Nvidia’s Jensen Huang bets big on token economics as AI’s new currency

TL;DR: Nvidia CEO Jensen Huang used the company’s annual tech event to promote tokens as the fundamental economic unit of AI, arguing that falling cost-per-token favours Nvidia’s chips. However, the link between producing cheap tokens and generating real customer revenue remains largely unproven. The pricing collapse from $33 per million tokens two years ago to 9 cents today raises commoditisation concerns reminiscent of early cloud computing.

At Nvidia’s flagship annual event this week, Jensen Huang made a forceful case for a new way of thinking about AI economics. Rather than focusing on the billions being spent on data centres or the competition closing in on Nvidia’s margins, he argued investors should watch one metric: the cost per token of output.

Tokens are the basic units that large language models produce, with roughly 1,300 tokens equating to 1,000 words of text. Huang’s thesis is straightforward: as long as Nvidia’s chips produce tokens at the lowest cost and demand continues to outstrip supply, the AI boom remains healthy.

The gaps in the theory

The argument sounds compelling as a defence of Nvidia’s position, but two significant holes remain. First, there is no established link between producing tokens cheaply and creating value for the businesses buying AI services. Falling production costs do not automatically translate into useful products or customer revenue.

Complicating matters, newer reasoning models consume vastly more tokens than their predecessors. OpenAI’s o1 models, which emerged in late 2024, perform far more computation per answer. AI agents, which automate complex workflows, promise an explosion in token consumption and potentially enormous bills for companies giving workers unlimited AI access.

Some software engineering teams have begun tracking how token usage maps to productivity, but the broader vision of every white-collar worker carrying a monthly token allocation alongside their salary remains speculative.

The commoditisation question

The second gap is profitability. If every AI factory runs on Nvidia’s latest chips, no single provider can establish a meaningful cost advantage. The price trajectory tells the story: when OpenAI launched GPT-4 two years ago, it charged $33 per million tokens. Its cheapest model now costs 9 cents for the same volume.

This mirrors concerns from the early days of cloud computing, when sceptics questioned how Amazon Web Services could profit from selling basic computing as a commodity. AWS eventually built higher-value platforms on top of those raw services. Whether OpenAI, Anthropic and their competitors can achieve the same remains an open question.

Looking forward

For UK businesses planning AI budgets, the token economics debate matters practically. Falling token prices make experimentation cheaper, but the lack of proven links between AI spending and business returns should temper expectations. The real test is not how cheaply tokens can be produced, but whether they deliver measurable outcomes.