Imagine a bakery that already sells more bread than every other bakery in the city combined, and the owner walks out front to announce that next year he expects to sell 70% more. Not hopes. Expects. That is roughly the energy Jensen Huang brought when he said Nvidia could grow revenue 70% year over year, and that he’s confident about it.
“I think we could grow 70% year over year. We’re confident about that,” Huang said. Analysts have Nvidia landing around $400 billion in revenue for its current fiscal year. Apply that 70% and you get somewhere near $680 billion. His stated reasons are straightforward: Nvidia’s position in AI and demand for its products that keeps outrunning what it can ship.
Why a toolkit reviewer cares about a chip company’s forecast
I spend my days testing AI tools. Agent frameworks, orchestration layers, vector databases, the endless parade of wrappers that promise to make your workflow effortless. Most of what I review has nothing to do with silicon. But every single one of those tools runs on compute someone else is paying for, and the price and availability of that compute quietly sets the rules for everything I evaluate.
When Huang says demand exceeds supply, that is not an abstract macro observation. That is the reason your API calls get rate limited during peak hours. That is the reason a startup’s generous free tier evaporates three months after launch. That is the reason the “unlimited” plan you signed up for in January has usage caps by June.
A 70% growth projection built on demand outpacing supply tells me the squeeze isn’t easing next year. If anything, it’s the opposite.
What this does to the tools you’re evaluating
Here is how I’ve started reading product pages differently. When a tool advertises aggressive pricing on a frontier model, I now assume that pricing is a marketing expense, not a business model. Someone is eating the difference between what you pay and what the compute costs. That someone is usually a venture-funded company buying market share.
That’s not automatically bad. Cheap compute subsidized by investors is genuinely useful while it lasts. But it changes how you should plan:
- Assume your per-token costs go up, not down. Efficiency gains in models have been real, but so has the demand curve Huang is describing. Budget for the pricing you’d tolerate, not the pricing you’re getting.
- Check whether a tool locks you into one model provider. Portability is the closest thing to insurance you have. Tools with a clean abstraction layer over multiple providers survive pricing shocks. Tools hardwired to one endpoint do not.
- Test the smaller model first. I’ve reviewed plenty of agent setups that default to the largest available model for tasks a mid-tier one handles fine. That default is a cost decision someone made for you.
- Watch for silent degradation. When compute is tight, providers sometimes route to cheaper models or lower quality settings without announcing it. If your outputs get worse and nothing in your prompt changed, that’s a real possibility worth logging for.
The honest caveat
Huang is the CEO of the company he’s forecasting. Confident public projections are part of that job description. I’m not treating $680 billion as a settled number, and neither should you. Analysts project, executives project, and both get revised.
What I take seriously is the underlying claim, because it matches what I observe from the tool side. Demand for AI computing exceeding available supply is not something Nvidia has to convince me of. I see it in every rate limit error and every quietly revised pricing page.
How I’d play it
If you’re building on AI tools right now, the useful move is boring. Instrument your usage so you actually know what you’re spending and where. Pick tools that let you swap the model underneath without rewriting your application. Run the cheapest model that clears your quality bar, and re-test that bar periodically because the cheap models keep getting better.
None of that is exciting advice. It’s the same advice I’d give about any dependency you don’t control. The difference is that with AI tooling, the dependency is enormous, the pricing is still unsettled, and the company at the center of it just told the market to expect demand to stay hot enough to support 70% growth.
Plan your stack like compute stays scarce and expensive. If it gets cheap instead, you’ll have built something efficient for no reason, which is a pleasant problem to have.
🕒 Published: