How Could Google’s TPUs Affect Nvidia Stock?

TL;DR

Nvidia faces a concentration risk because 39% of its revenue reportedly comes from two customers that can develop competing AI chips. Google’s specialized TPUs are presented as cheaper and more power-efficient for AI workloads, while Gemini 3 was reportedly trained entirely on them. The central investor question is whether major customers gradually reduce Nvidia purchases, potentially pressuring its growth and valuation.

Transcript

If you have a 401k or you own the S&P or the NASDAQ or straight out Nvidia stock, then what I'm about to show you could save you or cost you thousands, if not hundreds of thousands of dollars in your retirement. Nvidia's seemingly unstoppable dominance in AI chips is being challenged by the most unexpected competitors, their own biggest customers. ... Read More

Key Insights

• Nvidia’s customer concentration is a material business risk because two unnamed customers reportedly account for 39% of total revenue. One contributes 23% and the other contributes 16%, meaning purchasing changes by either customer could significantly affect Nvidia’s reported growth and market valuation.
• Nvidia’s profitability gives major customers a strong incentive to pursue custom chips. The company is described as earning a 75% gross margin, so cloud companies spending tens of billions of dollars may seek to remove part of that margin by designing processors for their own workloads.
• Google’s TPUs are specialized processors built specifically for the mathematics that powers artificial intelligence. Unlike flexible, general-purpose GPUs, TPUs focus on narrower workloads, which the presenter argues can provide substantial advantages in cost, energy consumption, and task-specific performance.
• Google’s custom-chip program began in 2013 after the company estimated that widespread Android voice-search usage could require twice its entire global data center capacity. That projected computing burden motivated Google, primarily known as a software company, to design its own silicon from scratch.
• TPUs are presented as four to 10 times more cost-effective and two to three times more power-efficient than Nvidia GPUs for relevant AI workloads. These advantages matter because electricity is described as both a major data center expense and a constraint on further computing capacity.
• Google’s seventh-generation TPU, called Ironwood, reportedly provides 100% better performance per watt than the previous generation. The presenter also cites SemiAnalysis as arguing that Google’s seventh-generation TPU is approximately on par with Nvidia’s newer Blackwell chip.
• Gemini 3 was reportedly trained entirely on Google TPUs without Nvidia chips. The presenter treats this as real-world evidence that custom processors can support highly demanding model-training workloads, rather than serving only small, limited, or experimental artificial intelligence tasks.
• Nvidia’s competitive position is not described as disappearing because GPUs remain flexible, capable processors that customers will continue buying. The proposed shift is more gradual: major cloud companies may direct selected workloads toward specialized internal chips, weakening Nvidia’s share of their total infrastructure spending.

Questions & Answers

Q: Why could custom AI chips threaten Nvidia’s stock?

Custom AI chips could threaten Nvidia’s growth because the companies buying large volumes of its GPUs are also capable of becoming competitors. The presenter says two customers generate 39% of Nvidia’s revenue, while major technology companies collectively planned more than $380 billion of spending in 2025. If these customers shift selected workloads to internal processors, Nvidia could face slower sales growth and a valuation reset.

Q: How concentrated is Nvidia’s customer revenue?

Nvidia’s own presentation reportedly shows that two unnamed customers provide 39% of total revenue. Customer A represents 23%, while another customer represents 16%. The presenter says industry observers overwhelmingly believe these customers are Microsoft and Amazon, but emphasizes that Nvidia does not disclose their identities. This concentration means reduced orders from either customer could have an unusually large revenue impact.

Q: Why are large technology companies developing their own AI chips?

Large technology companies have a financial and operational incentive to design their own chips because Nvidia is described as earning a 75% gross margin. A customer spending $100 on Nvidia processors is therefore paying for approximately $75 of Nvidia gross profit, according to the presenter’s explanation. Internal chips may lower costs, optimize performance for specific workloads, and provide a competitive advantage through cheaper computing.

Q: What is the difference between a GPU and a TPU?

A GPU is described as a flexible, general-purpose processor capable of handling many different tasks, similar to a Swiss Army knife. Google’s Tensor Processing Unit is presented as a specialist designed specifically for the mathematics that powers artificial intelligence, similar to a professional chef’s knife. This narrower design can improve efficiency and economics for selected AI training and inference workloads.

Q: How much more efficient are Google’s TPUs than Nvidia GPUs?

The presenter says Google’s TPUs can be four to 10 times more cost-effective than Nvidia GPUs for the same work and two to three times more power-efficient. A former Google executive is cited as saying the sixth-generation TPU was 60% more efficient than GPUs for AI. These figures are presented as workload-dependent advantages, not proof that TPUs replace GPUs in every application.

Q: Why does power efficiency matter for AI data centers?

Power efficiency matters because electricity is described as both a major operating cost and a primary constraint on data center expansion. Google, Microsoft, Meta, and Oracle reportedly face limits related to electricity availability. A processor that completes more AI work with less power can reduce expenses and allow companies to expand computing capacity without requiring an equivalent increase in electrical supply.

Q: Can Google’s TPUs train major artificial intelligence models?

Google reportedly trained Gemini 3, described as one of the most powerful artificial intelligence models, entirely on TPUs without using Nvidia chips. The presenter argues that this result demonstrates that TPUs can handle demanding training workloads rather than only small or specialized tasks. SemiAnalysis is also cited as considering Google’s seventh-generation TPU approximately comparable with Nvidia’s Blackwell chip.

Q: Does the rise of TPUs mean Nvidia’s business is over?

The presenter explicitly rejects the conclusion that Nvidia’s business is over. Nvidia GPUs remain flexible, powerful processors, and customers are expected to continue buying them. The risk is that major customers may allocate a growing portion of specific AI workloads to custom chips. Such a gradual shift could reduce Nvidia’s share of spending even if its products remain important and widely used.

Summary & Key Takeaways

• Nvidia became central to the AI boom because major systems, including ChatGPT, Gemini, and Meta’s Llama, were trained using its GPUs. The company reportedly generated $51 billion in quarterly data center revenue, but its success has encouraged large customers to consider custom chips that could reduce costs and dependence on Nvidia.

• The principal risk discussed is customer concentration. Nvidia’s presentation reportedly shows that two unnamed customers contribute 23% and 16% of total revenue, respectively. Industry observers cited by the presenter believe they are Microsoft and Amazon, although Nvidia does not identify them. Reduced purchasing by either customer could materially affect revenue expectations.

• Google’s alternative is its Tensor Processing Unit, a specialized chip designed for the mathematics behind AI. The presenter says TPUs can offer four to 10 times better cost effectiveness and two to three times greater power efficiency than Nvidia GPUs. Google reportedly trained Gemini 3 entirely on TPUs, demonstrating their suitability for demanding workloads.