What is the Meta Iris chip?

It is an in-house AI accelerator and one of four generations in Meta’s MTIA chip roadmap. Designed in collaboration with Broadcom and manufactured by TSMC, it is scheduled for mass production in September 2026 to handle AI workloads within Meta’s own data centers.

Will Iris replace NVIDIA GPUs?

No, at least not for now. The memo explicitly positions Iris as a supplement to NVIDIA and AMD GPUs. Meta intends to use its own chips for specific, optimized workloads to reduce costs, while continuing to rely heavily on GPUs for cutting-edge AI training.

Which parts of the Taiwanese supply chain will benefit?

Key beneficiaries include wafer manufacturing (TSMC’s advanced nodes), advanced packaging (CoWoS), and AI server assembly (Foxconn, Quanta, Wistron, etc.). A common trend among hyperscalers is that while chip designs vary, the manufacturing almost exclusively takes place in Taiwan.

What is the significance of the 7GW and 14GW compute targets?

GW (gigawatts) represents the power consumption scale of data centers and has become the standard metric for measuring AI infrastructure investment. Meta plans to bring 7GW online by 2026 and reach 14GW by 2027—a power demand equivalent to several large-scale nuclear power plants, all dedicated to AI infrastructure.

Meta’s In-House AI Chip "Iris" Set for September Production: Broadcom Design, TSMC Manufacturing, and a Two-Year Compute Doubling Goal

A Reuters exclusive reveals an internal Meta memo: the proprietary AI accelerator "Iris" will enter production in September 2026. Designed by Broadcom and manufactured by TSMC, the chip is set to power Meta’s ambitious plan to double its compute capacity from 7GW to 14GW, marking another major win for the Taiwanese supply chain.

In NVIDIA’s earnings calls, the question analysts love to ask most is about the "threat of custom silicon." The other half of the answer, however, is hidden in Hsinchu. According to an internal Meta memo obtained by Reuters last week, the tech giant—which serves billions of users worldwide—will move its custom AI accelerator, "Iris," into production this September. The design partner? Broadcom. The manufacturer? TSMC, as always.

Background

Meta’s custom chip initiative, known as MTIA (Meta Training and Inference Accelerators), has been in development for several generations, with Iris representing one of the four planned iterations. The details revealed in the memo are remarkably rare: Iris completed its debugging and testing phase in just six weeks with "no major issues"—a pace that can only be described as a smooth delivery for a data-center-grade ASIC. The motivation is equally straightforward: to drive down compute costs and reduce reliance on NVIDIA and AMD.

Key Takeaways

Production Timeline: Iris will enter mass production in September 2026, with Broadcom serving as the design partner and TSMC handling manufacturing.
Scaling Compute: Meta plans to bring 7GW (gigawatts) of compute capacity online in 2026, with a goal to double that to 14GW by 2027.
Strategic Positioning: Iris is intended to "complement," not "replace," NVIDIA and AMD GPUs.
Efficiency: The six-week debugging cycle marks an unusually smooth development progress.
Source: The information stems from an internal memo obtained by Reuters; Meta has not issued an official press release.

Market Impact Analysis

For Taiwanese Users: While you won't be buying Iris chips directly, the services you use daily will run AI more affordably. As the cost structure for AI features across the Meta ecosystem (IG, FB, WhatsApp) improves, the ceiling for free AI functionality will continue to rise.

For Taiwanese Enterprises: This is the most "Taiwan-centric" international news of the week. Every path toward "de-NVIDIA-fication" by hyperscalers leads back to Taiwan: Broadcom’s ASICs require TSMC’s advanced process nodes, advanced packaging consumes CoWoS capacity, and the final server assembly relies on AI server ODMs like Foxconn, Quanta, and Wistron. Meta’s move from 7GW to 14GW translates into guaranteed orders for the Taiwanese supply chain for the next two years—orders that don't depend on NVIDIA’s approval.

For Developers: In an era of fragmented custom silicon, abstraction at the framework level is becoming increasingly critical. Engineers who master compilers and cross-hardware toolchains (like Triton or XLA) will see their bargaining power rise, while those tied exclusively to the CUDA ecosystem are accumulating risk.

Future Trends

Google has its TPUs, Amazon has Trainium, and OpenAI’s collaboration with Broadcom has long been public knowledge. Meta’s Iris simply completes the puzzle of "every hyperscaler having their own chip." For Taiwanese investors and industry professionals, three indicators are worth watching: the scheduling of TSMC’s advanced process and CoWoS capacity, Broadcom’s ASIC business guidance, and shifts in the order structures of AI server ODMs. This custom chip race, much like the Taiwanese chip export controls we analyzed previously, points to one reality: "compute sovereignty" has become a shared anxiety among major powers and tech giants alike.

TheAI Academy Summary & Commentary

Bottom Line: Every step major tech companies take toward de-NVIDIA-fication adds more orders to the Taiwanese supply chain. Iris is not the death knell for NVIDIA; it is another long-term contract for TSMC.

Advice for Taiwanese readers: Industry professionals should monitor job trends in ASIC design services and advanced packaging. Investors should remember that this article is not investment advice; please refer to individual company financial reports and investor conferences for the extent of supply chain benefits.

Sources

The above is compiled from public information and is subject to subsequent official statements. Please evaluate investment decisions independently; this article does not constitute investment advice.