A full rack is not a business model.

A useful answer creates value. A paid hour captures some of it. The difference is where the business lives—or fails.

The answer has to be worth more than the work behind it.

An AI that saves someone an afternoon can create real value. But that does not tell you whether the company running its chips makes money. Customers must pay for enough capacity, at a high enough price, to cover the equipment, operation, and cost of capital.

  1. 01 · Useful task“This saves me work.”Benefit after checking and integration
  2. 02 · Paid capacity“I will pay for it.”Revenue someone can actually collect
  3. 03 · A viable business“We can do it again.”Costs recovered; the next investment justified

Each step is a condition, not an automatic conversion. One useful answer is not one GPU-hour: models, batching, and tasks consume different amounts of compute.

Can this box pay its way?

Illustrative dollars, not market prices

Start with eight GPU slots, a $240,000 installed purchase, and a four-year working life. “Box” is our teaching unit, not an industry-standard machine. Change only what you can sell.

Sell 70% of the hours at $2.50 per GPU-hour.

The business test

This price pays for this plan.

$1.28 collected for every $1 required.

Surplus after the required return $88,965
Money coming in$406,199
Money required$317,234
Both bars use year-0 purchasing-power dollars over 4 years, discounted at 8% real per year.

70% sold means 49,056 paid GPU-hours a year. At $2.50, year-one receipts are $122,640 before costs.

Buy the first box: $240,000. Run it: $77,234. Buy replacements within this horizon: $0. All in the same present-value dollars.

This horizon pays for the original purchase and operation only. No replacement is purchased at the endpoint; resale value is zero. Recovery is not a claim that the next box is already paid for.

Try “Busy, but cheap.” More sold hours can still lose money. A lower price may help customers; it does not automatically grow demand in this experiment.

Boundary: fixed capacity, before-tax cash flows, no debt or resale proceeds. Energy load stays fixed when billable hours change. Customer benefits are not measured by these bars. Change the deeper assumptions below.

Follow the cash through time
Net present value (NPV)$88,965
Break-even starting $/GPU-hour$1.95
Year 1 operating cash$99,321

This 4-year scenario recovers its cash costs and the 8% real hurdle.

8 illustrative GPU slots; $240,000 paid at year 0. 70% of capacity is billable, starting at $2.50 per GPU-hour.

Cumulative discounted cash, including initial purchase and replacements Cumulative discounted cash starts at -$240,000 and ends at $88,965 after 4 operating years. Squares mark replacement purchases. The full annual table follows. -$300K -$200K -$100K $0 $100K 01234
Solid line: cumulative discounted cash. Dashed line: zero, meaning cost recovery at the selected return. Squares: replacement purchases. Horizontal axis: years. A first crossing of zero is not permanent payback.

No replacement is needed within this horizon. No purchase at the end of year 4; terminal resale value is zero, even for a partly used final generation.

60% physical activity gives 57,185 facility kWh/year. Annual cash costs: $23,319, including $5,719 electricity.

Every year: the cash-flow table
4-year unlevered cash flow · constant purchasing-power USD · costs shown positive; cash flows signed
YearPrice / GPU-hourBillable GPU-hoursRevenueElectricityOther cash costsCapital purchasesNet cashDiscounted cashCumulative discounted cash
0 · buy—0$0$0$0$240,000-$240,000-$240,000-$240,000
1$2.5049,056$122,640$5,719$17,600$0$99,321$91,964-$148,036
2$2.5049,056$122,640$5,719$17,600$0$99,321$85,152-$62,884
3$2.5049,056$122,640$5,719$17,600$0$99,321$78,845$15,961
4$2.5049,056$122,640$5,719$17,600$0$99,321$73,004$88,965

Rounded for display only. The initial outlay is at year 0; all later cash flows are at year-end.

All assumptions & how the numbers work

A slot includes its proportional host/network equipment. Other annual cash costs include allocated rent, staff, maintenance, and networking service; electricity is separate. Space is rented, not purchased. No taxes, debt, working-capital timing, or construction are modeled.

Starting price / GPU-hour
2.5 USD
Paid capacity
70 %
Installed cost / GPU slot
30,000 USD
Operating horizon
4 years
GPU-equivalent slots
8 slots
Economic life / refresh
4 years
Real discount rate
8 % / year
Real price decline
0 % / year
Physical activity
60 %
Idle IT power / slot
0.2 kW
Full-load IT power / slot
1 kW
Facility overhead (PUE)
1.2 × IT energy
Electricity price
0.1 USD / kWh
Other cash costs / slot
2,200 USD / year

Paid hours are not energy load

Reserved capacity can be billed while idle; an unpaid experiment can consume power. Physical activity interpolates between idle and full-load IT power. PUE multiplies that energy to include facility overhead such as cooling.

