Google TPU Output May Outpace Nvidia GPU Sales by 2028

A Bold Forecast for Google's Chip Ambitions

A new analyst projection is fueling speculation that Google's custom AI silicon could eventually outproduce Nvidia's flagship chips in sheer volume. According to Fubon Research, Google is on pace to manufacture more Tensor Processing Units (TPUs) by 2028 than Nvidia sells in its GPU lineup, a claim that has rippled through tech and financial media as a signal of how quickly the AI hardware landscape is shifting 1.

The forecast underscores a broader theme: custom silicon, once a side project for cloud giants, is becoming central to their AI strategy and financial future. Alphabet's TPU program, paired with its Gemini models and Google Cloud infrastructure, is increasingly viewed by analysts as an underappreciated asset that could reshape the company's cost structure and long-term valuation, rather than a niche engineering effort 3.

Manufacturing Bottlenecks and Partners

Behind the scenes, the physical production of these chips depends heavily on foundry partners, and Samsung appears to be playing an outsized role. Samsung is believed to already manufacture the input/output die for Google's TPUs, and reports now suggest the company is so overwhelmed with AI chip orders that it may need to outsource some of that TPU-related work to keep pace with demand 7. This dovetails with separate reporting that Samsung is also developing its own AI accelerator, known as GAIA, a move that could complicate Microsoft's Copilot+ PC certification push by giving PC makers an alternative silicon path 2.

Custom Silicon Becomes a Strategic Battleground

Google is not alone in betting on proprietary chips. Amazon founder Jeff Bezos has said custom silicon such as Trainium and Graviton could grow into one of Amazon's major businesses as AWS scales its AI infrastructure, reflecting how hyperscalers increasingly see chip design as a way to control costs and reduce reliance on Nvidia 4. Together, these moves illustrate an industry-wide push by Google, Amazon, and others to internalize hardware development rather than depend solely on merchant silicon vendors.

Competitive Pressure From China

The stakes around AI hardware and inference costs are heightened by fast-moving developments in AI models themselves. A new Chinese model has surprised Silicon Valley with its capability, adding urgency to the U.S. industry's efforts to justify massive infrastructure spending 5. Separately, Moonshot AI's latest release has drawn attention for strong coding performance, low-cost API pricing, and an open-weight release strategy, intensifying pressure on both American AI labs and policymakers in Washington who are watching China's rapid progress 6.

Why It Matters

Taken together, these developments point to an AI hardware market in flux: custom silicon from Google, Amazon, and Samsung is challenging Nvidia's dominance, foundry capacity is becoming a strategic chokepoint, and cheaper, more capable Chinese models are pressuring the economics of AI inference. If Fubon's projection holds, the balance of power in AI chip production could look markedly different within just a few years.