Ming-Chi Kuo: Google is developing an upgraded TPU v9 chip, MediaTek secures the exclusive order, mass production in 2028.

```

Google and MediaTek's partnership in the AI chip sector is advancing to a deeper level.

On June 22, renowned Apple supply chain analyst Ming-Chi Kuo revealed on social platform X that Google is developing an upgraded chip based on the TPU v9 architecture, reportedly codenamed "Triggerfish", with MediaTek exclusively securing this new order. The target applications are AI agents and reinforcement learning scenarios, with mass production expected to scale in 2028.

According to Kuo's latest industry-chain investigation, Triggerfish is an upgraded derivative project based on the existing TPU v9 "Humufish". Its core specifications are comprehensively strengthened: SRAM capacity is increased to 2-3 times that of Humufish, memory is upgraded from HBM4 to HBM4E, and a simulation die has been added. These changes aim to enhance inference capability and alleviate CPU and memory bottlenecks.

Commercially, Triggerfish’s unit price is about 30% higher than Humufish. Google plans to maintain cumulative shipments of 4–5 million Humufish chips, and additionally order 1–2 million Triggerfish chips. Production is expected to start at the end of 2027, with mass production ramping up in 2028. Kuo points out that this order is expected to become a growth driver for MediaTek's 2028 performance.

MediaTek deeply tied to Google's TPU v9 generation

Kuo states that the Triggerfish project further affirms MediaTek’s status as Google’s preferred partner for TPU v9 development. MediaTek exclusively undertakes this upgraded order, continuing the relationship built in the Humufish project, demonstrating that both parties have established strong barriers in custom AI chip cooperation.

Triggerfish is defined as a variant of v9 rather than a completely new generation, representing targeted enhancements on the existing platform. This "in-platform iteration" strategy enables Google to quickly respond to evolving AI workload demands without starting from scratch, while giving MediaTek ongoing opportunities for incremental orders.

Technically, the core differences between Triggerfish and Humufish focus on three dimensions: massively expanded SRAM (2–3 times), an added simulation die, and memory upgraded to HBM4E.

Kuo notes that the main function of the added simulation die likely centers on reinforcement learning (RL) and coordination of AI agents. Larger SRAM allows the active working set required by reinforcement learning and AI agents to be retained locally on the TPU, reducing data movement costs and improving ultra-low latency decoding efficiency.

In addition, Triggerfish has also been strengthened in local TPU management and training/inference mode switching, with the overall design aiming to maximize effective compute to meet increasingly complex AI inference task demands.

For shipment forecasts, Kuo maintains his estimate of Humufish's lifecycle shipments at 4–5 million chips, while Triggerfish, as an incremental project, is expected to add another 1–2 million units.

In terms of pricing, Triggerfish’s unit price is about 30% higher than Humufish, meaning that even with comparatively limited shipments, its contribution to MediaTek's revenue will surpass Humufish for the same number of units.

Considering the timeline of production starting at the end of 2027 and mass ramp-up in 2028, Kuo believes Triggerfish is likely to become an important incremental source of business momentum for MediaTek in 2028.

Risk Disclosure and DisclaimerThe market carries risks and investments should be made cautiously. This article does not constitute personal investment advice and also does not take into account the specific investment objectives, financial circumstances, or needs of individual users. Users should consider whether any opinions, viewpoints, or conclusions in this article are suitable for their particular situation. Any investment made based on this is at the investor's own responsibility. ```