Nvidia posts another record as demand for tokens explodes

Nvidia has rewritten its own records again. The company reported results for its latest quarter that beat analyst expectations and confirmed the central thesis of the past two years: the AI race is gathering pace, not slowing down. CEO Jensen Huang summed it up in one line: "Global demand for tokens has become fully exponential."

To see why that statement reads like a financial diagnosis of an entire era, it helps to understand what a token is in the context of language models. A token is the basic unit that systems such as GPT-4 or Claude work with when processing and generating text. Every request to a model, every chatbot reply, every API call from a corporate application adds up to billions of tokens that have to be processed somewhere. And processing them takes Nvidia GPUs. That chain is what delivered yet another triumphant quarter for the company.

The context matters as much as the numbers. Over recent quarters, the world's largest tech companies (Microsoft, Google, Amazon, Meta*) have steadily raised capital expenditure on building and expanding data centers. The sums sound almost implausible: each of these companies spends tens of billions of dollars a year on AI infrastructure. And none of them plans to ease off. Quarterly reports keep coming, and each time capex forecasts go up. For Nvidia, that means steady, predictable, large-scale demand for its core products: the H100 accelerators and the next-generation Blackwell.

Notably, Huang attributes the company's success less to a competitive edge in hardware than to a fundamental change in how AI economics work. Infrastructure spending in the tech industry used to be largely cyclical: companies built capacity for specific workloads and then stopped. Today the logic is different. Language models are not simply used; they are constantly retrained, fine-tuned and scaled. Every model improvement requires more compute. Every new generative AI application creates additional token traffic. The system feeds itself, and demand for compute is growing faster than supply can keep up.

This has several important implications for the industry. First, Nvidia is cementing its role as the infrastructure monopolist of the generative AI era, a company without whose products no serious AI system can be built. Competitors exist: AMD is actively developing its accelerators, Google builds its own TPUs, and startups such as Cerebras and Groq offer alternative architectures.

But the CUDA ecosystem, which Nvidia has built up over years, creates a barrier to entry that is extremely hard to overcome in the short term. Second, the record capex of the big players means the whole market is betting on long-term monetization of AI products. Companies would not spend hundreds of billions of dollars on infrastructure without being confident that the investment will pay off.

For end users and developers, the picture is mixed. On one hand, growing infrastructure investment ultimately translates into faster, smarter and more affordable AI services. On the other, the concentration of the compute market in one company's hands creates risks for both pricing and hardware availability. The GPU shortage that became a byword in 2023 has not fully gone away; it has simply shifted to newer chip generations.

Nvidia's story over the past two years is not just the story of one successful company. It is a mirror of the current moment in AI: a period in which gigantic investment is outrunning our understanding of how exactly it will pay off. As long as demand for tokens stays exponential, Jensen Huang and Nvidia will remain at the center of the equation. The main question is not whether growth will continue; it will. The question is who will ultimately skim the cream from trillions of processed tokens.

*Meta has been designated an extremist organization and is banned in Russia.