INSERT COIN

Organizations are discovering that AI token pricing produces exactly the kind of behavioral distortions that metered communications produced a generation ago. Uber Technologies burned through its entire 2026 AI budget by April, driven by Claude Code adoption that jumped from 32 to 84 percent of its engineering organization, with heavy users running up monthly API costs between $500 and $2,000 per engineer; the company’s CTO told The Information they were “back to the drawing board” on budgeting. Microsoft’s Experiences and Devices division is winding down most of its Claude Code usage before the end of June, with financial considerations cited as a primary driver alongside toolchain strategy. One company reportedly accumulated a $500 million AI bill in a single month after failing to set usage limits. The Linux Foundation has announced its intent to launch the Tokenomics Foundation, a new standards effort aimed at establishing open standards for AI cost management.

The paradox is simple: per-token prices have actually fallen. Token prices dropped roughly 80 to 90 percent over the past two years. Total bills rose anyway, because agentic workflows consume far more tokens per task than earlier uses did. A single moderately complex agent loop can consume hundreds of thousands of tokens before returning a result. Gartner predicts that by 2030, inference on a one-trillion-parameter model will cost providers more than 90 percent less than in 2025. Gartner also notes that falling token costs will not fully pass through to enterprise customers, because agentic workflows demand significantly more tokens per task than earlier uses did.

Teams are responding the way AOL subscribers responded in 1993. They are watching the counter. They are batching interactions to reduce token consumption. They are making discrete cost calculations about whether a given prompt is worth running. The behavioral signature of metered technology is unmistakable, and it is shaping how AI gets used in ways that have nothing to do with the underlying value of the technology.¹

This is not a stable state. The question is where the transition leads, and the answer depends on physics as much as economics.

We Have Been Here Before

We have been here before. Long-distance telephone calls were metered by the minute, then cellular airtime, then text messages, then internet bandwidth itself. Each arrived expensive and tightly measured; each eventually became effectively unlimited or flat-rate. The pattern feels inevitable in retrospect.

But that communications story leaves out an older and more directly relevant history. Before online services metered your connect time, mainframe operators metered your compute time. Businesses in the 1960s and 1970s paid for CPU cycles, memory, and I/O through time-sharing services like Tymshare and GE Information Services. They bought access to computation by the unit, not by the month. The business model looked a great deal like today’s AI token pricing.

Mainframe time-sharing did not commoditize to flat-rate. It was disrupted. The personal computer made centralized time-sharing obsolete before it could follow the communications path. The meter did not come off; the metered product ceased to exist. That is the less comfortable historical analogy for AI token pricing, and it belongs in the analysis alongside the cheerful communications story.

The Thermodynamic Floor

Every physical process has a thermodynamic floor: a lower bound on energy consumption that engineering can approach but never eliminate. In computation, Landauer’s principle captures this precisely: it establishes the minimum heat dissipation required to erase a bit of information, a limit experimentally verified and now well established. The floor varies by many orders of magnitude across different processes, and where it sits determines everything about how a technology prices its output over time.

For transmitting a bit of information through a fiber optic cable, the thermodynamic floor is nearly negligible. The energy cost per bit is tiny relative to the fixed cost of infrastructure. Once the cable is in place, the marginal cost of one more byte becomes economically invisible. At the consumer layer, metering became less attractive not because anyone got generous but because the marginal cost of transmission fell well below the administrative cost of measuring it.

Electricity is the counterexample. We still meter electricity because the thermodynamic floor is the product itself. When you run the dishwasher, you are not transmitting information about energy; you are consuming energy. The floor cannot be engineered away, because you are selling the floor. Metering persists because the underlying physics requires it.

Cloud compute has been sitting between these two cases for nearly two decades without resolving. Amazon Web Services has been in the market since 2006, has faced intense competition for over a decade, and still bills by the compute unit. The thermodynamic floor for running a server is real and non-trivial, as anyone who has attempted to mine Bitcoin on AWS will confirm: the electricity cost exceeds the mining revenue, which means the marginal cost of cloud compute is not negligible. General cloud computing has not broadly followed the consumer communications path toward flat-rate pricing.

AI inference runs on that infrastructure. It inherits the cloud compute floor before adding its own inference costs on top.

Current inference runs many orders of magnitude above Landauer’s limit.² The headroom is vast in principle. Whether hardware and architectural improvements can close that distance is what the pricing question ultimately depends on. Transformer-based architectures have known efficiency constraints, and the appetite for longer context windows and deeper reasoning keeps pushing per-task costs upward even as per-token costs fall. The physics permits a much more efficient future; it does not guarantee one.

The Netflix Model

There is another escape hatch: hide the meter from the user.

Netflix runs its backend almost entirely on AWS, paying per compute unit for storage, databases, and application logic. It charges consumers a flat monthly subscription. The metering does not disappear; it gets absorbed into the product margin and becomes invisible to the end user. Netflix also built Open Connect, its own global content delivery network, specifically to reduce dependence on metered third-party infrastructure. Vertical integration as cost containment.

This two-layer structure, metered infrastructure underneath and subscription product on top, is already appearing in AI. Consumer-facing products from the major AI providers charge flat monthly rates. The per-token anxiety that shapes API usage is largely invisible to subscribers who use those products every day. The decoupling has already begun at the consumer layer. What has not decoupled is the developer and enterprise tier, where organizations are still making discrete cost calculations about individual interactions.

Three Scenarios

Three scenarios follow from this analysis, each with different governance implications.

In the first scenario, inference costs follow the communications trajectory. Marginal cost falls far enough that it becomes economically invisible against fixed infrastructure, and flat-rate pricing takes over at all layers. Token anxiety disappears even at the API level. At that point, the policy problem is access: who reaches AI services, the same way broadband equity became the central internet policy question. Known policy tools apply.

In the second scenario, inference costs stabilize at the cloud compute level: real and non-trivial, but low and predictable. The consumer layer stays on subscription pricing; the developer tier migrates from per-token to aggregate billing. The friction of thinking about individual inferences disappears, but the meter does not. Governance looks like a hybrid of utility regulation and access policy.

In the third scenario, the current generation of AI infrastructure gets disrupted before it commoditizes, the way mainframe time-sharing was disrupted. Local inference, models small enough to run on consumer hardware without cloud dependency, renders the per-token cloud model obsolete rather than affordable. The governance questions shift entirely: hardware access, local compute policy, and a much more decentralized AI landscape.

All three scenarios are improvements over the current model, in which individuals and organizations make discrete cost calculations about individual interactions. That kind of transactional friction shapes behavior in ways that have nothing to do with the value of the interaction. It favors short prompts over thorough ones, discourages iteration, and creates artificial scarcity in a technology whose value proposition depends on fluency and depth of use.

The Prodigy subscribers who batched their messages to save minutes were not behaving irrationally. They were rational actors under early-stage cost constraints. The constraint shaped their behavior; removing it changed their behavior. The same dynamic is already visible at the consumer layer of AI, where flat-rate subscribers use the technology differently than developers watching a token counter.

The physics will settle which scenario arrives at the infrastructure layer. We do not know the answer yet. What we can say with confidence is that the current per-token model is a way station, not a destination, and governance frameworks built on the assumption that it is permanent are going to need revision.