Annual billable hours = slots × 8,760 × paid fraction
Revenue(t) = billable hours × starting price × (1 − price decline)t − 1
Facility kWh = slots × [idle kW + (full kW − idle kW) × activity fraction] × 8,760 × PUE

Recovery, replacement, and the time value of money

Everything is in constant purchasing-power dollars. The real discount rate expresses the required return above inflation; it is not a nominal rate. Both bars are discounted to the same date, so receipts and costs are comparable. Their difference is the project’s unlevered net present value.

Year 0 pays the installed cost. At the end of each completed economic life, the next generation is purchased only if another operating year follows. Replacements buy the same capacity at the same real cost and energy profile. Terminal resale value is zero, even for partly used equipment. The four-year default recovers the first purchase; it does not book the next purchase at year 4.

NPV = −initial purchase + Σ [revenue(t) − operating cash costs(t) − replacement(t)] / (1 + real rate)t

The sum runs through the selected horizon. The break-even starting price sets NPV to zero along the same price-decline path. At zero paid hours, no finite price can recover costs. Depreciation is an accounting allocation, not an additional cash payment; subtracting it here would count capital twice. Interest and principal are excluded, not forgotten: this is a project return before financing.

Economic life is an assumption, not an accounting schedule. There is no automatic performance improvement, demand growth, replacement downtime, or optimal shutdown. Costs and physical load are fixed; the box keeps operating even when loss-making. Those simplifications make this a sensitivity experiment, not a company valuation.

Follow the invoice

The cloud still has a landlord. And a power bill.

A neocloud is a specialized compute-cloud provider here, often focused on GPUs. It sells access to expensive machines. A model lab may buy that capacity rather than build everything itself; a hyperscaler offers a broader cloud platform. These roles can overlap.

One illustrative financing chain
  1. Investors & lendersSupply equity and debt
  2. Compute providerBuys equipment; pays for a place to run it
  3. Cloud customerPays for capacity; still needs useful work to sell

Customer payments run back through operating bills and debt service. Equity gets what remains.

A conceptual cash flow, not a particular deal’s legal payment priority.

A long customer commitment can help finance a fleet. Nebius’s Microsoft agreement describes expected funding from contract cash flows and debt secured against the contract. But a contract is not cash already collected, and financing is not revenue.

One large customer can disappear or renegotiate. New chips can weaken old chips’ resale value. Debt payments can arrive before the hoped-for business does. Borrowing rearranges the risk; it does not manufacture demand.

Who actually owns the compute? It is not a two-company club.

Compute ownership is not limited to two companies. Microsoft and Amazon report equipment purchases as well as leases. CoreWeave describes buying hardware. That does not mean every GPU is owned outright or every occupied building belongs to its operator.

Hardware, buildings, power, and money are separate layers. A provider can own servers in a leased facility. A cloud-capacity purchase buys a service, not necessarily a lease of identified servers. A power-purchase agreement is not ownership of GPUs or unlimited access to the grid.

Pledging its GPUs gives a lender security rights; it does not by itself erase the borrower’s ownership. CoreWeave’s financing announcement distinguishes debt from equity. Neither should be counted as customer revenue.

xAI’s Colossus account documents infrastructure development and operation, not unencumbered title to each asset. Microsoft and OpenAI’s October 2025 agreement illustrates capacity purchases alongside an equity relationship. Buying shares in a model lab is not paying its compute invoice.

Dated 2023–2026 disclosures, not an ownership census. Private terms leave gaps; accounting assets are not GPU-only market values.

The next twenty years are a relay, not one very long GPU rental.

The equipment will be replaced; the useful tasks must keep arriving. Falling prices can spread benefits to customers while squeezing suppliers. More demand helps only if the resulting receipts cover the new cost structure. The lab’s longer horizons make those repeat purchases explicit.

At the customer’s end, value is not “tokens generated.” It is better work after integration, review, errors, and organizational change. Time saved matters when it improves outcomes or frees resources for something worthwhile. Those complementary changes take time—and may fail to pay off.

At the economy’s end, adding provider revenue to tenant revenue double-counts an intermediate input. A large compute bill is a cost to one business and revenue to another; it is not automatically new prosperity. Durable gains require useful additional output, not merely more invoices.

AI can be useful and some AI investments can be bad. Both can be true.

Three accounting traps—and what the evidence can really tell us

Depreciation is not a replacement fund. It allocates a past purchase across accounting periods. Adding it back does not pay for the next machine. Our cash model counts purchases once and does not subtract depreciation again. Debt service is separate from the unlevered project return. GDP is gross of depreciation; net domestic product deducts it.

