Nvidia’s $750B AI bet: a circular bubble or a real shift?

The circularity critique is the strongest argument against Nvidia today, and it’s also the easiest to oversimplify. The worry goes like this: Nvidia sells GPUs to hyperscalers, hyperscalers lend those GPUs out as cloud instances, AI startups rent them to train models, and then those models sell AI capabilities back to the same hyperscalers. Everyone is paying everyone with venture capital, and nobody is making real money from an end customer who actually needs the output. If the loop breaks at any point, you get a hard landing.

I get the fear. The numbers are genuinely absurd. We’re talking about a $750 billion wave of AI infrastructure spending, and the revenue supporting it still leans heavily on a handful of companies that can write billion-dollar checks. That’s not diversification; it’s concentration with a capex-driven circularity. If one of the big four clouds decides to decelerate, Nvidia’s growth story gets a brutal re-rating. So I don’t dismiss the bubble scenario. It’s plausible.

But the circularity argument misses something important: the actual usage inside those loops. I run a small team that uses LLM agents for code review and support triage. A year ago that was a toy. Now it’s part of our daily workflow, and the compute bill is real. Tools like Claude Code have genuinely cut my time to ship a feature. That demand exists, and it compounds. The problem isn’t that nobody uses these models. The problem is that the supply of capacity is being priced as if every organization on earth will adopt AI at once, which is a timing problem, not a value problem.

Here’s how I see the two sides:

• The circular argument: AI companies are each other’s customers. OpenAI buys Azure capacity, Azure buys Nvidia chips, Microsoft uses OpenAI models inside Office, and the profit eventually comes back to Microsoft. It’s a closed loop that inflates reported revenue without adding much to GDP.
• The fundamental argument: The compute underneath all of it—training clusters, inference pipelines, model serving—is becoming basic infrastructure like electricity or the internet. Even if 70% of the startups die, the surviving companies and the big platforms will keep buying capacity because that’s their operating cost now.

I lean toward a messy middle. We’ll likely see a shakeout in overvalued model labs and middleware companies. But the chip layer, especially if you’re looking at inference-heavy workloads, has a long tail of real demand. The "bubble" is in the abstract valuations of AI-native companies, not in the physical hardware. Nvidia sells the shovels, and the gold rush might end, but the shovels are still worth something.

What keeps me cautious is the sheer scale. A $750B bet means the market is pricing in near-perfect execution. If any major consumer of Nvidia’s products hiccups, the correction won’t be gentle. But a correction isn’t a collapse. The infrastructure will remain, the cooling systems will hum, and the next wave of cheaper, more efficient models will still need to run somewhere. I’d rather own the pick-and