Meta is expected to mass-produce its self-developed AI chip Iris in September! Joining forces with TSMC to expand its "14 gigawatt" computing power empire.

The custom chip war is heating up! According to an internal memo exclusively disclosed by Reuters, tech giant Meta is expected to start mass production of its in-house AI chip codenamed "Iris" in September 2026, with TSMC handling manufacturing. To build a global AI empire, Meta plans to invest $145 billion in infrastructure, aiming to double its total computing power capacity to a staggering 14 GW next year, thereby significantly reducing reliance on external suppliers like NVIDIA.

(Previous: Meta launches AI image generation tool Muse Image, but uses your Instagram photos without your consent) (Background: Meta wakes up? AI head: New model "Watermelon" scores can match GPT-5.5)

Table of Contents

Toggle

• Teaming up with Broadcom and TSMC, Iris chip to start production in September
• Spending $145 billion, targeting doubled computing power to 14 GW
• Locking in three major suppliers to fully counter "Chipflation"

Amid the exponential explosion in global AI computing power demand, tech giants are racing to seize control of the hardware layer, and Meta, led by Mark Zuckerberg, is undoubtedly one of the most aggressive players.

According to an exclusive internal memo obtained by Reuters on July 9, 2026, Meta is preparing for an unprecedented large-scale upgrade of its AI infrastructure, with the top priority being the full launch of its self-developed AI chip mass production plan this fall.

Teaming up with Broadcom and TSMC, Iris chip to start production in September

The report states that the data center AI chip codenamed "Iris" is expected to officially enter production in September 2026. As a key piece of Meta's fourth-generation custom chip plan, "Meta Training and Inference Accelerators (MTIA)", Iris will be specifically used to enhance AI inference and generation performance on Facebook and Instagram.

Notably, Meta's hardware development capabilities have significantly improved. According to the memo, the testing phase for the Iris chip was completed in just 6 weeks without any major defects found — a milestone success for Meta's chip team, which has faced setbacks in recent years. In terms of supply chain分工, Meta has chosen to co-design the chip with networking chip giant Broadcom, and have it manufactured by foundry leader TSMC, aiming to significantly reduce reliance on and huge procurement costs from external GPU suppliers like NVIDIA and AMD.

Spending $145 billion, targeting doubled computing power to 14 GW

Beyond the progress on custom chips, the memo also reveals Meta's staggering computing power expansion blueprint. The company plans to deploy up to 7 GW of computing infrastructure this year, with a target of doubling to 14 GW by 2027.

To support this massive computing beast, Meta's budget for AI infrastructure this year is estimated at $145 billion. Amid the global Big Tech AI capex arms race exceeding $700 billion, Meta's investment scale is among the top in the industry. The company has even set a goal to release a new AI processor every 6 months before 2027, a iteration pace far exceeding the current industry standard of one per year.

Locking in three major suppliers to fully counter "Chipflation"

While expanding rapidly, Meta has also keenly recognized the risks of raw material shortages. To cope with "Chipflation" caused by surging demand in the market, Meta has already made early moves, signing multi-year long-term supply agreements with several top component manufacturers:

| Supply Chain Partners | | --- | | Supply Core Components | | Strategic Significance & Market Impact | | --- | --- | --- | | Samsung Electronics | | Advanced memory chips | | Ensuring stable supply of High Bandwidth Memory (HBM), breaking AI computing bottlenecks. | | SanDisk | | Flash storage | | Meeting extreme demands for high-speed storage during massive AI model training and inference. | | Sumitomo Electric | | Fiber optic equipment components | | Enhancing transmission speed and bandwidth stability within ultra-large data centers. |

Although Meta officially declined to comment on the leaked memo, it is foreseeable that with the Iris chip going online in September, it will form a powerful complementary matrix with the large quantities of NVIDIA and AMD GPUs Meta purchases. The "hardware chip-building movement" driven by tech giants to reduce costs and increase underlying autonomy will have a profound impact on the global semiconductor and AI industry landscape in the coming years.