Investment returns

The equipment can pay back before the whole investment does

A disclosed IREN–Microsoft case now gives the recovery test a clearer capital boundary. The original Childress plan budgeted $5.8bn for equipment and ancillaries plus $2.8–3.2bn for 200 MW IT of facilities. Its approximately 85% project EBITDA estimate includes direct power, repairs, maintenance and site costs, but is not complete owner free cash flow. Budgets and margins remain company estimates. Original cost perimeter and source notes.

The model synchronizes four phases into 60 equal service months, places capital and the contractual advance at the modeled start, and credits that advance against months 25–60. It conserves the $9.664199424bn tranche-fee total. Actual payment dates, calendar-hour billing, taxes and as-built costs are not observed; this is not a realized project IRR.

The 80% and 75% cases add five or ten percentage points of costs outside the project measure, applied to original fees. Those are sensitivities, not measured company margins. Routine power and maintenance are already included and are not deducted twice. Net end value means sale proceeds after transition/realization costs, or the equivalent value of continuing operations, before allocating between creditors and equity. The model does not sell the assets and then reuse them.

The result does not say the site is worthless or already impaired. It says that a continuing asset must do additional economic work before the whole initial investment clears the stipulated hurdle. The equipment-only cash payback in month 31 omits the facility; it is also not the company’s approximately two-year claim for a different set of newer three-year contracts. Complete assumptions, sensitivity figure and reproducible calculation →

The earlier general recovery illustration remains a different test

The following $14.117bn example explains how life, timing and cash margins affect recovery. It remains useful, but it is not the newly documented IREN cost perimeter or an actual CoreWeave cohort valuation.

A gross margin deducts some direct costs. Operating profit includes more expenses and accounting depreciation. Operating cash flow adjusts noncash items and working capital but does not deduct all investment. None alone tells us whether a particular investment has recovered its capital and earned a return for the risk taken.

The whole-system assessment uses $14.117bn—CoreWeave’s disclosed first-half cash-investment scale—to illustrate an investment cohort. It does not assert that all of that mixed spending entered service together or that it describes a new GPU-only fleet. The public evidence does not supply the cohort-specific prices, utilization, maintenance and asset mix needed to underwrite the actual investment. Disclosed investment input.

At four operating years and an assumed 50% complete pre-financing cash margin, the required level annual revenue is $8.907bn. Approximately $4.454bn of annual cash then has a present value equal to the $14.117bn investment at a 10% hurdle. This is a conditional break-even calculation including return on capital, not simply cost divided by years.

The margin means cash after operating costs, tax and maintenance, before financing payments. It is not the reported EBITDA margin. Interest is not deducted a second time when the discount rate already represents the all-capital hurdle. New investment needed after the chosen earning life is not supplied free by the model.

Price, cost, delay and useful life decide the outcome

These are different scenarios, not a measured revenue shortfall. Delay also can create construction overruns, financing needs or loss of competitive earning time that this simple calculation does not include. A longer building life cannot be assumed to extend an accelerator’s useful earning life. Assumptions, complete table and reproducible equations.

The demand study adds a cost-sensitive test. With revenue normalized to 100, assumed variable cash costs of 35 and fixed costs of 15, cash is 50. A 20% price cut and 25% volume increase restore revenue to 100, but variable costs rise to 43.75 and cash falls to 41.25. At the inherited four-year investment scale, the scenario has a −$2.470bn net present value. It is not an actual project loss.

A 20% fall in unit variable costs restores that cash. Without the cost improvement, the assumed structure needs 44.4% more volume—not 25%—to restore original cash. Extra work must fit available capacity; expansion capital is not included. The stress is not calibrated to a mix-sensitive token-price series. See the revenue-versus-cash visual and complete sensitivity table.

Cheaper service can help buyers without rescuing every supplier

Lower unit costs can make more useful tasks worth buying. Competition can also pass savings to customers, and customers can change their model mix or supplier without stopping AI use. The crucial conjunction is adequate paid volume, attainable prices, controlled all-in costs and sufficient duration. More use by itself is not enough.

The office buyer example shows that a moderate license cost can be justified by a fraction of saved-time value. The support example shows that a billable outcome need not eliminate an equivalent human expense. Good customer economics support willingness to pay; they do not allocate every dollar of customer surplus to the model or infrastructure provider. Office buyer hurdle; support buyer hurdle.

