The AI Boom Has Entered Its 'Wait, Is This Worth It?' Era
The great AI cost panic of 2026 is upon us
A very brief history of AI—and, by extension, the debate over whether it is a bubble—might go something like this:
The age of scaling and praying (Q4 2022 - 2025): After ChatGPT swept across the internet, hyperscalers poured hundreds of billions of dollars into AI infrastructure. Despite enormous hopes for the technology, actual AI revenue lagged spending. There was a strong case that supply was outrunning demand, just as it does in every industrial bubble.
The age of agents (End of 2025 - ???): With the arrival of Anthropic's Claude Code, OpenAI's Codex, and other autonomous agents, corporate spending on AI went so berserk that Anthropic's servers started buckling under the load. Suddenly, the story flipped: Demand for AI was demonstrably outstripping the supply of compute, complicating the traditional bubble narrative.
The reality check (circa Q2 2026): After months of running up token bills in the millions of dollars, some companies started wondering whether the productivity gains from autonomous agents actually justified the cost. The question shifted from “Can AI generate demand?” to “Can AI backfill supply?” to “Hold on a sec, what are we spending all this money on?”
Thus, the center of gravity of the AI discourse has shifted from concerns about demand, to worries about supply, to a freakout over value. In the last few days, several executives of Fortune 500 companies have confided to reporters that they’ve been out spending on AI tokens like drunk gambling addicts slumped over at the Bellagio craps table. An AI consultant told Axios that one of its clients spent half a billion dollars in one month on Claude. That is roughly what the entire US population spends monthly on shampoo or contact lenses. Uber and Microsoft have reportedly discontinued some of their Claude Code licenses, as costs became unmanageable and productivity proved elusive. AI critics and bubble watchers are circling the news and claiming that this is the backlash that could take down the industry. Gary Marcus, the cognitive scientist who has been warning about AI since ChatGPT launched, wrote this week that “if enough other companies report the same [productivity disappointments], the bubble pops.”
So, what has happened here? And what level of a crisis is this for the trillion-dollar bet that the US economy has made on AI?
It’s worth backing up a half-step to recall how we arrived at this point. The shift from chatbots to agents that took place over the last six months is the central economic event in AI. No matter what story you’re trying to understand—whether it’s the surge in AI use at white-collar companies; the unprecedented increase in AI revenue at frontier labs, such as Anthropic; the degrading of Claude for some Anthropic customers; or even the stock boom in memory chips—everything comes back to the age of agents.
Compared to O.G. tools like ChatGPT, agents are practically another species of technology. While chatbots take one question and give one answer, an agent runs a loop: planning, calling up tools, retrieving results, updating context, and planning again. While this ability to loop back makes agents useful for complicated tasks, such as coding or complex data analysis, every step of every loop burns tokens, or units of computation. Add up all the steps, and you’ll find that agents eat tokens like mammals breathe oxygen. According to data from SemiAnalysis, the typical agent job uses 96,000 tokens before generating an answer, which is more text than the entire novel “The Great Gatsby.”
This explosion of token consumption has several consequences. First, tokens are money, and agents that use a lot of tokens owe a lot of money. The average business is spending 13x more on AI tokens than in January 2025,” said Ara Khazarian, the lead economist at Ramp. At several firms, AI agent spending quickly ran into the tens of millions of dollars—in many cases, more than the team of talented software engineers that some people feared that this technology would replace.
Second, to justify their surging token spending, companies initially conflated AI use and productivity. Firms such as Meta created internal leaderboards that ranked employees by token consumption, which encouraged workers to spend all day throwing projects at agentic AI without regard for efficiency, a practice dubbed “tokenmaxxing.” As token costs soared, the benefits from tokenmaxxing lagged. In many cases, it probably resulted in more work and more burnout for many teams. The engineering analytics firm Faros AI found that “code churn,” or lines of code deleted versus added, increased by more than 800 percent under high AI adoption.
