Google Could Deploy 15 Million TPUs by 2028, Challenging NVIDIA’s AI Scale
A new research note suggests Google is preparing one of the most ambitious AI infrastructure expansions in the industry’s history. According to a report from Fubon Research, the company plans to deploy 12 million to 15 million TPU v9 accelerators by 2028—potentially matching or even exceeding NVIDIA’s projected annual shipments of data center AI GPUs.
If these estimates prove accurate, Google would become not only the world’s largest operator of AI accelerators but also a leading force in vertically integrated AI infrastructure. The projected manufacturing scale would likely exceed the production capacity of a single foundry, potentially requiring support from both TSMC and Intel Foundry.
Although Google has not confirmed these figures, the report highlights the accelerating shift among hyperscalers toward designing and deploying custom AI silicon optimized for their own software ecosystems and cloud platforms.
🚀 Google Plans a Massive TPU v9 Rollout #
Google has invested in proprietary Tensor Processing Units (TPUs) for roughly a decade, making it one of the earliest hyperscale cloud providers to develop custom AI accelerators.
According to the Fubon Research memo, Google intends to transition to its TPU v9 architecture in 2028 while dramatically increasing deployment volumes.
The report estimates that Google will own between 12 million and 15 million TPU accelerators during that year.
Unlike previous generations, TPU v9 is expected to adopt a four-chiplet compute architecture, significantly increasing silicon complexity as well as manufacturing requirements.
Fubon also noted that this architectural transition is expected to more than double Google’s semiconductor capacity consumption compared with 2027.
📈 Deployment Scale Could Rival NVIDIA #
For comparison, the report estimates NVIDIA’s annual shipments of data center AI GPUs at approximately:
- 8.2 million units in 2026
- 12.4 million units by 2028
If Google’s projected deployment reaches the upper end of the reported range, its annual TPU production could exceed NVIDIA’s projected GPU shipments during the same period.
It is important to note that these figures compare Google’s internal deployment plans with NVIDIA’s market-wide shipments rather than direct product sales.
Nevertheless, the comparison illustrates how rapidly hyperscale cloud providers are expanding proprietary AI infrastructure.
🏭 TSMC Alone May Not Meet Production Demand #
Producing up to 15 million advanced AI accelerators within a single year would place enormous pressure on semiconductor manufacturing capacity.
According to the research note, TSMC alone may not be able to satisfy Google’s projected production targets.
Instead, Fubon believes Intel Foundry could become an essential manufacturing partner by 2028.
This assessment aligns with recent industry reports indicating that Google has been evaluating Intel’s advanced packaging technologies and may have placed orders for approximately 3 million TPUs manufactured through Intel.
While neither company has officially confirmed these reports, such a partnership would provide Google with additional manufacturing flexibility during large-scale production.
Advanced Packaging Becomes a Critical Factor #
The TPU v9 architecture reportedly uses four compute chiplets, making advanced packaging technologies just as important as wafer fabrication.
Unlike monolithic processors, chiplet-based accelerators require sophisticated interconnect technologies to deliver high bandwidth and low latency between compute dies.
This also means packaging decisions must be made early during chip design.
Intel’s EMIB and EMIB-T packaging technologies differ fundamentally from TSMC’s CoWoS-L, preventing packaging approaches from being freely interchanged after silicon has been designed.
As a result, any multi-foundry manufacturing strategy would require careful co-design of compute chiplets and packaging infrastructure.
🌐 Vertical Integration Continues to Reshape AI Infrastructure #
If Google ultimately deploys between 12 million and 15 million TPUs, it would become the largest owner of AI accelerators globally.
Even if Google continues purchasing NVIDIA hardware, the majority of its AI infrastructure would increasingly rely on internally designed processors optimized for Google’s own workloads.
This reflects a broader trend across hyperscale cloud providers.
Rather than relying exclusively on commercially available GPUs, companies are increasingly investing in custom silicon tailored for:
- Large language model training
- AI inference
- Cloud infrastructure
- Internal software frameworks
- Data center optimization
Custom accelerators provide greater control over performance, energy efficiency, deployment costs, and long-term platform evolution.
⚖️ Does This Threaten NVIDIA? #
A larger TPU deployment does not necessarily imply a decline in NVIDIA’s market position.
Global demand for AI computing continues to grow at an unprecedented pace, allowing multiple hardware ecosystems to expand simultaneously.
NVIDIA’s GPUs remain the dominant platform for enterprise AI due to their mature hardware ecosystem and software stack, while hyperscalers increasingly deploy proprietary accelerators alongside commercial GPUs.
Instead of competing solely on hardware shipments, the industry is gradually shifting toward platform competition.
For NVIDIA, the long-term strategic challenge may be less about Google’s deployment volume and more about the continued expansion of alternative AI software ecosystems that reduce dependence on CUDA.
As custom accelerators become increasingly capable, hyperscalers have greater incentives to optimize their AI frameworks around proprietary hardware rather than third-party GPU platforms.
🔮 Outlook #
Although the reported deployment targets remain unconfirmed, they underscore the extraordinary scale at which hyperscale cloud providers are investing in AI infrastructure.
Should Google achieve production volumes approaching 15 million TPU v9 accelerators, it would represent one of the largest deployments of custom AI silicon ever undertaken. Realizing such ambitions would also require unprecedented coordination across chip design, advanced packaging, foundry capacity, and software optimization.
Regardless of the final shipment figures, the broader direction is becoming increasingly clear. The future of AI infrastructure will not be defined solely by raw GPU performance but by vertically integrated ecosystems that combine custom silicon, high-performance networking, advanced packaging, and tightly optimized software stacks. In that environment, Google’s TPU roadmap represents not just a hardware strategy, but a long-term effort to reduce reliance on third-party accelerators while strengthening its position as one of the world’s leading AI infrastructure providers.