Who pays for the empty capacity?

Picture a single invoice with two worksheets clipped behind it. The invoice covers one month of a shared technology platform: the reserved servers, the support shift that keeps them running and the software licences on top, $120,000 in all. The platform can practically handle about 12,000 comparable units of work a month, and this month 6,000 arrived. The first worksheet divides the bill by the work that came in and charges each unit $20. The second divides the bill by what the platform could have done, sets a rate of $10 a unit, assigns $60,000 to the work performed and leaves the other $60,000 on the page under its own name: unused. The platform, the invoice and every number on it are invented, the units are assumed to be alike and the capacity figure is an assumption rather than a measurement, so that the arithmetic stays in plain view. Anyone who has read an internal chargeback report will still recognize the two sheets.

Both worksheets reconcile to the same $120,000. They differ in what they divide by. The first spreads capacity supplied, the commitment the organization paid for whether or not anyone used it, across capacity consumed, the work that happened to arrive. The second keeps the two apart. My argument is that a cost record which fuses them will make a shared service look less efficient whenever demand falls, though no server, shift or supplier invoice has changed. Separate the cost of the capacity supplied from the capacity consumed, and then decide, on purpose, who owns the part that went unused.

This is not a claim that the second worksheet saves money. Moving $60,000 from the customers' charges to a line called unused capacity changes no cash flow at all; the bill is paid either way. Nor is it a claim that $10 is what the next unit of work really costs, or that every shared platform needs an elaborate costing system. The claim is narrower. The first worksheet answers a bookkeeping question, how to make the month balance, and in answering it hides the question that has a decision behind it: why the organization is paying for twice the capacity it used.

The same bill, a different denominator

Press the button under the figure, and one business unit that sends 2,000 units a month moves its work elsewhere. The invoice does not move, because the reserved servers, the shift and the licences were bought for the term. On the first worksheet, the remaining 4,000 units now carry $30 each, a 50 percent increase for customers whose own work did not change at all. Nothing about that work became more expensive to perform. There are simply fewer customers left to share a fixed commitment.

The second worksheet records the same event in a different place. The rate stays at $10, because the capacity and its cost did not change. The charge for work performed falls to $40,000, and the unused line grows to $80,000. Each remaining customer sees a stable price, and the organization sees a growing number it now has to explain. That presentation is not kinder to anyone. It is more informative, because it puts the change where it happened, in the relation between what was bought and what was used, instead of spreading it over the people who stayed.

Rate the capacity, not the month

The second worksheet is a simplified version of an idea set out by Robert Kaplan of Harvard Business School and Steven Anderson, founder of the software and consulting firm Acorn Systems. Their working paper, Time-Driven Activity-Based Costing, is dated November 2003; a version appeared in the Harvard Business Review in November 2004, and a book of the same name followed, which the school described as new in April 2007.¹²⁴ Kaplan had helped develop activity-based costing in the 1980s, and the new method was partly a repair of it.

The original method asked employees, by interview or survey, how they divided their time among activities, and used the answers to spread a department's cost. Kaplan and Anderson found that this was time consuming and costly to run, and that the models were often not maintained afterward.¹ In the 2007 interview Kaplan called the time surveys "the Achilles heel of ABC."⁴ The paper also names a quieter flaw, the one that matters here. People asked what share of their time goes to each activity report percentages that add up to 100, and few report much idle time. The survey therefore assumes, without saying so, that the department was fully used, and every dollar of its cost lands on the work that happened to be done.¹

The time-driven version replaces the surveys with two estimates: the cost per unit of time of supplying capacity, and the time each transaction or activity consumes.¹³ The first is the department's cost divided by its practical capacity, the time its people or machines can actually work, which is deliberately less than the theoretical maximum. The paper suggests that practical capacity is often estimated at 80 or 85 percent of the theoretical figure, allowing for breaks, arrival and departure and communication unrelated to the work, and for machines, maintenance and repair.¹ It states plainly that rates should be calculated at practical capacity, not at actual utilization.¹ With those two numbers, each transaction is charged for the capacity it used, and the cost of capacity paid for but not used appears separately, where managers can review it and consider reducing the supply of resources, or decide to keep the capacity for expected growth.¹

That is the whole of the second worksheet. The $10 rate is the bill divided by practical capacity; the $60,000 is the rate times the work done; the unused line is what remains. Kaplan said in the same interview that the insight about the central role of capacity came to him only around 1990, years after the first activity-based systems.⁴ It is easy to see why it was late. A system built to explain where cost goes is naturally built to make all of it go somewhere.

How the first worksheet feeds itself

Run the first worksheet forward. The $30 charge is a signal, and signals change behavior. A business unit facing a 50 percent price rise for unchanged work will reasonably ask whether it can buy the same service more cheaply elsewhere, and sometimes it can, because an outside supplier with other customers is not asking it to fund this platform's idle commitment. If a second unit leaves, the same bill is spread over fewer units again, and the rate rises again. Whether such a loop actually runs depends on how easily customers can move and how long the commitments last, so treat it as a possible feedback loop, not a forecast about every shared platform. But it needs nothing more than arithmetic and a customer acting sensibly on the price in front of it.

The ruler below holds the bill and the practical capacity fixed and lets the volume that actually arrives move from 3,000 to 12,000 units. Drag across the sheet, or tab to it and use the arrow keys.

