Meta’s in-house developed chip Iris will begin production in September, manufactured by TSMC, as the company looks to move away from reliance on NVIDIA.

An internal memo obtained by Reuters reveals that Meta's self-developed AI chip, codenamed "Iris," is expected to enter production in September 2026. Co-designed with Broadcom and manufactured by TSMC, it is the fourth-generation chip in the MTIA series, primarily targeting AI inference workloads for Facebook and Instagram.

(Background: Meta spends billions of dollars to bind with Amazon AWS! Snaps up hundreds of thousands of Graviton5 chips to counter NVIDIA's AI computing monopoly) (Background supplement: OpenAI joins Broadcom to launch its first AI chip "Jalapeño," achieving the milestone in nine months to challenge Nvidia's dominance)

Table of Contents

Toggle

• Why not keep buying from Nvidia?
• What exactly is Iris?
• Computing arms race: A bet from 7 GW to 14 GW

An internal memo obtained by Reuters has leaked that the social media giant plans to move the heart driving its AI empire from Nvidia into its own hands. The memo specifies a clear timeline: in September 2026, the self-developed chip codenamed "Iris" will officially enter production. On one side are the general-purpose GPUs that Nvidia and AMD have sold for a decade; on the other are custom chips tailored for Facebook and Instagram. Meta is choosing to pursue both paths in parallel.

Why not keep buying from Nvidia?

The answer lies in Meta's own admission: deploying the latest generation of GPUs on its massive infrastructure scale is "extremely difficult." Nvidia's chips are general-purpose products designed for customers worldwide. Meta wants custom chips that fit its data centers precisely, saving not just procurement costs but also hidden costs of scheduling and cooling.

When computing power reaches a certain scale, the efficiency loss of general-purpose chips becomes magnified, making customization the more cost-effective choice.

It is understood that this self-development journey has actually been ongoing for five years, not without its obstacles along the way. The MTIA (Meta Training and Inference Accelerator) project, Meta's own series of training and inference accelerator chips, has encountered multiple delays in the past, falling behind external expectations.

In March of this year, Meta publicly showcased Iris along with three other AI processors for the first time, marking a preliminary milestone after five years of work. External procurement and self-development have never been an either/or choice. Meta states that the self-developed chips are meant to "supplement," not replace, its purchases from Nvidia and AMD. After all, the daily recommendations, translations, and content moderation handled by social platforms are so massive that no single chip can handle all the work alone.

What exactly is Iris?

Iris is the fourth-generation chip in the MTIA series, co-designed by Meta and chip design giant Broadcom, and expected to be manufactured by TSMC.

Its primary task is focused on "inference." Simply put, it handles the daily operations where already-trained AI models actually answer questions, generate recommendations, and judge content, rather than training cutting-edge large models from scratch.

In other words, training is like burning money to build a brain, while inference is having that brain show up to work every day. The former is a one-time astronomical investment, while the latter is a recurring bill that happens every minute and only grows larger. The "You might like" recommendations, automatic translations, and content violation detection on Facebook and Instagram's feeds are all supported by this type of inference computing.

The memo shows that Iris successfully passed its approximately six-week testing phase without major defects. For a project that had previously stalled, this represents a rare smooth stretch, and it has given Meta the confidence to put the mass production timeline directly into the internal memo.

Computing arms race: A bet from 7 GW to 14 GW

On another front, Meta plans to deploy about 7 GW of computing infrastructure this year, and double it to 14 GW by 2027. Its spending on AI infrastructure this year is estimated at around $145 billion. More aggressive is the pace: Meta plans to launch a new AI processor every six months by 2027, far faster than the industry norm of one per year.

Self-developed chips are by no means unique to Meta. OpenAI has already partnered with Broadcom to launch its own chip; Google has its TPU; Anthropic is reportedly in talks with Samsung for custom chips; and even DeepSeek is developing its own inference chip.

Nvidia's moat has never been about being unable to build better chips, but that everyone has to buy from it. However, when Meta, OpenAI, and Google all start replacing the option of "buying chips" with "building their own," the rules of the game are being quietly rewritten. No one has placed all their bets yet.