Fubon: Google Plans 12–15 Million TPU v9 Chips by 2028, Intel Foundry Required
A Fubon Research note has put a number on Google’s 2028 TPU ambitions: 12 million to 15 million ninth-generation chips. For comparison, Fubon projects Nvidia’s data center GPU shipments at approximately 12.4 million units in the same year. If Google reaches its target, a single customer would deploy AI accelerators at a volume approaching the market’s dominant merchant supplier.
Fubon’s text: “Based on our checks, Google plans to have 12–15 million TPUs in 2028. Entering 2028, Google’s TPUs will enter the V9 generation with four compute dies, suggesting that their capacity consumption will more than double in 2028 versus 2027.”
The v9 Architecture Step
Moving from three to four compute dies per package is a significant manufacturing step. It requires advanced die interconnect and packaging at scale. Google has been using Intel’s EMIB (Embedded Multi-Die Interconnect Bridge) technology to stitch large multi-die packages together. For v9, Fubon says TSMC alone cannot meet Google’s production targets — Intel Foundry is required.
Intel reportedly holds orders for approximately 3 million TPU v9 chips. Running two foundries for a single customer’s internal chip program is unusual at this volume and signals that Google is treating supply chain resilience as a first-order constraint, not an afterthought.
The Scale Comparison
| Accelerator | 2026 projected units | 2028 projected units | Source |
|---|---|---|---|
| Google TPU (all generations) | — | 12–15 million | Fubon Research |
| Nvidia data center GPUs | 8.2 million | 12.4 million | Fubon Research |
The comparison requires care. A TPU and an H100 or Rubin GPU are different units of compute. Performance per accelerator varies significantly by workload, memory architecture, and interconnect topology. Direct unit comparisons do not establish equivalent capability.
What they do establish: Google is building toward compute self-sufficiency at a volume that removes its exposure to the Nvidia supply chain for internal AI workloads. The v9 generation is the first time an analyst note places Google’s deployment volume in the same range as Nvidia’s merchant sales.
Alphabet’s Capex Context
Alphabet spent $44.9 billion in Q2 2026 capex and $80.6 billion in the first half of the year. The company is explicit that the primary use is technology infrastructure — servers, data centers, networking, and the supply chain that feeds them. Not all of that is TPU program spending. Google is still procuring Nvidia hardware in parallel for workloads where GPUs are better matched.
Google first announced eighth-generation TPUs in April: TPU 8t for training and TPU 8i for inference, with general availability planned for H2 2026. The fact that Fubon is modeling v9 production volumes for 2028 suggests foundry and supply chain discussions are already underway — standard lead time for custom silicon at this scale is roughly two years.
Performance comparisons between TPU v9 and Nvidia’s next generation (Rubin, Rubin Ultra) are not yet available. What the Fubon note adds to the picture: Google’s bet on custom silicon is not a hedge against GPU scarcity. It is a structural infrastructure strategy operating at a scale that could reshape the AI hardware supply chain.
Key Numbers
- 12–15 million: Fubon estimate for Google TPU v9 deployment in 2028
- 12.4 million: Fubon projection for Nvidia data center GPU shipments in 2028
- 8.2 million: Fubon estimate for Nvidia data center GPU shipments in 2026
- 4: Compute dies per TPU v9 package (up from prior generation)
- ~3 million: Estimated Intel Foundry orders for TPU v9
- $80.6 billion: Alphabet H1 2026 capex