moving from an experimental project to an essential element of competition among global technology companies.
An internal Meta assessment acknowledged that adopting each new generation of commercial graphics processors across an organisation of its size had proved costly and time-consuming. By controlling more of its chip design, the company could adapt processors directly to its software and deploy them on a schedule determined by its own priorities.
The effort is part of an extraordinary expansion of AI infrastructure. Data centres are consuming increasing quantities of electricity, water, memory chips, networking equipment and fibre-optic cables as companies race to build more powerful models and integrate them into everyday services.
The industry is therefore shifting its attention from simply creating larger models to making artificial intelligence less expensive to operate. At an AI summit held in Paris this week, chip designers, cloud providers and start-ups placed particular emphasis on reducing inference costs and extracting more performance from each unit of electricity and computing power.
Meta’s target of 14 gigawatts demonstrates the magnitude of the challenge. One gigawatt can represent electricity demand comparable to that of hundreds of thousands of homes, meaning the company’s expansion will require not only processors but also new power agreements, transmission infrastructure and enormous data-centre construction projects.
The investment also carries considerable financial risk. Custom processors can reduce long-term operating expenses, but developing them requires billions of dollars, highly specialised engineering teams and access to limited manufacturing capacity. A design that arrives late or performs below expectations can quickly become obsolete in a rapidly changing market.
Nevertheless, control over computing infrastructure is increasingly viewed as a source of competitive power. Companies that own their models but rely entirely on rivals for the chips needed to operate them may struggle to control prices, product schedules and access to new technology.
Meta’s production timetable therefore represents more than another semiconductor launch. It signals a broader transformation in which the world’s largest AI companies are attempting to control the entire technological chain—from models and software to chips, data centres and electricity.
Nvidia is unlikely to lose its central position soon. Meta’s continued purchases from the company underline how difficult it is to replace the performance and software ecosystem surrounding its processors.
But Iris demonstrates that Nvidia’s largest customers no longer want dependence to be permanent. As demand for artificial intelligence expands, the next phase of the industry’s competition may be determined not only by which company develops the smartest model, but by which one can operate it most efficiently and independently.
