Mizuho: Google TPU Shipment Forecast Raised to 35 Million Units
Miles Bennett
Mizuho's latest channel checks raise the 2028 Google TPU shipment ceiling above 35 million units, with cumulative 2026–2028 volumes approaching 50 million — a shift that moves the custom-chip pricing story beyond Broadcom's quarterly prints and into the full cloud-vendor supply chain.
Why are cloud vendors suddenly building their own chips at scale?
As AI moves from training to deployment, inference is expected to become the dominant workload — reaching 50%–60% of compute by late 2026.
This means → the buying metric flips from raw horsepower to cost per token, power draw, and total cost of ownership.
Custom chips (ASICs — silicon designed for a specific task) are not meant to replace GPUs across the board. They target large-scale, steady-state inference loads to cut operating costs systematically and reduce single-vendor dependence on Nvidia.
What makes Broadcom's bundling model so sticky?
Broadcom does not just sell chip design — it packages Ethernet solutions, switches, and rack-level networking into a single AI-infrastructure stack.
Citi's review confirms Broadcom reiterated AI semiconductor revenue exceeding $100 billion in fiscal 2027, with its AI networking-to-chip bundle ratio holding at roughly 30%.
In plain terms = buying Broadcom's chip means buying its network plumbing too. That bundling makes it very hard for a customer to swap out just the chip vendor.
How did MediaTek break into Google's TPU roadmap?
Morgan Stanley notes Google runs a dual-supplier strategy across TPU generations: v7 and v9 go to Broadcom; v8 and v10 bring in MediaTek.
MediaTek has raised its 2026 AI custom-chip revenue guidance to $2 billion; Goldman Sachs projects the addressable market at $70–80 billion by 2027.
This reflects a structural pivot — MediaTek is migrating from legacy smartphone silicon into high-value cloud ASIC design, a shift that will reshape its margin profile.
Does custom-chip growth hurt or help memory makers?
Rising custom-chip volumes do not weaken HBM (high-bandwidth memory — ultra-fast memory purpose-built for AI chips) demand; they spread it across more platforms.
As next-generation custom chips migrate to HBM4e, memory makers like Micron see their demand base diversify beyond the Nvidia-only pipeline.
This means → customer-concentration risk falls for memory suppliers, and the growth curve becomes more resilient.
Where does this chain ultimately land in the physical world?
Goldman Sachs forecasts custom-chip share in AI servers will rise from 38% in 2025 to 50% by 2027.
Scaling custom silicon must translate into deliverable rack capacity — second-order elasticity flows directly to TSMC advanced packaging, full-rack server integration, high-speed PCB substrates, liquid cooling, and power management.
In plain terms = the best chip blueprint is worthless until someone turns it into a rack you can plug in and cool. These "last-mile" physical links are the next bottleneck — and the next investment theme.
Content is for reference only, not financial advice.