Meta Bets on Its Own Silicon With September Launch of Iris Chip

When Apple designed its first in-house A-series processor more than fifteen years ago, the logic was that a company betting its future on a product could no longer afford to rent the most important component of it. Meta Platforms has now reached a comparable threshold in the data center. According to CNBC, an internal Meta memo reviewed by Reuters and reported in the second week of July shows the company will begin manufacturing its custom artificial-intelligence chip, code-named Iris, in September, a decision that marks the moment its long-running silicon ambitions cross from prototype into production.

Iris Moves From Laboratory to Fabrication Line

Iris is the fourth-generation processor in Meta's Meta Training and Inference Accelerator program, the internal effort the company has pursued for several years to build accelerators tuned specifically for the recommendation systems and generative models that power Facebook and Instagram. According to CNBC, the chip was designed in partnership with Broadcom and will be fabricated by Taiwan Semiconductor Manufacturing Company, the same foundry that produces the most advanced processors for Apple, Nvidia and Advanced Micro Devices.

The memo cited by CNBC describes an unusually smooth path through validation. Chip testing took only about six weeks and surfaced no major issues, an outcome that engineers rarely enjoy on a first pass through silicon of this complexity. That result matters beyond a single product. It suggests Meta's design and verification pipeline has matured to the point where the company believes it can sustain a rapid cadence, aiming to launch a new chip roughly every six months through 2027, according to CNBC. By comparison, most semiconductor firms measure their release intervals in years rather than months.

Loosening a Costly Dependence on Nvidia

Behind the engineering milestone sits a commercial calculation that has preoccupied every large operator of AI infrastructure. Training and serving frontier models has, until now, meant purchasing enormous quantities of Nvidia's graphics processors at premium prices and on Nvidia's timetable. For a company spending at the scale Meta is, that dependence is both a budgetary line item and a strategic vulnerability.

Custom silicon offers a route around it. A chip designed for Meta's own workloads can be optimized for the specific inference and ranking tasks that dominate its traffic, potentially delivering more useful computation per watt and per dollar than a general-purpose accelerator bought on the open market. The Iris program does not eliminate Nvidia from Meta's supply chain, and the company has given no indication that it intends to abandon merchant silicon outright. What production of Iris does accomplish is optionality, the ability to shift a growing share of internal demand onto hardware Meta controls end to end.

Broadcom and TSMC as the Enabling Partners

The structure of the effort is worth noting because it reflects how hyperscale companies now approach chip design without becoming chip manufacturers themselves. Meta supplies the architectural intent and the workload knowledge; Broadcom contributes the design engineering and intellectual property that turn that intent into a manufacturable layout; and TSMC provides the fabrication capacity. According to CNBC and Reuters, this division of labor is what allows a software company to field competitive silicon on a compressed schedule.

Seven Gigawatts This Year, Fourteen by 2027

The chip roadmap is inseparable from an infrastructure buildout of extraordinary scale. According to CNBC, Meta is targeting seven gigawatts of compute capacity in 2026 and intends to double that figure to fourteen gigawatts in 2027. Gigawatts have become the working unit of AI ambition, a proxy for the raw electrical draw of the server halls a company can bring online, and doubling that capacity within a single year implies a construction and procurement program measured in tens of billions of dollars.

The spending numbers bear that out. According to CNBC, Meta's AI infrastructure outlay could run as high as one hundred forty-five billion dollars this year. Placing Iris into production against that backdrop reframes the chip as more than a research trophy. It becomes a lever on the single largest cost center in the company's expansion, and the degree to which internally designed silicon can absorb the coming demand will shape how efficiently those dollars translate into usable capacity.

• Iris enters manufacturing in September 2026 as the fourth MTIA-generation chip, according to CNBC.
• Meta targets seven gigawatts of compute in 2026, doubling to fourteen gigawatts in 2027, per CNBC.
• AI infrastructure spending could reach one hundred forty-five billion dollars this year, according to CNBC.
• The company aims to ship a new chip roughly every six months through 2027, per CNBC.

Investors Reward the Vertical Integration Story

Financial markets read the disclosure as a validation of Meta's strategy rather than a warning about its cost. According to CNBC, optimism over the AI chip plan helped push Meta's stock to its best week since early 2024, with shares surging around six percent. That reaction reflects a shift in how investors weigh capital-intensive AI programs. Where heavy infrastructure spending once drew skepticism about returns, evidence that a company can control more of its own hardware stack now reads as a durable competitive advantage.

According to CNBC, testing of the Iris chip took only about six weeks with no major issues, and Meta aims to launch a new chip roughly every six months through 2027.

The enthusiasm is not without qualification. A first production run is not the same as deployment at scale, and the true measure of Iris will be how it performs across Meta's fleet once it displaces merchant accelerators in live workloads. Yield rates at TSMC, the software maturity needed to run existing models on new hardware, and the pace at which Meta can actually stand up gigawatts of capacity all remain open questions that a single memo cannot answer.

Positioning Within the Hyperscaler Silicon Race

Meta is not alone in this pursuit, and the competitive context sharpens the significance of the September timeline. The largest cloud and platform companies have each built or commissioned custom accelerators to reduce their exposure to a single supplier and to tailor hardware to their own economics. Meta's decision to compress its release cadence to roughly twice a year is an attempt to keep pace with a field moving faster than traditional semiconductor cycles allow.

Whether that cadence proves sustainable is the question the industry will now watch. Designing a chip every six months demands not only engineering talent but foundry allocation, packaging capacity and a validation pipeline that can keep errors from reaching production, the very pipeline the six-week testing result is meant to demonstrate. If Meta holds to the schedule described in the memo, Iris will be the first of a series rather than a standalone bet, and the September start date becomes a reference point against which each subsequent generation is measured.

For now, the fact reported by CNBC stands on its own terms. A company that built its business on software is committing to manufacture its own advanced silicon, and it is doing so at a moment when the cost of not controlling that hardware has become too large to ignore. The details in this draft rest on the CNBC and Reuters reporting cited throughout and remain subject to human verification before publication.