Sam Altman, CEO of OpenAI, said at the Intelligence at Work event that the increasing costs of operating artificial intelligence systems have become a serious concern for companies. He noted that, for the first time, customers are regularly raising the price of running AI systems as an issue, so OpenAI is actively seeking ways to make its models more efficient.
"People these days half jokingly, half seriously say: 'My company spent the entire 2026 budget in the first quarter. Could you make this a bit more efficient?'" Altman said.
A question that wasn't apparent earlier
According to Altman, this was not a central topic at the start of the year: customers then were satisfied with AI spending. Since then, however, cost levels have become one of the most important concerns, and OpenAI is working to deliver more value at lower cost with its models.
"We are continually working to provide more value with our models at lower cost. At the start of the year this wasn’t a topic; suddenly it has become a huge problem," he said.
"Tokenmaxxing" and concrete examples
Recent reporting has highlighted companies facing very large bills while trying to maximize their AI usage. The trend has a name: "tokenmaxxing," where firms intentionally maximize AI-model usage in the hope of boosting employee productivity and revenue.
Jensen Huang, CEO of Nvidia, has said engineers should consume at least the equivalent of half their salary in AI tokens per year or he would worry about efficiency. Peter Steinberger, developer at OpenClaw, reported that his team spent $1.3 million on OpenAI API tokens in a single month — roughly 603 billion tokens.
High usage doesn't always pay off
Early experience suggests that unlimited AI use doesn't always deliver expected returns. Some Amazon employees admitted using AI agents for unnecessary tasks simply to rank higher on internal AI-usage leaderboards. Microsoft has reportedly limited distribution of Claude Code developer licenses after costs rose. Even the CEO of Uber has said there is no clear link at present between extremely high AI spending and successful product development.
These examples are prompting companies to ask whether massive AI budgets are really yielding proportional business outcomes.
Token consumption keeps accelerating
Altman nonetheless expects token consumption to accelerate over the long term. He said that about six and a half years ago, OpenAI’s largest token-using customer consumed roughly 100,000 tokens per month; today that amount is approximately equal to the global per-capita average token usage. By contrast, current top customers use on the order of 100 billion tokens per month, and Altman acknowledged—somewhat embarrassed—that he knows users who consume even more.
He suggested that if growth continues at a similar pace, it is conceivable that the world’s per-capita average monthly token usage could reach the order of 100 billion.
Prices must fall faster than usage grows
That scenario is only realistic if token prices fall faster than usage increases. Several companies have concluded that, in some cases, it's cheaper to pay humans than to run certain tasks on AI.
This dynamic echoes the Jevons paradox: when a resource becomes cheaper to use, consumption tends to rise rather than fall. Although model training and operation are becoming more efficient, increasingly sophisticated agent-based systems are driving exponential growth in token demand. At present, usage growth appears to outpace the efficiency gains AI labs are able to deliver.
Source: tomshardware.com
