Google Upgrades TPU V9 to Bet on the AI Agent Era. MediaTek Secures Exclusive Orders for the First Time, Will It End Nvidia's Chip Dominance?

TradingKey - On Monday, June 22, well-known Apple supply chain analyst Ming-Chi Kuo posted that Google (GOOGL) (GOOG) will develop an upgraded v9 chip, potentially codenamed "Triggerfish," based on the TPU v9 (codenamed "Humufish"), with MediaTek securing the exclusive order. This chip targets AI Agent and reinforcement learning (RL) scenarios, and is expected to enter the volume production phase in 2028.

While this is not their first collaboration, it marks the first time MediaTek has exclusively secured an order for Google's core TPU. Previously, in the TPU v7/v7e generations, MediaTek had already entered Google's ecosystem, primarily handling some I/O (input/output) solutions. This latest partnership will further deepen the relationship between the two companies.

Kuo pointed out that compared to the v9-generation Humufish chip, this revised TPU v9 chip has several major differences. The revised v9 chip significantly increases SRAM (static random-access memory) capacity to 2 to 3 times that of the v9, and adds a new simulation die. It also upgrades the High Bandwidth Memory (HBM): while the v9 uses HBM4, the revised v9 chip utilizes the more advanced HBM4E.

This series of upgrades is aimed at enhancing the chip's inference capabilities while alleviating the CPU wall (processor bottleneck) and memory wall (memory bottleneck) issues. The CPU wall refers to the speed of the central processing unit (CPU) in handling task scheduling failing to keep pace with the computational speed of AI accelerator chips (GPUs and TPUs), which similarly drags down overall computing speed.

The reason for needing a larger SRAM capacity is that SRAM is the fastest and lowest-latency cache inside the chip. Expanding its capacity allows more of the active working set required for reinforcement learning and AI Agents to be kept locally on the TPU, thereby reducing data movement costs and improving operational efficiency during the ultra-low-latency decoding phase, pushing real-time decision-making and response speeds to the limit.

The primary function of the newly added simulation die is highly likely focused on reinforcement learning and AI Agent coordination, as reinforcement learning often relies on virtual environment simulations.

Upgrading HBM4 to the higher-speed HBM4E is intended to boost memory bandwidth, fundamentally addressing the memory wall issue.

Regarding shipments, Kuo maintained his forecast of 4 to 5 million units for Humufish's full lifecycle, while Triggerfish is viewed as an incremental project expected to contribute an additional 1 to 2 million units in shipments. In terms of pricing, as Triggerfish's unit price is approximately 30% higher than Humufish's, this partnership will yield a higher revenue contribution for MediaTek even with relatively limited shipment volumes.

Can Google's TPU v9 break Nvidia's monopoly?

The main competitor for Google's self-developed TPU is currently Nvidia's GPU. Compared to the latter, Google's current upgrade of its v9 chip architecture is expected to enable it to compete with Nvidia (NVDA) through differentiated competition, or even achieve a relative lead. Nvidia's current GPUs (such as Blackwell) are still designed around the logic of general-purpose computing power and have shown no additional advantage in handling continuous decision-making for AI Agents and reinforcement learning environment simulations, still facing CPU and memory wall issues. Through this upgrade, Google may establish a generational lead in dedicated AI Agent chips.

In terms of ecosystem, Google's TPU is bundled with its own Google Cloud, allowing it to offer the most cost-effective, lowest-latency cloud services for Agent applications.

The partnership with MediaTek is essentially an attempt by Google to restructure its TPU supply chain, which is expected to help the company cope with future computing power price wars and build a moat through lower production costs and higher yields.