Bubble watchers insist that this burnout and backlash could bring the AI spending boom crashing down to earth. And, since I have no special access to the future, it’s worth pausing for a moment to consider the possibility that they’re right. If agentic AI is broadly seen as a waste of money by the major Fortune 500 clients, the great revenue surge of 2026 will slow down and add to the already significant doubt that the hyperscalers can make back the money that they’re outlaying on chips, data centers, electricity, and AI engineers.
Rather than see the agent backlash as a clear sign that AI is a scam, or that it is doomed, it might make more sense to see this development in the context of a normal technological adoption curve. As Aaron Levie, the CEO of Box, has said, chief executives who valorized AI without understanding it were also subject to a special kind of AI psychosis—that is, the belief, disconnected from reality, that more AI is always better; that every line of code produced by agentic AI was ready to ship; and that each prototype that Claude Code spits out was another billion-dollar idea.
As SemiAnalysis’s Doug O’Laughlin told me in an interview last week, every new technology requires an extended period of trial and error, as organizations toggle between (a) not enough experimentation or spending, followed by (b) too much experimentation and spending, followed by (c) too dramatic a pullback, followed by (d) the repetition of steps (a) through (c), until firms figure out a long-term balance between labor spending and tech spending. Whether AI skeptics like Marcus are right that the bubble is about to pop depends entirely on a question that, as of today, nobody can definitively answer: Is the bill worth it?
This is just one of the questions that I put to O’Laughlin in a recent interview. The following is an edited transcript of our conversation. Among other things we talk about:
The panic over AI costs and its historical echoes
The state of the AI-bubble debate
The business lesson behind Anthropic’s historic success—what whether the company would already be doing $100 billion in annual revenue if it weren’t constrained by compute
Whether OpenAI is doomed to get boxed in by Anthropic’s enterprise business and Google’s consumer access
Why agents are to AI what the website was the internet
The AI Cost Panic
Derek Thompson: Everybody is freaking out about AI agent costs right now. JPMorgan recently published a note entitled “Al Token Costs are Eating Internet Profits Alive.” Several companies, including Shopify, Spotify, ServiceNow, and Roku said in their recent earnings calls that AI is surging as a share of operating expenditures. Now we hear that Uber and Microsoft have blown through their 2026 token budgets in a matter of months, and some of these companies are reportedly dropping their Claude contracts. Where is this heading?
Doug O’Laughlin: I want to take the historical view.
Before the Industrial Revolution, for a lot of companies, all costs were labor costs. Over time, with the rise of machines, companies had to think about capital equipment being a permanent part of the cost of doing business. Now we’re seeing the beginning of a new operating cost. Let’s call it automated intelligence. Even if labor isn’t displaced, this category of automated intelligence is going to be a permanent part of the picture, much the same way that traditional IT didn’t exist 100 years ago and now every company factors it into their operating expenses.
While automated intelligence seems incredibly expensive today, it’s going to make sense for many companies to spend a lot of money on agentic AI in the long run. The challenge will be figuring out the right ratio of labor costs to AI costs. Today there are no benchmarks. Everybody is trying to figure it out on the fly. But despite the news stories, I’d argue that numbers that seem big today might seem small tomorrow. Everyone is going to figure out the effective ratio of compute to labor.
Thompson: How is your company SemiAnalysis thinking about token spending and this ratio between labor costs vs. AI costs?
O’Laughlin: We’re a fast-growing company. But even we’re like, “man, this is a lot of tokens.” [In April, SemiAnalysis acknowledged in a newsletter that the company “reached as high as $10.95 million dollar annual spend rate” on Anthropic Claude tokens.] So we’re figuring out the ratio of compute to labor, just like everybody else. I think it’s important to stress again that agentic AI is not even a year old. It’s been five months! No one knows the right ratio. Clearly there is value to be derived. But if you just go crazy, there are places where you’re boiling the ocean for GPUs, and you’re clearly wasting money.
Tokenomics and Bubble Fears
Thompson: When I first changed my mind about AI being a bubble, I got a lot of notes from people arguing that the unit economics of AI were broken; that consumers weren’t paying full freight for tokens; that the whole thing was being supported by venture capital subsidies; and that AI demand was going to collapse when prices rose.
Do you buy this story?
O’Laughlin: No, I don’t buy it.
