Google’s approach to custom silicon is moving incredibly fast. Not long ago, the company split its custom AI hardware into specialized lanes, offering separate chips optimized strictly for training or inference workloads. However, the industry is rapidly pivoting toward AI agents—systems capable of reasoning and executing complex tasks autonomously—which requires a completely different hardware philosophy. To tackle this, Google is reportedly partnering with MediaTek to build an ambitious, all-in-one TPU AI chip.
According to FundaAI, this next-generation silicon is part of the upcoming TPUv9 series and goes by the internal codename “Triggerfish.” Unlike previous generations, Google wants Triggerfish to handle both heavy AI training and daily inference workflows on a single piece of hardware, eliminating the need to jump between separate chips (via Wccftech and @jukan05).
Fusing the brain together
MediaTek is pulling off a pretty interesting engineering trick to make this hybrid approach work seamlessly. The semiconductor company is integrating a brand-new, dedicated CPU tile directly onto the same physical package as the main compute die. This specialized CPU acts like a traffic controller, quickly switching the chip’s internal focus between processing data streams and training models on the fly.
On top of the innovative layout, Triggerfish will feature a massive SRAM cache upgrade. This will boast capacities up to two to three times higher than previous chips. To handle the intense bandwidth requirements of autonomous AI systems, the hardware will also utilize next-gen HBM4E memory infrastructure. Mass production for the Triggerfish layout is expected to fire up in the final quarter of 2027, with volume shipments scaling wide into 2028.
Navigating supply chain gridlock
Triggerfish isn’t the only piece of silicon Google has cooking in the kitchen. The tech giant is also working on a companion chip named “Humufish.” While Google is designing the main compute core for Humufish internally, MediaTek is handling the backend input/output setup.
What makes Humufish particularly interesting to hardware enthusiasts is the manufacturing strategy. While the main production will likely remain at TSMC factories, Google plans to leverage Intel’s advanced EMIB packaging technology to put the components together. EMIB is generally viewed as a highly scalable, cost-effective alternative to traditional packaging schemes.
Spreading the manufacturing footprint across different companies is a smart business move. Right now, TSMC’s assembly lines are absolutely overwhelmed with massive hardware orders from industry titans like Apple, NVIDIA, and AMD. Turning to Intel Foundry services will help Google build a clever safety valve to bypass global supply chain bottlenecks.
The post Google and MediaTek Team Up on ‘Triggerfish’ AI Chip to Supercharge Agentic Hardware appeared first on Android Headlines.
