A new report from the well-known analyst Ming-Chi Kuo carries big news for MediaTek. The company has secured the exclusive design project for Google’s next AI accelerator. That chip is the enhanced version of the TPU v9, codenamed Triggerfish. It targets the ultra-low-latency inference needs of AI agents and reinforcement learning. Production is set for late 2027, with mass production in 2028. As a result, it could become a strong growth engine for MediaTek.
Built for Ultra-Low-Latency Inference
According to Kuo, Google is developing its ninth-generation Tensor Processing Unit on two tracks. The standard TPU v9 carries the codename Humufish. Meanwhile, MediaTek handles the enhanced version, codenamed Triggerfish.
The core task of the enhanced chip is clear. It tackles the two biggest pain points in AI compute today: the “memory wall” and the CPU transfer bottleneck. In short, it pushes the chip’s underlying inference ability much higher.
Key Hardware Upgrades
Compared with the standard TPU v9, the enhanced version brings several major spec jumps:
- Much larger SRAM: the built-in static memory rises to two or three times the standard version.
- Upgraded memory: it drops the standard HBM4 and moves straight to the newer HBM4e high-bandwidth memory.
- A new simulation die: this added die should accelerate reinforcement learning and the coordination between multiple AI agents.
By greatly expanding the SRAM, Google can keep the active working state of RL and AI agents fully inside the TPU chip. So the design avoids the hidden cost of shuttling data between the main CPU and the TPU. In turn, it maximizes AI efficiency during the ultra-low-latency decoding stage.
Process and Packaging
The core TPU v9 specs include TSMC’s 2nm process. For packaging, however, capacity is tight on TSMC’s CoWoS lines. Because of that, Google is also evaluating Intel’s EMIB-T packaging technology.
Lower Volume, but a 30% Higher Price
On shipments and pricing, Kuo offers some numbers. The mainstream standard TPU v9 will still take the larger share, with an estimated 4 to 5 million units. Yet MediaTek’s potential orders for the enhanced version look strong too, at roughly 1 to 2 million units.
Most important is the price. Because the enhanced version uses more complex packaging and higher-grade memory, its average selling price runs about 30% above the standard model. Kuo’s report frames this as both a high-margin win and a sign of MediaTek’s rising standing. The world’s top cloud giants now trust its custom ASIC skills. Notably, MediaTek had already entered Google’s TPU v8i project for lightweight inference. This new win pushes it higher still.
From Phone Chips to Cloud ASIC Heavyweight
The exclusive Triggerfish win means a great deal for MediaTek. It marks a real break from the old “mobile processor maker” label. Instead, the company has transformed into an indispensable compute architect for top global data centers.
In the past, tech giants looking for an AI ASIC design partner thought first of Broadcom, or of traditional design houses like Alchip and GUC. In recent years, though, MediaTek has shown sharp strategic ambition. It holds strong hardware IP, including advanced-process coordination and high-speed interface technology. On top of that, it enjoys the scale advantage of close work with top foundries like TSMC.
A Shift Toward Inference and AI Agents
On a technical level, Google’s choice to push an enhanced TPU v9 reflects a wider shift. The AI industry is moving from large-model training toward difficult inference and AI agents.
Future AI will not just answer a single question. Instead, it will need to reason like a human, with reinforcement learning, multi-step thinking, and agent coordination. Traditional chip designs would stall under all that traffic to system memory. By helping Google pack in 3x the SRAM plus HBM4e, MediaTek builds a high-speed “thinking buffer” right inside the chip.
Once confirmed, this partnership could add major revenue and profit for MediaTek in 2028. More than that, it lights another beacon for Taiwan’s chip-design strength on the global AI map.
Support Our Threat Intelligence
If you find our technology report and cybersecurity news helpful, consider supporting our work.
Buy Me a Coffee
PayPal
