Alphabet is developing an upgraded version of its TPU v9 AI chip, codenamed "Triggerfish," with Taiwan's MediaTek as its exclusive development partner, according to supply chain analyst Ming-Chi Kuo. The chip is expected to deliver significantly stronger inference capabilities for next-generation AI workloads such as AI agents and reinforcement learning. GOOGL stock fell 1% in premarket trading Monday, following a 2.3% gain last week that snapped a five-week slide.
What Triggerfish brings to the TPU lineup
Triggerfish builds on Google's existing TPU v9 platform, known internally as Humufish, but with a substantial memory upgrade. Kuo said the new chip will feature SRAM capacity two to three times larger than Humufish and support for next-generation HBM4E memory, compared with HBM4 in the current design.
The larger SRAM capacity allows more active workloads to remain on-chip, reducing data movement and improving efficiency during inference. That could help Google address growing bottlenecks in AI computing, often called the "CPU wall" and "memory wall."
Kuo estimates lifetime shipments of Humufish at roughly 4 million to 5 million units. Triggerfish is expected to add another 1 to 2 million units, with production starting in late 2027 and volume ramping in 2028.
MediaTek's expanding role in custom AI silicon
The reported move cements MediaTek's position as one of Google's preferred partners for the TPU v9 and underscores the Taiwanese company's increasing business ties with U.S. Big Tech firms. Over the past year, MediaTek has moved beyond its traditional smartphone roots and emerged as a strategic silicon partner for several major U.S. technology companies building custom AI hardware.
The company recently deepened its partnership with Nvidia for the GB10 Grace Blackwell Superchip and Nvidia's broader AI computing initiatives. MediaTek and Microsoft have also expanded cooperation around edge AI, showcasing cloud-to-device AI workflows using Microsoft's Phi models on MediaTek silicon.
While Broadcom and Marvell remain dominant in the custom AI silicon market, MediaTek's growing role with Google, Nvidia, and Microsoft positions it as a credible challenger. For developers working on AI infrastructure, the shift matters: more players in custom silicon means more variety in how AI workloads get deployed, and more pressure on Nvidia's pricing.
Google's TPU strategy and what it means for developers
Google has increasingly relied on custom TPUs as a key part of its AI strategy, using them internally for products such as Gemini while also offering them through Google Cloud. The company has emerged as one of the few hyperscalers pursuing a large-scale in-house AI-chip roadmap rather than relying solely on Nvidia accelerators.
That approach has implications for anyone building on Google Cloud: TPU availability, pricing, and performance directly affect what AI workloads cost to run. The memory upgrades in Triggerfish also signal where Google sees bottlenecks - inference workloads with large active datasets that need to stay on-chip to avoid slow, expensive data movement.
For those working in AI for IT & Development, the technical details matter. Larger SRAM and HBM4E support mean future TPUs will handle bigger models and longer context windows more efficiently. That translates to lower latency and cost for production inference, which is what teams building Generative AI and LLM applications will feel directly.
Why this matters for IT and development professionals
Google's custom chip roadmap is not just a hardware story - it's an infrastructure story. If Triggerfish delivers the claimed memory improvements, it will change the cost profile for running large-scale inference workloads on Google Cloud. Teams planning AI agent systems or reinforcement learning pipelines should watch TPU availability and pricing closely, since those workloads are exactly what the chip is designed to handle.
The timeline matters too. Production begins in late 2027 with volume ramping in 2028. That means decisions about cloud providers and AI infrastructure made now should account for what TPU capacity will look like in two to three years, not just what's available today.
Your membership also unlocks:
