AI agents are rewriting
the economics of computing
Powerful AI tools are burning through budgets, prompting a rethink of how the technology is priced ahead of frontier lab IPOs
Barely four months into 2026, Uber had already hit its annual AI spending limit.
“We budget based on what we know, but things are changing fast in this space,” said Praveen Neppalli Naga, Uber’s chief technology officer. “Every week there’s a tectonic shift.”
Uber was not alone. As companies raced to adopt AI agents, bosses introduced leaderboards and targets to encourage staff to use them. Their ability to carry out complex tasks with limited human input is “the real unlock for AI”, said Ara Kharazian, chief economist at fintech company Ramp.
That autonomy has made AI costs harder to predict. Agents can burn through tokens — the units of information a model processes — with little human oversight. One Amazon project ran 860 per cent over budget after incomplete tasks worked in the background for five months, costing more than $1mn, according to multiple people familiar with the matter.
By shifting some enterprise services from fixed subscriptions to uncapped usage-based pricing earlier this year, leading AI labs pushed the risk of longer workflow costs on to customers. Early adopters including Uber, Meta, Cisco and Walmart have responded with spending caps. Intensifying competition between labs has also brought down the price of newer models.
The bigger problem is how AI is priced in the first place. Most charges reflect the computing power consumed rather than the value created, prompting customers and providers to seek alternatives. Linking prices more closely to outcomes while securing durable revenues will be critical for the industry — and investors weighing anticipated initial public offerings from Anthropic and OpenAI.
“Pricing in tokens doesn’t make any sense,” OpenAI president Greg Brockman said at the September launch of its Astra model. “Price per task is what matters. Can you get the thing done at an appropriate price and appropriate speed?”
Tokenomics
Anthropic’s long-awaited IPO prospectus highlighted the challenging economics of building state-of-the-art AI models – alongside delivering warnings about the “existential risks” they pose. The start-up recorded an operating loss of more than $8bn last year as it increased spending on computing power and told investors its rapid expansion would require $518bn in infrastructure commitments over the coming years.
At the same time, growing demand for agentic AI, particularly coding tools, has driven up revenues at the leading labs in recent months. Yet, even as AI systems become more powerful, taking on increasingly complex tasks autonomously, every interaction with a large language model starts with the same basic process.
But this process cannot expand indefinitely. The amount of information a model can handle at once is constrained by its context window — effectively its working memory.
The GPT-3.5 model that sparked the ChatGPT revolution in 2022 initially had a maximum context length of 4,096 tokens, roughly equivalent to a 3,000-word English-language essay or 20 minutes of spoken conversation.
Since then, context windows have expanded dramatically, allowing models to process far more information and enabling agents to tackle more demanding tasks. This has led to wider adoption, more token-intensive agentic workloads and greater use of more capable — and often pricier — models. Technological advances have drastically reduced the computational cost of each token, with the price for achieving the same level of AI performance falling roughly 13-fold a year since 2023 according to research group Epoch AI. But surging usage has outweighed those savings, racking up costs for companies.
Token use varies by task and model, but higher volumes generally mean greater demand for specialist chips, data centre capacity and electricity. That makes tokens a useful proxy for AI use — and a basic accounting unit of the AI economy.
“The best comparison to a token is a kilowatt-hour for electricity,” said Aleh Tsyvinski, a Yale economics professor who in July published a paper analysing the consumption habits of users who collectively consumed more than 380tn tokens across 400 different models. “It is fundamental to how we think about and measure AI’s development . . . this is the basic building block of AI.”
Companies use tokenisers — software that splits raw text into tokens — to train and run their models. The same sentence may produce a different token count depending on the tokeniser used. In every interaction with an AI system, tokens serve different functions and are priced differently depending on how efficiently they can be processed.
Cached tokens may be cheaper, but the sheer volume of usage is still pushing up costs for businesses adopting the technology. There are also other ways to manage long conversations. Systems can summarise previous exchanges, discard material not required or retrieve previously stored information, reducing the context brought forward. But summaries, retrieved material and new outputs still consume tokens — and can build up over subsequent steps.
Unlike a chatbot that returns a simple answer, coding agents such as Anthropic’s Claude Code or OpenAI’s Codex can also plan tasks, inspect files, write and revise code, and check their own work. A single request can therefore trigger far more model activity — and consume far more tokens — than a conventional prompt-and-response exchange.
OpenAI’s enterprise data illustrates the scale of the shift. As of June, Codex was used by just 17 per cent of customers, yet generated 63 per cent of all output tokens across ChatGPT and Codex.
A study involving researchers from Stanford University, Massachusetts Institute of Technology, Google DeepMind and Microsoft found that coding agents consumed 1,000 times as many tokens as simpler coding interactions.
Much of this extra activity occurs before a final answer is produced, as models repeatedly process context accumulated during a task. Yale’s Tsyvinski said this new era of agentic capabilities was driving “exponential growth” in token consumption as individuals and companies look for increases in productivity. “Agents are not writing out a basic recipe prompted by you or me, they are collecting and analysing vast amounts of data.”
This iterative cycle has become a key driver of rising corporate bills. Models do not always take the shortest route, often trying several approaches before reaching a conclusion. “You can see a meandering path of how the models get to where they have found the answer,” said Aakanksha Chowdhery, a Stanford professor and former Google DeepMind researcher. “It feels magical, but it’s kind of amusing.”
Goldman Sachs forecast in May that agentic AI would drive a 24-fold increase in global token consumption by 2030, with enterprise applications accounting for the largest share.
So far agents have seen the greatest uptake among software developers, but they will need to prove useful to other parts of a business if vast investments in AI infrastructure are to pay off. “One big driver of how quickly agentic AI disseminates through the economy will be how quickly the non-technical parts of a firm adopt agentic workflows for their tasks,” said Ronnie Chatterji, OpenAI’s chief economist.
At OpenAI, agent use spread from research and engineering to legal, finance, HR and marketing in about five months. “We’re almost all now using it,” Chatterji said, though whether other companies adopt the technology as quickly “remains to be seen”.
The rate of AI spending is rising at a faster rate among the heaviest users, according to Ramp’s analysis of 70,000 businesses. Since the start of the year, monthly spend per employee has more than doubled among the top 1 per cent of companies, to an average of $7,200, compared with a doubling across its platform as a whole.
More difficult for companies is the unpredictability of those costs. The study on agentic token consumption found that token use could vary sharply. Across repeated attempts of the same task, the most intensive run typically consumed twice as many tokens as the least, and in some cases 30 times as many. Human estimates of task difficulty were a poor guide, while frontier models also tended to underestimate how many tokens they would need.
AI agents were “not trained to be budget-aware”, said Jiaxin Pei, assistant professor at the University of Texas at Austin and one of the paper’s authors.
Taming costs
Despite several high-profile incidents — and a bout of “tokenmaxxing” in which staff faked activity to climb internal leaderboards and boost usage metrics — Amazon has pressed ahead with its enthusiastic AI rollout.
But Uber is among a number of companies that have imposed spending limits. The ride-hailing group has introduced token budgets, capping monthly spending at about $1,500 per AI coding tool.
AI usage has increased ninefold since February, the company told investors in August. In response, Uber has sought to cut costs by using services such as OpenRouter to switch between models and choose cheaper, more efficient options for different tasks. It is also caching data and asking engineers to write more focused prompts and clearer instructions — including where to find files — to avoid wasting tokens on routine work.
“Our spending has stabilised,” Neppalli Naga said. “In April [rising usage] would have been a big problem for us. We look at budgeting problems as engineering problems.”
Uber is also deploying “open” models, which can be customised and run at a fraction of the cost of the more sophisticated “closed” models from OpenAI and Anthropic. As their performance narrows the gap with leading closed models, these cheaper systems are gaining market share. Some of the most efficient have been developed by Chinese AI labs including Alibaba, DeepSeek, Moonshot AI and Zhipu AI, also known as Z.ai, drawing increasing scrutiny from regulators.
Airbnb chief Brian Chesky last year said that the company preferred to use Alibaba’s “fast and cheap” Qwen model for certain workloads, triggering a wider conversation about the defensibility of leading US models.
George Hu, president of Nvidia-backed AI infrastructure start-up Fireworks, compared using the most powerful AI for every query to asking a company’s highest-paid employee to perform every task. “The idea that every task should be done by a super expensive model doesn’t match anything we understand in the real world,” he said.
Smaller models fine-tuned for industries such as law and finance — or even individual companies — could make AI cheaper to run. Trained on specialist or proprietary data, they may require less computing power while outperforming general-purpose frontier models on narrow tasks.
As the top AI companies prepare for multitrillion-dollar IPOs and compete for enterprise customers, they are beginning to reduce costs for their top-tier models. Claude Opus 5.5 costs 40 per cent less to use than its predecessor, while OpenAI’s Sol and Luna GPT-6 versions are 50 per cent cheaper than their previous iterations.
The breakneck speed of model development, combined with usage-based billing, makes future revenues very hard to predict for Anthropic and OpenAI. “At this point, if you are looking at 2027, no one knows what they’re making,” said Joey Brookhart, analyst covering AI labs at research firm SemiAnalysis.
As well as competition from cheaper Chinese open models, a strong new release from a US hyperscaler such as Google or Meta that threatens the dominance of Anthropic and OpenAI and forces prices down could prompt a “much different conversation about these AI labs and the IPOs very quickly”, Brookhart added.
But bullish investors argue that surging demand for AI will allow both labs to thrive even as prices fall, with rising usage more than offsetting lower prices.
Yet the fundamental problem of measuring returns on AI usage remains unresolved. Current systems are akin to judging employees by time spent in the office rather than what they accomplish, said Quanyan Zhu, a professor at NYU Tandon School of Engineering and author of a paper on token economics.
Zhu and other researchers expect some pricing to shift towards results — or service-based — models, where customers pay for completed tasks or agreed outcomes rather than the tokens consumed delivering them.
OpenAI is already testing such an approach, following companies including Salesforce and Sierra. As businesses move towards a more disciplined use of AI, they will likely need to go beyond tokens and work out what AI output is worth to them, said OpenAI’s Chatterji.
“This idea that all tokens are the same is just an artefact of us being able to count them,” he said.