NextFin News - OpenAI and Anthropic are running into a new constraint that is less glamorous than frontier-model performance and more important for business: customers are starting to care not just about how powerful a model is, but how efficiently it uses tokens, compute and budget. That shift is beginning to change the economics of the AI boom. It also raises an uncomfortable question for the two labs that have become the market’s clearest private-company standouts: can growth keep compounding at the same pace once enterprise buyers start tightening the screws?
The Market Is Repricing AI From Brute Force To Efficiency
The immediate issue is not demand disappearing. It is demand becoming more selective. Enterprise customers that once treated larger token bills as a sign of serious AI adoption are now asking whether they are getting enough value for what they spend. That matters because OpenAI and Anthropic have both built their recent growth on a simple formula: more users, more usage, more inference, more revenue. If those same customers start capping spend, routing work to cheaper models or shifting lower-value tasks to open-weight alternatives, the growth curve gets harder to sustain.
The concern is visible in the numbers that have been circulating around the two companies. Anthropic last reported a $47 billion annualized run rate in May, up from roughly $10 billion in revenue for all of last year. OpenAI’s run rate was said to be pacing closer to $25 billion earlier this year, up from $13.1 billion in revenue in 2025. Those figures are eye-catching, but they also underline the problem: the higher the base gets, the harder it is to keep compounding at the same speed.
That tension is showing up not only in customers’ behavior but also in the market’s conversation about AI infrastructure. Companies are now talking less about maximum usage and more about return on investment, model routing and work-specific efficiency. Microsoft, Amazon and Google are all leaning harder into products that emphasize cheaper or more efficient model selection. In practice, that means the market is moving from a “use the biggest model for everything” mindset toward a more disciplined allocation of tokens and compute.
Some of the clearest evidence is anecdotal, but it is still revealing. One AI startup chief said his company moved its entire traffic off Anthropic’s Claude models and onto DeepSeek, a cheaper open-weight alternative, and said the cost curve went “down, like, crash to the ground.” That kind of shift does not require a mass exodus to matter. It only needs enough large customers to start testing alternatives to pressure the economics of premium-model usage.
The broader implication is straightforward: the AI market is no longer being measured only by who can sell the most powerful model. It is increasingly being measured by who can make deployment efficient enough for corporate finance teams to tolerate.
Why CFOs Are Suddenly In The Driver’s Seat
The new pressure point is the chief financial officer. AI spend has often been treated like software spend, but in many enterprises it behaves more like variable infrastructure cost. Usage can spike quickly, forecasting is hard, and the bills are tied to how much work employees and applications send through the models. That makes token spend visible in a way that earlier-stage software adoption was not.
Gil Luria, an equity analyst at D.A. Davidson, said the biggest customers may begin limiting what he called their “out-of-control token spend.” His argument is not that AI adoption will stop. It is that the phase of reckless expansion is giving way to budget discipline. That distinction is important. A company can keep using AI and still reduce the amount of money flowing to the most expensive providers if it is routing more routine tasks to cheaper models or smaller, specialized systems.
“Some of their largest enterprise customers may start limiting their out-of-control token spend.”
That is exactly why the efficiency narrative matters so much. If a customer can get 80% of the performance at a fraction of the cost, the purchasing decision changes. The premium models still win on the hardest tasks, but they lose the default position on everyday work. Over time, that can compress usage growth even if AI adoption remains strong in headline terms.
There is also a practical finance issue. Companies usually budget annually, but AI usage can grow much faster than planning cycles. One market participant described the problem as CFOs not having great tools to manage this new category of spend. That is a structural vulnerability for premium AI providers: the more successful they become, the more likely customers are to create internal controls around them.
That is why the current phase looks less like a demand collapse and more like a normalization. The first wave of enterprise AI adoption was about proving that usage could scale. The second wave is about proving that usage can scale efficiently enough to survive budget scrutiny.
The Competitive Threat Is Not Just Open Source
Open-weight and open-source alternatives are part of the pressure, but they are not the only force. Big tech platforms are also pushing efficiency-focused offerings of their own. Microsoft has announced lower-cost models and says its coding products route users to the most appropriate model for a task. Amazon and Google are also emphasizing efficiency in their AI product strategy. That matters because enterprise buyers often prefer a familiar cloud vendor if it can solve the same problem at a lower cost and with simpler procurement.
At the same time, startups built around token efficiency are making a business case directly against brute-force model usage. Engram, a memory-focused startup, raised $98 million to build models that it says can match or outperform frontier systems using up to 100 times fewer tokens. The company’s pitch is an explicit attack on the assumption that better AI necessarily means more expensive AI.
The market is therefore seeing a layered response. On one side, frontier labs are racing to build the most capable models and secure the biggest enterprise relationships. On the other, customers, cloud providers and smaller startups are all trying to make those models cheaper to use. The result is a market where capability remains valuable, but efficiency is becoming a gating factor for adoption.
That should not be mistaken for a loss of strategic relevance for OpenAI or Anthropic. Their models still sit at the top end of the market, and their brand value is enormous. But brand value does not guarantee pricing power forever. In software and cloud markets, the provider that starts as the obvious default can eventually become the expensive default, and once that happens, procurement teams begin to redesign the stack around cost.
The key risk is that the companies’ growth rates may be peaking just as they are trying to justify the massive compute and capital commitments that frontier-model development requires. That is the same tension investors are now applying to the broader AI trade: if spending is rationalized faster than model monetization scales, the economics tighten even if the technology keeps improving.
What This Means For The Next Phase Of AI
The most important implication is not that AI is losing momentum. It is that the spending pattern is maturing. Early-stage enthusiasm rewarded scale for its own sake. The next phase will reward systems that can prove a better output-to-cost ratio. That shift favors model routing, memory systems, domain-specific tools, and product designs that minimize unnecessary token usage.
For OpenAI and Anthropic, that means the growth story may remain impressive but less linear. Customers will still pay for frontier performance where it matters, but they are likely to become more selective about when they use it. The companies that can preserve pricing power while lowering perceived waste will be in the best position. The ones that depend on maximum usage everywhere may find that enterprise finance teams become their toughest customers.
The next catalysts are straightforward: new enterprise contract data, any disclosure about usage growth or monetization mix, and evidence that cloud and model-routing strategies are actually lowering customer bills. If efficiency-focused offerings keep expanding while open alternatives keep improving, the market will continue to move away from tokenmaxxing and toward a more disciplined model of AI consumption.
The biggest lesson is that the AI race is entering a phase where growth is no longer enough on its own. In the next chapter, the winners will not just be the models that think best. They will be the ones that help customers spend least for every useful answer.
Explore more exclusive insights at nextfin.ai.