A sunk purchase is not a reason to run forever. Once a box is bought, compare future avoidable costs, receipts, and alternatives such as resale. For the next box, justify the full investment again. The simple lab does not optimize shutdown or resale.

A task gain is not a GDP forecast. A 2023 NBER working paper studied 5,179 support agents and reported 14% more issues resolved per hour on average, with heterogeneous effects. Those numbers are from the working paper, not its revised 2025 journal version—and not a 14% economy-wide forecast.

Acemoglu’s May 2024 task-based scenario estimated no more than 0.66% cumulative total-factor-productivity improvement over ten years, refined below 0.53% for harder-to-learn tasks. That is not annual GDP growth or a universal ceiling on what AI could eventually do. Neither study calibrates our illustrative box.

Sources, dates & methods

Original evidence review: . Second edition: . Individual source and review dates are listed below; none is a scenario forecast date.

The lab is a sensitivity experiment, not a valuation, market-price survey, or macroeconomic forecast. Corporate sources support specific asset and contract distinctions. Financial commitments are not treated as realized cash or output. The source ledger retains the original evidence dates; the second edition changes the teaching sequence, not those documents.

  1. Microsoft Annual Report 2026

    Microsoft · Fiscal year ended June 30, 2026

    Notes 6 and 13 distinguish property and equipment, including servers/network equipment/software, from operating and finance leases. These are aggregate accounting categories, not a GPU inventory, market valuation, or proof that every asset is unencumbered.

  2. Amazon.com Announces Fourth Quarter Results

    Amazon · February 6, 2025 · results for 2024

    The cash-flow statement separately reports purchases of property and equipment and equipment acquired under finance leases; the release describes AWS server infrastructure. The financial totals cover Amazon, not exclusively GPUs or AWS. They cannot establish title to an individual accelerator.

  3. CoreWeave Secures $2.3 Billion Debt Financing Facility

    Michael Intrator · CoreWeave · August 3, 2023

    The article explicitly describes purchasing and paying for hardware for executed customer contracts, as well as buying the company’s first GPU. This supports hardware acquisition, not an inference of ownership from collateral alone. It does not establish ownership of every data-center building.

  4. CoreWeave Secures $7.5 Billion Debt Financing Facility

    CoreWeave · May 2024

    Announces a signed debt facility to develop compute for existing customer contracts, alongside separately described equity funding. Facility size is not revenue, profit, money necessarily drawn, or proof that all underlying assets have identical collateral terms.

  5. Nebius announces agreement with Microsoft for AI infrastructure

    Nebius · September 8, 2025

    Dedicated capacity at Vineland, New Jersey; expected financing from contract cash flows and debt secured against the contract. The release describes an intended financing strategy and identifies deployment, financing, and pricing risks. It is not evidence that every planned dollar has been spent.

  6. Colossus: the build and timeline

    xAI · Undated page · timeline May 2024–February 2025

    xAI describes building the cluster and reaching 200,000 GPUs in February 2025. This is evidence of a frontier lab directly developing and operating infrastructure, not audited proof of legal title to every GPU, building, or power asset. The page’s scale and performance claims are the company’s.

  7. The next chapter of the Microsoft–OpenAI partnership

    Microsoft and OpenAI · October 28, 2025

    Separately describes Microsoft’s equity investment and OpenAI’s commitment to purchase additional Azure services. A services purchase is not hardware ownership, and the investment is not the compute invoice. This dated agreement is not a complete or current inventory of OpenAI’s arrangements.

  8. Generative AI at Work · NBER Working Paper 31161

    Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond · April 2023 · revised November 2023

    This working-paper version studies 5,179 customer-support agents and reports 14% more issues resolved per hour on average, 34% for novice/low-skilled workers, with minimal impact on experienced/highly skilled workers. These are not macroeconomic estimates. The final 2025 QJE article (doi:10.1093/qje/qjae044) is a different version; the figures here belong to the specified working paper.

  9. Artificial Intelligence and the Modern Productivity Paradox · NBER Working Paper 24001

    Erik Brynjolfsson, Daniel Rock and Chad Syverson · November 2017

    Discusses implementation lags, complementary innovations, organizational changes, skills, and intangible capital, alongside other explanations for weak measured productivity. This is a framework for interpretation, not a guarantee that delayed returns will arrive.

  10. The Simple Macroeconomics of AI · NBER Working Paper 32487

    Daron Acemoglu · May 2024

    A task-based scenario gives no more than 0.66% cumulative total-factor-productivity growth over ten years, refined to less than 0.53% under harder-to-learn-task assumptions. These are conditional estimates, not annual GDP growth, a universal upper bound, or inputs to this unit-economics lab.

An independently built learning series. Models are teaching tools, not research findings or investment advice. No prompts or lab inputs are sent to an AI service.