The ink curve is the first worksheet: the bill divided by the volume. It climbs slowly at first and then steeply. Halving the volume from 12,000 to 6,000 doubles the charge, and halving it again to 3,000 doubles it again. The gold line is the second worksheet's rate, flat at $10 wherever the volume sits. The two meet at one point only, practical capacity, where nothing is left unused. Everywhere else the hatched gap between them is unused capacity, divided among the customers who stayed and presented to them as the cost of their own work.

A rate is not the price of the next request

The second worksheet needs its own warning label. A $10 rate says what capacity cost, measured against what the platform could do. It does not say what the next request will cost. Within slack capacity, an extra unit of work is nearly free: the servers are paid for and idle. At a bottleneck the same unit can be expensive, because it delays other work or forces overtime. At the edge of a contract tier, one more unit can be very expensive, because it triggers the next block of reserved capacity or the next licence band. The question behind a real decision, such as whether to take on a new workload, is what changes in total cost and value if you do. An allocation rate, however carefully built, answers a different question.

Time is also a better driver for some resources than for others. Kaplan and Anderson built their method around people and processes, where minutes are a natural measure of what a transaction uses, and extended it to machine time.¹ A database query's cost may depend instead on memory, disk reads, how many other queries run at once, or a reservation paid in advance. For a technology platform, the honest version of the second worksheet measures consumption in whatever unit actually limits the pool, and says which one it chose.

Practical capacity, finally, is an estimate, not a reading from a meter. Set it high and the unused line looks large; set it low and the line shrinks toward nothing. A manager who would rather not explain idle servers has an obvious reason to lower the estimate, and a manager who wants a bigger budget next year has a reason to raise it. A more detailed allocation is not automatically a more truthful one. The test of a better estimate is which decision it changes: whether a commitment is renewed, resized or filled with other work.

Give the residue an owner

If the unused $60,000 is not the customers' problem, whose is it? The useful answer names a decision rather than a department, and there are four to choose from. Keep the capacity, because the slack buys something the organization wants, such as room for a seasonal peak, a recovery target or a service level; then write that purpose down and have someone with budget authority sign it, so it can be tested later. Redeploy it, by bringing other work onto the platform, which shrinks the unused line without changing the bill. Let it expire, by not renewing the commitment at the end of its term. Or remove it now, if the contract allows, and accept the cost of leaving early.

Only the last two save cash, and only when a commitment actually ends. Until then, moving cost from a customer's charge to an unused line changes who sees the number, not how much is paid. That distinction ran through an earlier dispatch on the time that AI saves: released capacity is an opportunity, and it becomes money only when someone acts on it. Servers and shifts work the same way. What the second worksheet offers is a better place to make the choice, with the number in front of the person who can change it.

What the choice should not be is invisible. When the first worksheet balances, the unused capacity has not gone anywhere. It has been folded into each customer's charges and from there into their profitability reports, where nobody who reads them can do anything about it. A product line that looks marginal may be carrying a commitment it never chose. The spreadsheet needed every dollar to land somewhere, and it put the dollar where it was least likely to be questioned.

Objections, considered

Someone has to pay the whole bill. Yes, and both worksheets agree on that: every dollar is reconciled in each. The disagreement is about which line the dollar sits on, and therefore which decision maker sees it. A central budget that carries the unused line is still paying for it. It is also the budget that can renew, resize or cancel the commitment, which a business unit's profitability report cannot.

Slack is waste. Some of it is. Some of it is a purchased service level: spare capacity is what lets a platform absorb a surge, or lose a machine, without slowing down for everyone. The difference is whether anyone chose it. Name the slack, state its purpose and review it on a schedule, so that capacity kept for resilience is defended as resilience instead of being discovered, years later, as waste.

The model is too simple for cloud pricing. It is, on purpose. The example is a fixed pool of identical units so that the arithmetic stays visible. Real platforms mix reserved commitments, which behave like the invoice here, with usage charges that rise and fall with demand, and with step costs at tier boundaries. Each deserves its own line. Usage charges can be passed through as consumed; reserved commitments are capacity supplied and need the treatment described above; the steps need a decision at each boundary. Folding all three into one blended rate brings back the first worksheet's problem, with more decimal places.

This teaches departments to treat technology as free. It is a real risk. If business units see only the $10 rate and never the unused line, they may use capacity as if it were unlimited, and a platform team left holding the residue may keep buying more. Price signals matter, and that is exactly the point: a chargeback policy is an incentive design, not a neutral statement of cost, and it should be designed for the behavior the organization wants. A rate based on practical capacity tells customers what their use costs. A visible capacity budget tells the owner what it costs to keep the rest. Both signals are needed, aimed at different people.

An unused commitment needs an owner before it needs a customer to absorb it.

Before the next chargeback review

• Which denominator does the rate use? The volume that actually arrived this month, or the practical capacity of what was bought.
• Who estimated practical capacity, and how? An estimate with no named author and no method can be moved to make any month look good.
• Which costs change with the next unit? Separate usage charges from fixed commitments and from the steps where one more unit buys a new tier.
• Who approved the unused capacity? A name, a purpose and a review date, or only a residue in someone else's charges.
• When does each commitment end? That is the date on which keeping or cutting the capacity becomes a choice that saves cash.

Back to the invoice with its two worksheets. Press the button again and the worksheets trade places on every line that depends on volume: $20 becomes $30, the used line shrinks and the unused line grows, while the rate holds at $10. The invoice does not change. It said $120,000 before the business unit left, and it says $120,000 after. Everything that moved was a choice about where to write the number down. The only figure that can actually get smaller is the total at the top, and it shrinks only when the person who owns the commitment decides that it should.