Meta to Put Its Iris AI Chip Into Production in September as It Targets 14 Gigawatts of Compute

Meta plans to begin mass production of its in-house AI chip, codenamed Iris, in September 2026, as the company pushes to reduce its reliance on Nvidia and take greater control of the infrastructure powering its artificial intelligence ambitions. The plans, first reported by Reuters from an internal company memo, mark one of the most concrete steps yet by a hyperscaler to put a custom accelerator into the field at scale.

The disclosure also lands at a moment of intensifying competition in custom AI hardware, where every major cloud and consumer technology company is racing to design silicon tailored to the demands of large language model training and inference. For Meta, which spends tens of billions of dollars annually on computing, even modest gains in efficiency or cost per token translate into enormous savings.

Doubling Compute Capacity to 14 Gigawatts

According to the memo described by Reuters and corroborated by outlets including Quartz, Firstpost, and Moomoo, Meta intends to roughly double its computing capacity to about 14 gigawatts by 2027. Gigawatts are the standard yardstick for the colossal data center buildouts now underway across the industry, and a 14-gigawatt target places Meta among the most aggressive infrastructure spenders in the world.

The Iris accelerator is central to that expansion. While Reuters did not disclose detailed specifications, the chip is understood to be designed for the inference and training workloads that dominate Meta's recommendation systems, generative AI features, and internal research. Moving it into production means Meta is preparing to deploy its own silicon alongside, and potentially in place of, the Nvidia GPUs that currently anchor its data centers.

Several technology publications, including Times Now and NewsBytes, framed the move as evidence that Meta's AI push is gathering pace and maturing beyond pure procurement of third-party chips.

Why a Custom Chip Matters

Meta is not the first company to build its own AI silicon. Google has shipped its Tensor Processing Units (TPUs) for roughly a decade, Amazon has developed its Trainium and Inferentia accelerators, and Microsoft has unveiled its own custom AI chips. But Meta's scale of deployment and the breadth of its AI workloads make the Iris program especially consequential for the competitive balance in the accelerator market.

Custom chips offer two principal advantages. First, they can be optimized for a company's specific software stack and model architectures, squeezing out performance that general-purpose GPUs leave on the table. Second, they reduce dependency on a single supplier at a time when Nvidia's high-end GPUs are expensive and frequently supply-constrained. For a company planning to double its compute footprint within roughly two years, that strategic leverage is difficult to overstate.

Spending Plans Overshadow the Chip Milestone

Despite the positive signal on custom silicon, investors reacted cautiously. Yahoo Finance reported that Meta shares fell as the company's massive AI infrastructure spending plans overshadowed the chip progress, a pattern echoed by TradingPedia, which noted that Meta stock slipped as the spending commitments dwarfed the manufacturing milestone.

The tension reflects a broader industry question that has weighed on the entire hyperscaler cohort: whether the capital pouring into AI computing will generate returns quickly enough to satisfy markets. Meta's stated ambition to reach 14 gigawatts of capacity implies continued, heavy capital expenditure on land, power, cooling, and chips for years to come.

The Energy and Power Constraint

Reaching 14 gigawatts of computing capacity is as much an energy challenge as a silicon one. A single gigawatt is roughly the output of a large nuclear reactor, meaning Meta's target implies securing power on the scale of multiple new power plants within a few years. That has pushed hyperscalers, Meta included, into unprecedented negotiations over nuclear energy, renewable contracts, and dedicated grid connections.

The Iris chip program is partly a response to this constraint. More efficient silicon draws less power for the same computational work, easing the burden on an already strained electricity supply. If Iris can improve performance per watt, it could help Meta stretch its gigawatts further and reduce both its energy bills and its carbon footprint.

A Pivotal Moment for Hyperscaler Silicon

The September production target gives Meta a defined timeline to watch. If Iris delivers meaningful performance or cost advantages in real-world deployments, it could accelerate a shift in which the largest AI operators increasingly build rather than buy their most critical accelerators. That would reshape the economics of the entire AI hardware supply chain, from foundries to the graphics processor vendors that have dominated the last two years.

For now, the memo signals intent rather than proven results. But the combination of a production date, a doubling of capacity, and a clear strategic rationale makes Iris one of the most closely watched custom chip programs in the industry.

Stay Ahead of AI

For deep, independent coverage of the AI hardware race, data center buildouts, and the companies reshaping compute infrastructure, follow AI Buzz Wire for the developments defining the next era of artificial intelligence.

Read more AI hardware news →