The conversation around Nvidia Corp (NVDA) on Wall Street tends to focus on one question: how massive is AI demand? But here's a more interesting question that almost nobody seems to be asking: how long can Nvidia maintain 70%+ margins before its biggest customers decide they've had enough?
The threat isn't another company building a better GPU. The real danger is what Alphabet Inc. (GOOG) (GOOGL) has already done with its TPU—and what that signals about the future. It represents the moment when hyperscalers stop paying someone else to own the most lucrative piece of the AI infrastructure puzzle.
Custom Chips Aren't About Winning—They're About Margin Control
Google isn't building TPUs to prove it can design prettier hardware than Nvidia. It's scaling them because sending billions in compute spending to Nvidia every quarter isn't a sustainable strategy. With TPUs, Google runs AI workloads on its own terms, using its own infrastructure, at costs it controls directly.
Hyperscalers have learned the lesson that Apple Inc. (AAPL) taught the tech world years ago: if you own the platform, you shouldn't be paying premium margins to suppliers forever. It's basic business survival.
Here's the thing—TPUs don't have to outperform Nvidia's GPUs across the board. They just need to reach "good enough" for massive internal workloads while costing a fraction of what Nvidia charges. That's exactly how pricing power starts to crumble. Not in a dramatic collapse, but gradually, one training job at a time.
When One Hyperscaler Moves, The Rest Will Follow
This is the scenario that Nvidia bulls consistently brush aside. Once hyperscalers realize that custom silicon delivers structural improvements to gross margins, none of them will voluntarily stay on the sidelines as the company still writing full checks to Nvidia.
Amazon.com Inc (AMZN) is already deploying Trainium and Inferentia. Meta Platforms Inc (META) has MTIA. Microsoft Corp (MSFT) is backing Maia. This isn't some hypothetical future scenario—it's already in motion. Nobody wants to be the last player left paying the full GPU premium when everyone else has found a workaround.
The critical insight here is that Nvidia doesn't need to lose compute market share to lose its margin dominance. All it takes is for hyperscalers to build alternatives that are credible enough to establish a price ceiling—a reference point that makes Nvidia's pricing negotiable rather than inevitable.
What This Means For Investors
Make no mistake—the AI demand story remains rock solid. But the AI pricing power story? That's a different narrative entirely. Nvidia isn't facing the risk of obsolescence. It's facing the risk of becoming negotiable.
Once hyperscalers have genuine alternatives and real negotiating leverage, the idea of "70% margins in perpetuity" shifts from being an article of faith to something that sounds quaint in hindsight. The demand for AI compute will continue growing. But who captures those margins—and at what rate—is a question that's very much still being written.
