Rounds to Zero. The Fluid Economics Underpinning the Price of Compute.

In June of 2009, a Senate subcommittee sat down to find out why a text message cost twenty cents.1

The price had doubled from ten cents at all four major American carriers, at more or less the same time over a three year period.2 Srinivasan Keshav, a computer scientist at Waterloo, was called in to testify on what it actually cost a carrier. His analysis revealed that it cost about a tenth of a cent for the wireless channel and roughly double that for the billing.3 Against twenty cents retail, three-tenths of a cent is a 98 percent margin on the marginal message.4

Any other paper on the cost of AI would use this as a story about the perceived inanity of vendor pricing. Prices detach from costs when costs aren’t transparent (though competition usually closes the gap). Today’s token prices have been inflated the same way, so the conventional wisdom is to wait, negotiate hard, and let the market do its work.

But the obvious lesson is wrong because competition never closed that gap. The price per message never actually came down. The Senate hearing changed nothing.5 What happened instead was that the meter became irrelevant. First came the bundles, then unlimited, then the messaging apps.6 By about 2013 very few people were still counting their messages.

During the whole time that people were arguing about twenty cents, there was a different number that was actually driving behavior. If you are from a certain generation, you will recall that a text message was 160 characters. That number had been produced by a standards committee way back in 1985. And while it wasn’t a technical limit, it completely shaped how many of us wrote for the next twenty years.7

You may have noticed that there’s a new meter in town, watching tokens. Each week you can read posts that exude the same confidence and share roughly the same blind spot. Most organizations have adopted the same setup: a cost-per-token figure (even though very few can say what it should be measured against, or where the figure comes from).

I.Unmetered

Consider for a moment that it is rare that anyone calculates the cost to produce a slide deck.8 Because it isn’t easily measurable, it isn’t managed.9 The same goes for most of what knowledge workers do between meetings: drafting memos, first-pass contract reviews, summaries of summaries, and the constant rebuilding of decks to account for changes in numbers, approach or strategy.

Now LLMs have arrived to do some of that work, and unlike the unmeasurable knowledge worker, the cost is clearly visible. We haven’t started spending money on making documents, we’ve just started seeing what it costs.10

Measurement drives management regardless of magnitude.11 Or, in other words, cost that can be measured gets managed, and those that can’t get absorbed (irrespective of which of the two is larger). Companies therefore are running optimization programs against this small, new, and (most importantly) visible number, while the large one stays off the books.

II.Invisible Handiwork

There are two primary mechanisms that drive the unit price of technology. One is a consequence of production, and the other is a consequence of strategy.

The first is pass-through: production gets cheaper, competitors race each other to the customer, and retail follows the cost curve down. This is the story most commonly associated with the semiconductor. It’s the one that taught us to expect “it gets cheaper every year” is a fundamental law of the universe.12 Under pass-through we can forecast because the curve has physical determinants.

The second is abandonment: this is when price stops mattering to the seller, and the meter gets switched off. That’s how long distance13 and cost-per-text ended. The business model changes rather than the underlying cost. Under abandonment it is nearly impossible to forecast because the timing is a factor of strategy.

Practically every business case written since 2023 assumes the first mechanism. But neither mechanism describes what’s happened to inference this past year. Contract rates for the H10014 went from about $1.70 to $2.35 per GPU-hour between October 2025 and March 2026.15 Last generation’s silicon got a lot more expensive, which could be read as an ordinary market shortage. The key question is: what happens when the shortage clears? Under pass-through the price should fall back toward cost (which is the expectation underpinning most of the public business cases). Whether it holds this time depends on who owns the hardware, and on what terms.

III.New Class Unlocked

On August 10th, Nvidia announced memorandums of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR, to build financing platforms aimed at mobilizing more than $500 billion of third-party capital against AI compute.16

Jensen Huang describes Nvidia compute as “broadly adopted, flexible across models and workloads, fungible and transferable across customers and operators, and continuously improved through CUDA software,” (the software component being the piece that extends its useful life17). Larry Fink told CNBC it put him in mind of the mortgage-backed securities market he came up in during the 1970s.18

Underneath the vocabulary it’s a familiar instrument. Chips go into a special-purpose vehicle which then borrows against them and the notes get sold on.19 Nvidia says it may provide residual value support on up to twenty-five percent of an opportunity, project by project, if resale value comes in under projection when the financing matures.20

What that does to the price of inference isn’t complicated (or sinister). Once a large share of the world’s compute is bought with borrowed money, the price of compute becomes an input to a credit model rather than an output of a cost curve. Credit models need utilization at certain levels and revenue per gigawatt clearing certain thresholds, and the debt behind the hardware is denominated in dollars rather than tokens. What they don’t need is for the price to reflect the actual cost to run a model.