Power and delivery matter to that duration. The IEA’s April 2026 outlook estimated global data-center electricity use at 485 TWh in 2025 and projected 950 TWh in 2030. A forecast of energy demand is not delivered capacity or cash available for a particular project. Grid and construction constraints can restrain excess capacity while also delaying receipts. IEA scope and forecast.

Retail pricing does not automatically reprice a fixed cloud contract. A price decline may first reduce the contracted buyer’s margin rather than the infrastructure lender’s current receipts. The ability and obligation to keep paying—and the provider’s delivery performance—then matter. CoreWeave also reported higher annualized capacity pricing on short three-to-six-month contracts; that issuer claim cannot be extended across an entire fleet or asset life. Dated short-contract statement; contrasting contract horizons.

Creditors can be paid while the original investment disappoints

A negative investment net present value is not a dollar estimate of lender loss. An all-capital recovery model asks whether cash compensates for the full investment and required return. A debt-payment test asks whether the particular borrower pays the particular obligations on their dates. The financing structure decides which claim absorbs the difference.

DDTL 4.0 makes this distinction concrete. DBRS identifies Meta’s take-or-pay contract and expects it to fund full amortization without refinancing. Meta may honor reserved-capacity payments even with disappointing utilization; the operator may pay lenders but retain a thin equity residual. A project rating therefore does not certify Meta’s AI investment return, the operator’s equity return, or an attractive share price. The rating is also a projection, not current compliance.

Sources: Morningstar DBRS initial credit analysis; DDTL 4.0 executed credit agreement.

Liquidation is a backup case, not that expected repayment route. Against $2.837bn of June principal and approximately $3.3bn of noncurrent book collateral, assumed 5% realization costs imply required gross realizations of about 85.6% of book if all $155m of current collateral is recovered as cash, or 90.5% with none. Those cash outcomes and costs are assumptions, with prior-ranking and competing swap claims set to zero. They are not expected losses or market prices.

The threshold changes with principal amortization, drawings, depreciation, transfer costs, reserves consumed and claim priority. It does not value collateral at final maturity and must not be substituted for the $14.117bn hypothetical cohort used above. The collection, reserve and enforcement calculations keep those two models separate.

Time, replacement and a usable second life decide the remaining recovery

The service contract preserves aggregate fees by extending service end dates after a delayed start. The timing test therefore shifts all 60 service months: at the $8.8bn midpoint and 85% project measure, a twelve-month delay raises required net value from $2.806bn at year five to $3.912bn at year six. If the advance is delayed too, the requirement rises to $4.224bn. These are assumed timing cases before extra construction, idle-equipment, penalty or financing costs—not findings of a contractual breach. The extended-term timeline.

Another route is to keep the facility and buy a new equipment generation. The finite second-cycle case charges $5.8bn for replacement equipment plus $0.2bn for retrofit at year five, then five years of operations with no final sale value or new customer advance. Its $20m per IT MW-year price scale comes from a dated claim about different, newer contracts; using it five years later is explicitly an assumption, not a forecast.

With assumed full-use annual variable costs of $1bn, fixed costs of $0.2bn and an 80%-cash first term, 90% paid utilization at that price produces +$0.299bn whole ten-year NPV. At 65% paid utilization and a 20% lower price it produces −$2.825bn, despite still generating $1.23bn of annual operating cash. Both use a 10% hurdle.

At the selected unchanged price/cost scale the whole investment breaks even at 85.95% paid utilization. With a 20% price reduction and unchanged unit costs it would require 117.21%, beyond the fixed cohort’s capacity. Efficiency, cheaper replacement or a better service mix can change that boundary; unbounded extra volume cannot be assumed. Follow the replacement arithmetic and limits →

Facility-only continuation is a different business. The required fifteen-year net annual facility cash is about $353m in the 85% first-term case or $431m in the 80% case, with no extra initial tail capital. Those are reverse tests, not observed rents. Core Scientific’s gross billing-MW revenue cannot be used as a matched net IT-MW rent for Childress. Why the hosting comparison cannot supply an appraisal.

An ownership price asks a different cash question

A share buyer does not buy one year’s profit. The price reflects a claim on future distributions and realizable ownership value, affected by reinvestment, financing and dilution. High current earnings refute a claim that nothing commercial is happening; they do not justify an arbitrary price.

