The market's shift toward efficiency exposes the fragility of a cash-burning AI business model, forcing OpenAI and Anthropic to prove real value.
A year ago, the corporate attitude toward artificial intelligence resembled that of a newly minted heir arriving in Las Vegas: money was thrown around everywhere in the hope that something would turn a profit. It was the era of tokenmaxxing, where the goal was to inflate processing consumption just to make it onto a board slide. But the party is over. According to CNBC, companies are now tightening their AI budgets and obsessively focusing on return on investment. The market's great shift is no longer about which model has the greatest cognitive capacity, but about which delivers the lowest cost per solved task.
This shift in tone exposes a structural fragility that venture capital made a point of ignoring: the business model of cutting-edge AI startups is, until proven otherwise, a massive subsidy. OpenAI and Anthropic burn billions to train and run models, yet operate on margins that would depend on infinite growth in corporate consumption. When the customer finally does the basic math and realizes the token bill doesn't pay for itself, these companies' growth rates slow down. The capacity hype has hit the cold wall of efficiency.
In practice, the market is maturing at the worst possible time for those selling dreams. The transition from "we need to have AI" to "we need AI to actually cut costs" forces a ruthless audit of the technology. Fact: companies no longer want the smartest model in the world; they want the cheapest model that solves the problem reliably. This is terrible for those who priced their product as a magical commodity and justified astronomical valuations based purely on server expansion.
In my opinion, what we are witnessing is the end of the technological "sick leave" phase. AI companies had to subsidize market adoption with unrealistic pricing to prove the technology worked. Now that customers demand efficiency, the bill comes due. If OpenAI and Anthropic fail to demonstrate tangible business value—and not just laboratory feats—the current race will be downgraded to a commodity dispute where the winner is whoever has the cheapest data center, not the most poetic algorithm.
The final irony is that efficiency, the holy grail customers now demand, is the exact mechanism that could crush the margins of the AI providers themselves. By forcing companies to optimize every penny spent on processing, the market signals that artificial intelligence is not an infinite premium product, but a basic utility. And, as any energy company knows, utilities do not sustain trillion-dollar valuations on the charm of novelty alone.
Tokenmaxxing refers to the past corporate practice of inflating AI processing consumption just to show adoption, without regard for profitability. It is ending because companies are now tightening their AI budgets and obsessively focusing on return on investment (ROI) and cost efficiency.
These companies operate on a cash-burning, subsidized business model that relied on infinite growth in consumption. As customers demand the cheapest model that reliably solves problems, growth rates slow down, exposing structural fragilities and threatening their astronomical valuations.
By forcing the optimization of every penny spent on processing, the market signals that AI is a basic utility rather than an infinite premium product. Like energy companies, utilities generally cannot sustain trillion-dollar valuations based solely on novelty.