Unfinanced capacity can be sold for whatever the market is willing to bear. Financed capacity mandates a payment schedule, driven by a large and well-capitalized constituency whose interest is in holding the price up—the debt floor. It doesn’t drive prices down the way pass-through would, nor is it the seller walking away from the meter. This third driver is inverted relative to the other two. How much debt it carries depends on a key assumption about the working life of the chips.

IV.Different Clocks

Ask how long a graphics processor lasts and you’ll get three answers spanning a factor of four from the experts. Michael Burry’s position has generally been read as two to three years. IBM’s Arvind Krishna has said five. Huang points at the A100, a 2020 part still in commercial service six years on, and extrapolates that to an economic life stretching toward a decade.21

At the end of the day, these are accounting judgments, not measurements of a physical object. They’re attempting to judge obsolescence, which depends entirely on what comes next. That judgment also doesn’t stay in the accounts. Depreciation schedules are used to set borrowing bases which determine how much hardware gets bought. The hardware sets the supply of inference capacity. And the supply of inference capacity sets what a thousand tokens cost. In other words, AI budgets are being run off an accounting policy debate with unclear inputs. None of which gives a clear indication of where the money covering the losses comes from.

V.41 Minus 59

In early August, Apollo’s chief economist Torsten Slok published a note showing that margins across the AI stack rise the further you sit from the end user. Silicon and equipment run at plus 41 percent. Models and applications run at minus 59.22

That’s backwards from most industries where the business closest to the customer captures the most value. Here the layer selling to the end customer is the one losing money. In Slok’s summary: “The upstream margins are real, but they are paid for out of capital raised by the layer losing money, not out of cash generated by end demand.”

That ends up shaping the underlying return question. Boards typically want to know whether a particular AI investment paid off, rather than asking if the arrangement is making any money at all. The price is set by companies running at deep negative margins, funded by investors betting the price goes up or the cost comes down. Some part of what every buyer pays can be seen as a customer acquisition expense shifted onto a different balance sheet. And critically, all of this assumes the hardware could be sold on to somebody else.

VI.Children of the Boxcar

On the last day of May in 1886, crews along roughly 11,500 miles of Southern railroad pulled the spikes on one rail, shifted it three inches, and drove them back in. Within just three days, standard-gauge trains had begun running through the South.23

They did it so a car built for one track could earn money on any other track. That fungibility is what made the equipment trust certificate work. Rolling stock was bought through a trust that held title until the certificates were retired, so a lender looking at a locomotive was looking at something it could take away and sell to any of a hundred different operators.24 The instrument worked because a boxcar in receivership was still a functioning boxcar.

The same logic is being applied to this latest revolution as lenders are looking for collateral they can model and re-deploy, which means financing conditions are expected to push data center operators toward standardized installation design. Data center builds will be laid out for what a lender can resell, not necessarily tailored for the workloads they’ll be running.

Where the analogy fails is on the resale itself. Between 1893 and 1897, companies owning about a third of American railroad mileage passed through bankruptcy.25 And yet even as these companies went belly-up, the track still carried freight. The receivers ran the roads, the rates came down, and the capacity that bondholders had paid for got used by shippers. The only loser in the whole narrative was the bondholders.

Through it all, there were buyers for the equipment, because the operators who hadn’t failed still needed cars. In the compute story, who will the buyer be? The hyperscalers are building their own facilities and designing their own silicon. The neoclouds are levered. In a default large enough to put a real tranche of collateral on the market, the natural purchasers are the firms that are placing the same bets, meaning there might not be a bid. For a buyer, all of this arrives as a modeling problem.

VII.Assumption Debt

Software engineering has a name for the accumulated cost of shortcuts and deferred maintenance: technical debt. It compounds until the system becomes unmaintainable. Something similar is accumulating in how we model our AI spending.

