Google's custom silicon program has quietly become a four-vendor operation, and MediaTek just secured its most durable seat at the table. Reports from Taipei-based DigiTimes and TrendForce on August 26, 2026 confirm that Google's TPU v8 lineup will split its workload explicitly: Broadcom builds the training-oriented Sunfish variant, and MediaTek designs the inference-optimized Zebrafish chip.
Zebrafish Positions MediaTek For Alphabet's Inference Wave
Alphabet's shift toward inference is the strategic prize. Training runs are lumpy and concentrated, but inference bills grow with every Gemini query, every Search AI Overview, and every Workspace agent that fires off a background task. MediaTek CEO Rick Tsai told analysts in late July that the company's first data-center ASIC will enter mass production in Q4 2026, with data-center revenue expected to top $2 billion this year — a figure that assumes Zebrafish ships in serious volume.
Broadcom Keeps Training, Marvell Wedges In
The v8 split leaves Broadcom holding the training role it has anchored across earlier TPU generations, while Marvell's recently announced $12.2 billion warrant deal — which lets Google purchase up to 58.97 million Marvell shares at $206.58 through fiscal 2033 — gives Alphabet a third custom-silicon partner for adjacent workloads including networking silicon, SerDes, and high-speed I/O.
Why Google Wants Four Vendors
The multi-vendor structure is a negotiating tool as much as a technical hedge. Each partner knows the others exist, which caps ASIC design fees and keeps roadmaps competitive on power, area, and cost. It also insulates Google from a single foundry or IC-design vendor stumble as Alphabet accelerates its bid to displace NVIDIA GPUs inside its own data centers.
A New Reality For NVIDIA
Google's TPU 8t and 8i program is targeting 121 exaflops of aggregate compute in the next deployment wave — an internal number that reads directly as a $21 billion counter-move to NVIDIA's dominance in cloud inference. Combined with NVIDIA's Hugging Face acquisition, the MediaTek win crystallizes how AI's demand curve is being met by more than one architecture at a time.
Reporting based on coverage from DigiTimes, TrendForce and Futurum Group.