What if the smartest thing AI has done so far is reveal how fragile our idea of “cheap” really was? For years, executives treated software like sunlight: abundant, invisible, and difficult to ration. Now the meter is showing up in the room. Powerful AI tools are burning through budgets, and the new discipline is not invention but restraint. That is the sort of reversal markets hate. They like miracles. They dislike invoices.
The latest evidence is blunt. Amazon, Walmart, Cisco, Uber and Meta are imposing usage caps, curbing wasteful use, and steering employees to cheaper models to control AI spend, according to the Financial Times via Digital Today. Anthropic and OpenAI are also shifting some services from flat subscriptions to token-based billing, which raises enterprise costs. This is not a side effect. It is the business model becoming visible. Once the fog lifts, the old fantasy of limitless use meets the hard edge of scarcity.
The hidden risk is not that AI stops improving. It is that improvement can make demand less disciplined. When a tool becomes more capable, people do not always use it more carefully; they use it more often. That is the old problem of cheap abundance. In engineering, a stronger engine can still fail if the fuel line is misjudged. In markets, a clever product can destroy its own pricing power by encouraging consumption that customers never planned for. The result is familiar: the vendor calls it adoption, the buyer calls it drift.
The clearest warning came from Workato. The software firm, with about 1,300 employees, saw AI spend jump 7x in a single day after Anthropic switched to token pricing in May, according to Huffington Post Italia citing the FT. Carter Busse, the company’s CIO, said: “We created a monster”. That sentence is memorable because it is honest. A monster is not always a threat from outside. Sometimes it is a process that has learned to grow faster than governance. Budgets, unlike models, do not scale by magic.
Token pricing is the sort of change that seems technical until it reaches finance. A flat subscription feels like a bridge: predictable, quiet, easy to cross. Token-based billing is more like a toll road in fog. The journey is the same, but every turn of the wheel costs more than the user expected. Anthropic and OpenAI are moving some services toward that model, and the market is learning what happens when usage becomes granular. People do not merely pay for capability. They pay for curiosity, repetition and laziness too.
Costi Perricos, Deloitte’s global generative AI leader, put the cultural shift plainly: “Compute costs are starting to enter the minds of CFOs and boards… consumers and businesses have been taught that AI is cheap or free, but it is not”. That sentence matters because it exposes the psychological error at the heart of the boom. If a tool feels free, the user behaves as if marginal decisions do not matter. But they do. History is full of systems that collapse because the price signal arrives late, after the habit is formed and the bill is already due.
There is also a game-theory problem hiding in plain sight. In any organization, if one team uses AI aggressively and another restrains itself, the reckless team may appear more productive in the short run. The savings are private, the waste is socialized. So usage creeps upward until someone with budget authority intervenes. That is why caps appear. They are not just cost controls. They are attempts to restore a prisoner’s dilemma back into a cooperative game. Without limits, everyone is rewarded for spending the common pool faster than the system can replenish it.
The AI story was always partly a story about psychology. Users were told the tools were cheap or free, then asked to build workflows around them. That creates a strange dependency. Once a firm has reorganized tasks around a new technology, it becomes harder to pull back, even when costs rise. The technology turns from novelty into infrastructure. Infrastructure is where optimism goes to be audited. If the road, power line or payment rail gets expensive, the dream is not disproved; it is re-priced.
Sam Altman has acknowledged that cost is now a serious issue. According to Digital Today citing the FT, he said cost has become a “huge problem” for customers this year, “a problem that didn’t exist at all last year”. That is a useful admission because it undercuts the notion that AI economics have settled into a stable path. They have not. Rapid change in user demand can be more destabilizing than improvement in model quality. The product gets better, but the budget gets less forgiving. Markets often confuse those two lines on the graph.
The investor instinct is to treat every growth curve as if it will continue until it becomes destiny. But probability has a habit of humiliating linear thinking. A system can look antifragile because usage is rising, while actually becoming fragile because every new user increases the strain on the underlying cost stack. The more people depend on the service, the more dangerous a pricing reset becomes. That is true in cloud computing, energy, and now AI. The bill is not merely a line item. It is a stress test.
Uber provides a useful example of how quickly enthusiasm collides with arithmetic. Huffington Post Italia, citing the FT, reported that Uber exhausted its entire 2026 AI budget by April and imposed a $1,500/month per-employee cap on certain AI tools. That is not a trivial adjustment. It says the firm is no longer treating AI as an open-ended experiment. It is treating it like any other constrained input. That is what mature systems do when they discover that a tool once marketed as transformative can also become a leakage point.
The broader market is also confronting the difference between demand and efficiency. Goldman Sachs analysts forecast token consumption will grow 24x by 2030, and they said that could worsen chip shortages over the next 12–18 months, according to the Financial Times via Digital Today. That forecast matters not because forecasts are certain, but because it shows the strain can move from software budgets into physical supply chains. When tokens multiply, the effect is not abstract. It reaches chips, power, and capacity. The internet taught investors to think of scale as weightless. AI is reminding them that scale has a material appetite.
OpenRouter’s data suggests usage has expanded about 250x since the start of last year, according to FT Chinese. Even without turning that into a precise forecast, the direction is enough. Rapid expansion can make every assumption look sensible right up until the moment it is not. That is how speculative systems behave. They work beautifully while the denominator is growing. Then they become more delicate when growth slows or costs rise. The flaw was never only in the product. It was in the expectation that pricing could remain invisible forever.
This cost pressure arrives just as frontier labs are confronting the public market. OpenAI posted a $20.9bn operating loss on $13.07bn revenue in 2025, with a net loss of $38.5bn after a one-time charge, and it made a confidential IPO filing, according to Forbes. OpenAI and Anthropic have both delayed IPO plans. FT Chinese reported that Anthropic is reportedly targeting a November listing at a $2tn+ valuation, while OpenAI listing is unlikely before 2027, per Sam Altman, and the company is weighing a private round at a $1.2tn valuation.
That sequence matters because markets do not merely price growth; they also price the friction around growth. Investors can tolerate losses when they believe the unit economics are improving. They become less patient when they suspect the business is selling a dream that becomes more expensive each time it is used. The public market tends to punish uncertainty in a different way than private capital does. Private investors can wait for the narrative to catch up. Public shareholders eventually ask the blunt question: who pays when the miracle becomes routine?
There is a quiet irony here. AI has made some tasks faster, but it has also made the economics more legible. Flat pricing hid the scale of consumption. Metered billing reveals it. Usage caps reveal it more. Once that happens, boards stop asking whether the tool is impressive and start asking whether it is controllable. That is the moment at which a growth story becomes a governance story. The software may still be smart. The company must now prove it is disciplined.
The lesson is not that AI is broken. It is that every powerful system eventually meets the constraint it tried to abstract away. Rivers look smooth until they hit a narrow channel. Engines look elegant until the fuel bill arrives. In markets, the most dangerous phase is not failure but normalization, when a product has enough users to matter and enough cost to irritate, but not yet enough pricing power to justify either. The next phase of AI will belong to those who can make consumption look less like a fire and more like a furnace: controlled, measured, and contained.
