Cheaper intelligence, greater demand: The message from Nvidia

Two days after I published “Six things investors may be getting wrong about AI”, Nvidia’s latest results provided a test of many of the ideas in the article. The most important takeaway wasn't simply that Nvidia expects revenue to grow about 70% next year. It was the explanation for why demand continues to accelerate even as the cost of producing intelligence falls.

Vera Rubin (Nvidia's latest GPU generation) is expected to deliver token costs 35 times lower than Grace Blackwell (earlier GPU generation). Ordinarily, an improvement of that magnitude might sound bearish for infrastructure demand. If every unit of AI work requires dramatically less compute, why would the world continue building so many data centres?

Efficiency is accelerating demand

Nvidia’s answer is that useful AI work is growing much faster than hardware efficiency. Customer forecasts apparently imply that demand could support approximately 100% growth next year, although Nvidia currently expects supply constraints to limit growth to around 70%. Significantly, its compute is reportedly fully utilised across every cloud it serves. 

Agentic AI may amplify this effect enormously. Jensen Huang estimated that an agent completing a task can require between 15 and 100 times more compute than a conventional human-prompted interaction. An agent does not simply answer a question. It reasons, plans, calls tools, evaluates the result and revises its approach. 

If companies eventually operate millions of agents continuously in the background, AI demand will no longer be limited by the number of human prompts. Machines will increasingly generate work for other machines. That could unlock significant demand. 

Open and closed models can grow together

The call also supported the idea that intelligence may bifurcate rather than simply commoditise. Nvidia said adoption of both open and closed models is “skyrocketing”. Frontier labs continue to grow rapidly, while open models are becoming foundational to enterprises, startups and sovereign customers building proprietary AI systems.

Companies may rent frontier intelligence when a task requires it, while using cheaper open models for routine, domain-specific or high-volume work. A frontier model might supervise a task while dozens of cheaper models perform the underlying steps. Open models may therefore dominate token volumes while frontier models retain premium economics. Both can succeed simultaneously, and both still require compute.

The demand investors cannot see

Perhaps the most surprising aspect of the call was how broad demand has become. Investors tend to focus on hyperscalers and frontier labs, but Nvidia’s business also supports NeoClouds, sovereigns and enterprise customers. These customers generated $40 billion of revenue in Q227 and grew 138% over the past year. It represented nearly half of Nvidia’s data-centre revenue during the quarter.

Jensen described this as the part of the market many investors do not see. These customers generally lack both the desire and the technical capability to design custom chips or assemble AI infrastructure themselves. They need a complete computing platform, making Nvidia’s full-stack offering particularly valuable.

It also named Australian company Firmus among the regional NeoClouds combining local land and power with Nvidia’s platform. This suggests the build-out is broadening from the companies creating intelligence to those applying it. It also reduces Nvidia’s dependence on capital spending by a small number of hyperscalers and frontier labs.

The financing flywheel

Nvidia is increasingly helping finance the ecosystem through equity investments, take-or-pay commitments, credit enhancements and partnerships designed to mobilise outside capital.

Management argues that customer demand is real and its risk is limited because Nvidia’s compute is fungible and can be redeployed. That may be true, but Nvidia is no longer simply observing demand. It is helping create the financial structures that allow customers to purchase more infrastructure. This does not invalidate the investment cycle, but it makes utilisation and end-customer economics more important than purchase orders, announced gigawatts or GPU shipments. Investors should watch token consumption, cloud AI revenue, GPU utilisation, rental pricing and whether customers continue earning attractive returns without progressively greater financial support from suppliers.

Overall, the call strengthened the demand outlook. Intelligence is becoming dramatically cheaper. Open and closed models are expanding together. Agentic systems are increasing compute consumption, while demand is broadening beyond frontier labs and hyperscalers.

None of this proves every data centre being built today will earn an attractive return. But it reinforces the larger question at the heart of the article. If intelligence becomes as cheap and widely embedded as electricity, how much of it will the world ultimately choose to consume? The early evidence suggests the answer may be far more than we currently imagine.