Capital systems / AI & investment

Who gets paid
for the AI boom?

The supplier can collect before the buyer earns a dollar.

AI adoption can survive a financing shakeout; the returns will not be shared evenly.

28 October 2026 · About 20 min read

Listen to the summary

A delivery worker moves a server cabinet toward an unfinished facility while an engineer examines an unconnected power cable.

The argument in 30 seconds

AI adoption can continue through a financing shakeout. The key is whether stress stays with weak borrowers or reaches well-funded customers cutting orders because returns disappoint.
The same AI order can pay one company and burden another. Customer quality, contract terms and deployment timing determine who captures the return.
Strong demand is not enough to justify any valuation. Oil, rates and grid politics affect project economics; competition and replacement spending determine what reaches shareholders.

Executive summary

My working view is that AI adoption can continue through a financing shakeout, but the returns will not be shared evenly. The important distinction is between a weak borrower struggling to refinance and a well-funded customer deciding the economics no longer justify spending. Paid adoption supports demand; it does not establish the return on every project. Deployment delays, expensive funding and competition can weaken those returns without disproving the technology. I would therefore favour businesses that turn demand into cash, can meet their obligations and retain an advantage as customers get more choice. Oil, the Fed and disputes over grid costs make that assessment less forgiving. Valuation remains the final test: a strong business can still be a poor investment. I would become more constructive as cash conversion improves, and more concerned if weakness spreads to well-funded buyers cutting orders because returns disappoint.

A chip sale is someone else’s
cash requirement.

The immediate buyer of compute is not necessarily the person funding useful work. A customer pays for an AI service, the service buys compute, and someone has to build the capacity before those receipts arrive. Debt and equity can bridge that gap. The question is whether they are financing a productive asset or postponing a problem.

Paid adoption and cash-generating customers are a real counterweight to the financing concern. The test is whether weakness stays with borrowers who need new capital or spreads to well-funded buyers disappointed by returns.

Figure 1

Two kinds of money move through the same industry

End customerAI serviceCompute operator

Service payments

Capital providersCompute operatorHardware supplier

Compute operatorSite & power

Buildout finance · equipment, rent and power

Hardware supplierAI service

Optional strategic funding, not every project.

Conceptual relationships, not measured revenue shares. One company can perform several roles. Strategic supplier funding is optional, not universal.

1 / 3

Customer cash has to reach the system.

End customers pay an AI service; that service pays a compute operator. Those recurring receipts are different from the debt or equity used to buy equipment before the service earns anything.

Research question: are independent customers paying for useful work, or are orders still dependent on the next funding round?

2 / 3

The buildout is paid for before it earns.

Capital providers fund a compute operator, which buys equipment and commits to sites and power. That financing is not customer revenue. It bridges the gap between delivery and cash generation.

Research question: who bears the cost if capacity cannot open or customers do not use it as expected?

3 / 3

Separate contracts can share one dependency.

A supplier can also invest in a customer or support its financing. That may enable a viable project. It also means apparently separate orders can depend on the same customer’s access to fresh capital.

Research question: how many exposures ultimately rely on the same fundraising or capex decision? Dashed line = optional strategic funding, not every project.

Compute operator

Pays for capacity before it can sell the service. Delivery, utilisation and funding costs decide how much cash remains after keeping the facility running.

Read the relationships

End customers pay for useful work, not installed capacity: renewal depends on time saved, revenue earned or costs reduced. AI services collect that revenue and pay for compute; usage can rise while revenue per task falls, so margin after serving the workload matters.

Capital providers can fund both AI services and compute operators before customer receipts arrive. Strategic supplier support can also fund either business; the schematic omits some of these paths. Operators buy equipment and commit to sites and power. Financing bridges delivery and cash generation; it is not customer revenue. Supplier investment or financing support can enable a viable project, but apparently separate orders may share one customer’s need for fresh capital. The dashed route is optional, not universal.

Hardware suppliers can collect before buyers earn a return, but retain their own costs, commitments and exposure to later orders. Sites and power providers depend on tenant credit, availability obligations and who funds extra infrastructure. Lenders have a contractual claim; equity receives the residual. Their timing, loss absorption and upside differ.