The whole-system assessment’s Microsoft exercise uses a timestamped 17 September 2026 observation, 19:29:50 UTC, at $496.03 and about $3.692trn equity value. That observation lacked a durable vendor page and a same-instant share-count reconciliation. It is retained only as an illustrative input, not presented here as a verified current quotation or a fair-value estimate. Price provenance and full model.

Its annual equity-cash proxy is $63.886bn: operating cash less cash property/equipment purchases and finance-lease principal. Allowing for the full $12.405bn stock-compensation expense lowers the proxy to $51.481bn. That allowance is a rough treatment of dilution, not measured repurchase cash. Microsoft cash and compensation disclosures.

At the middle assumptions, the first path reaches roughly $448bn annual cash in year ten, and about 69% of present value comes from later cash. Terminal assumptions therefore matter greatly. The exercise does not precisely model nonoperating assets, future borrowing, acquisitions, taxes or dilution; recognizing some omitted assets consistently could lower the operating-growth requirement. Cash growth is not necessarily revenue growth. The model neither supplies a trading target nor establishes impossibility.

A financing-round value is not money already earned

The dated March OpenAI post-money value of $852bn and May Anthropic value of $965bn price ownership at those rounds. They are not the cash raised, enterprise values or audited sales. The May round includes $15bn of earlier hyperscaler commitments, so those commitments cannot be added again. Later reported revenue run rates do not silently replace a dated denominator. OpenAI round; Anthropic round.

One explicit scenario assumes no owner distributions for ten years, followed by cash growing at 3% indefinitely, discounted at 10%. It needs $154.7bn of OpenAI owner cash in year eleven, or $175.2bn for Anthropic. At an assumed 25% complete owner-cash margin, those correspond to $618.8bn and $700.8bn of revenue. The margin and timing are not observed lab economics. Private sensitivity and equations.

Earlier distributions would lower the endpoint requirement. Additional cash needs or new equity can raise an existing owner’s hurdle. Favorable reporting that Anthropic expected a second consecutive adjusted operating-profit quarter, excluding stock compensation, is relevant counterevidence to a perpetual-loss story. The full investor documents were unavailable and the article was paywalled; adjusted operating profit is not audited distributable cash after all reinvestment. Attributed profitability report.

An offering changes the financing—not the asset’s economics

The broader offering review retains its 18 September evidence boundary and did not establish publicly priced OpenAI or Anthropic offering terms. Company-announced confidential submissions, public exhibits and reported preparation dates are not completed financing. The selected CoreWeave update is narrower: already a public issuer, it priced $3.7bn of additional convertible notes on 18 September, with settlement expected 22 September. It is neither an IPO nor received cash at the financing investigation’s 20 September cutoff. Pricing source; dated stages and limits.

A listing can broaden funding access and reveal financial information. It cannot make an uneconomic project earn its cost simply by changing owners. A postponed offering is also not automatically insolvency: the question is available cash and obligations without assuming proceeds from an unclosed transaction.

If that issue closes without the purchaser option, estimated net proceeds less capped-call costs leave $3.1457bn before other expenses, while $3.7bn face principal remains. A nominal 2.875% coupon does not express the entire economic cost of conversion and dilution protection. Prospective proceeds calculation.

What would make the return case stronger?

The decisive evidence is realized cash from identifiable delivered investments: service prices and paid utilization, complete operating costs, maintenance and replacement needs, productive life and actual amortization. A backlog helps only to the extent delivery occurs and the counterparty can pay. Investor pricing then needs a separate test against that cash and dilution.

Customer demand makes that outcome plausible in important settings. It has not established the assumed 50% cash margin or a four- to six-year cash stream for the investment cohort. Evidence that would strengthen or weaken the answer →

The next useful recovery evidence is net transferable service and equipment value: power and lease rights, assignment, downtime, replacement cost and the time needed to redeploy. Better evidence could support the favorable amortization case or reveal a weak enforcement backup. A chosen recovery percentage alone is not a finding. What remains unmeasured.

The new case narrows the missing inputs: final Horizon costs and dated payments; acceptance, credits and collections for all phases; complete cash operating costs; net values of transferable sites or clusters; and actual replacement-generation economics. None of those quantities is established by a reported short-term capacity price. Cohort-specific revision tests.