Google is developing an upgraded variant of its next-generation Tensor Processing Unit, and MediaTek has secured the exclusive contract to build it. The chip is codenamed Triggerfish and sits within the TPU v9 family as an upgraded follow-on to the base Humufish design. Analyst Ming-Chi Kuo confirmed the arrangement on June 22, with production expected to begin in late 2027 and volume ramping through 2028.
Triggerfish builds on Humufish, though with some AI-targeted changes. SRAM capacity is two to three times higher than that of Humufish, allowing complex AI tasks to run on-chip rather than accessing slower memory. A new simulation die has been added, and the memory has been upgraded from HBM4 to HBM4E for higher bandwidth. This directly addresses the memory wall bottleneck that limits how fast AI accelerators can feed data to their compute units.
A dedicated CPU tile on the same package as the main compute die handles workload switching between training and inference. This makes Triggerfish the first Google TPU designed to handle both in a single chip rather than splitting them across separate silicon.
MediaTek previously handled I/O solutions and inference-optimized work on the TPU v7 and v8 generations, with Broadcom handling training-focused designs. Triggerfish marks the first time MediaTek has secured the exclusive order for a core Google TPU variant.

On shipments, Humufish is projected to reach 4 to 5 million units over its lifetime. Triggerfish adds an incremental 1-2 million units on top of that. The unit price premium for Triggerfish is around 30% higher than that of Humufish, making the smaller volume a more meaningful revenue contributor for MediaTek than the raw unit count suggests.
Humufish is also expected to use Intel's EMIB packaging rather than TSMC's CoWoS, with Intel handling packaging while TSMC manufactures the compute die. Reports of a three-million-unit Intel Foundry order for Google TPUs appear to refer to packaging rather than full wafer production. Google is now working across Broadcom, MediaTek, Marvell, and Intel for different parts of its AI chip supply chain.