Call it assumption debt: the gap between what our plans are modeled on and what is actually being controlled for. Namely unit prices flat or falling, capacity available on demand, the vendor still offering the service at a predictable price in the future, and finally the model version still existing in the same or compatible form that prompts were tuned against. None of those is unreasonable, but the ownership is distributed.

So a company that approves a three-year program in early 2025, conditioned on unit prices falling twenty percent a year, will be caught off guard when last generation’s silicon gets more expensive. That program is currently running against a curve trending the opposite way from the one the approval was conditioned on. No one is paying attention because they’re tracking spend, not exposure.

VIII.Skeptic’s Check

There is another, less skeptical reading of the industry, and that is that demand for inference may simply be inelastic and growing fast enough that the financing structure doesn’t actually matter that much. If every enterprise application acquires a model call in the next three years, the constraint will be capacity, not price. Lenders financing capacity into a shortage tend to do well. On that reading the debt floor never binds, because demand holds the price well above it.

Perhaps Huang’s estimates about a useful life are correct, and the A100 will continue working with regular software updates to the installed base to keep the edge. If history has taught us anything, it’s that we repeatedly underestimate how long hardware stays economically productive once the workloads adapt to it.

The mortgage-backed comparison might similarly fall apart. At its core, a mortgage-backed security was a claim on households’ willingness to keep paying for an asset that produced no cash. A chip in a data center has the potential to produce revenue every hour it runs. That’s a materially better collateral story, and the people arranging these deals know more about collateral than most of the less-informed commentary does.26

And abandonment could arrive here too, in our favor. If inference becomes a feature of something else (a la the way texting became a feature of a messaging app) the metered price stops being a metric any of us tracks. Several vendors have already begun pricing by seat or outcome instead of by the token.

Deliberate loss-making is a strategy, not a symptom. Amazon ran at a loss for the better part of a decade (and the people who understood why did very well). A layer losing money on purpose while it takes a market is a different animal from a layer losing money because it can’t charge enough, even though from the outside the two are nearly indistinguishable.

IX.No Matter What

A few things survive regardless of which reading turns out to be right. The baseline problem holds in every scenario. If we never measured what the displaced work cost, no future price makes the return calculation valid.27 Cheaper tokens don’t help an unanchored comparison.

The exposure problem also holds in every scenario. Third-party assumptions operate on a different incentive model from the ones we own, whether they turn out right or wrong. The debt floor remains on an external balance sheet, managed on an independent timetable. The portability problem holds as well: if a vendor doubles the price, how many weeks until companies are substituting away? If a model version is retired, how long? Those numbers are knowable today even if they aren’t being measured. Until someone opts to run the workload against a second provider, the answer is, at best, a guess.

The only thing that doesn’t hold across all scenarios is the confidence. A three-year model with a smooth declining cost curve is attempting to make a claim about the outcome of a financing experiment running at a scale of several hundred billion dollars without any better information than the rest of us.

X.Unknown Unknowns

We don’t know whether the debt floor holds. We don’t know whether the compute financing structure is the beginning of a durable asset class or a way of moving risk to holders who won’t examine it closely.28 We don’t know whether the depreciation schedules being written into these deals will look conservative or absurd in 2030. We don’t know whether inference demand grows into the capacity being built, or the capacity sits waiting for demand that arrives late. The latter was the failure of the dot-com period. The tech evangelists were right about what the technology could do, even though they were wrong about the timing.

What we can say is narrower. The price of compute is an output of a capital structure now, not a straightforward function of what compute costs to produce. The return our boardrooms are asking for is calculated against a baseline nobody measured, and an externally calculated rate. And the load-bearing assumptions are being considered as givens rather than hypotheses.

Which leaves the question of what the modern “160 character limit” is this time. The closest match is the rack standard the lenders are working toward. It is being settled now, by credit committees solving for resale value, without tuning toward actual workload demand. Everything built afterward will have to fit inside it anyway.

The depreciation schedule is more consequential, though with a looser parallel (hence the loud, very public debates). But it was meant to answer a question about accounting rather than anything about a company’s AI program, and it sets what those tokens cost three years out regardless. Neither decision will show up on a dashboard or get negotiated with the people who have to build within those constraints.

Twenty cents was never the number that mattered. Neither, in the end, is the price of a token.