Follow the customer’s money before adding another AI holding. Owning a chip designer, a compute landlord and a cooling supplier can look diversified, yet all three may depend on the same investment budget. They have different business risks, but shared exposure to the decision to build another cluster. Following the money is more revealing than counting the tickers.

These are familiar questions from sovereign credit. October has taken me from Hong Kong to Bangkok around the IMF–World Bank meetings, and now to Sri Lanka, with Milan next. The setting changes, but I keep asking who funds the investment, when it earns cash and what happens if those timelines stop matching. One company’s equipment sale is another company’s capital requirement; the supplier may collect long before the buyer earns anything from using it. ^[1]

Read the contract,
not just the order book.

The practical work is in the contracts. I want to know who owes the money, when they must pay and what lets them walk away. Supplier investment can make a viable project possible. It also deserves scrutiny when several apparently separate orders depend on the same customer raising more capital.

Figure 2

The same demand story. Different funding.

What supports the commitment?

Customer wants capacityYes

Cash sourceOperating business

Financing dependencyLower

Good project returns?Still to be proved

Not a company ranking

Cash generation buys more options.

A customer generating cash can absorb a delay without immediately returning to investors. That strengthens its ability to honour a commitment, but it does not prove the investment earns an attractive return.

Who owes the money?

Identify the legal payer, any guarantee and how much capacity depends on the same counterparty. Brand recognition is not the same as recourse to a strong balance sheet.

When is payment earned?

Look for installation, power, service-availability and acceptance conditions. An order can be real without being collectible today.

What can change or be cancelled?

Check volume flexibility, break clauses and price resets. A large commitment is not automatically an unconditional obligation.

A funding window can close before demand does. A customer may genuinely want more compute while relying on debt or equity that has not yet been raised. When funding tightens, an order can remain economically desirable but harder to complete. Contract protections and available liquidity become decisive.

A stronger customer and a better investment are related questions, not interchangeable ones.

Move the revenue.
The bills don’t wait.

A deployment delay can change the return on a project without changing anyone’s belief in AI. Equipment has been purchased, financing costs are running and the power connection is late. Revenue moves out while newer hardware keeps arriving. The question becomes how much useful earning time is left, who covers the extra cash requirement and whether the next project still clears the investment hurdle.

The exercise below isolates those moving parts. Change utilisation, borrowing costs or the delay before a hypothetical facility starts earning. The horizon remains five years from delivery, so waiting uses up earning time rather than moving the finish line.

Capacity actually earning revenue80%

30%100%

All-in annual borrowing cost8%

3%18%

Delay beyond planned start0m

0 months24 months

Change selling price as well

Compute selling price, relative to the base case100

60 · lower prices125 · pricing strength

This changes revenue per unit of utilised capacity. It does not automatically change demand, chip efficiency or power costs.

Earning window

60m

Out of 60 months from delivery.

Extra sponsor cash

$0m

To fill interim cash deficits.

Equity cash returned

$77m

1.54× the $50m contributed.

In the baseline, the facility earns for all five years and repays its debt. The sponsor receives $77m on $50m contributed. This is a cash multiple, not an annual return.

At an illustrative 12% annual equity hurdle, the base-case NPV is −$6.3m. A positive cash multiple does not, by itself, clear the required return.

Cash before extra funding, asset sale and debt repayment

Your scenarioBaseline

Negative values identify the funding gap, not a usable negative cash balance. The model then assumes the sponsor fills that gap. The terminal debt payment is shown separately below.

The baseline case is shown: utilisation 80%, borrowing cost 8%, no delay. The full model is defined below.

Find the combinations that return less than the equity invested

Each cell runs the same model at a different utilisation and borrowing cost. Your current delay and selling price are held constant. Select a cell to test it.

Cell values are total equity cash multiples, not annual returns. Below 1× means less cash returned than contributed; it is not a default probability. This grid samples the model, not real companies.

Compare the five project scenarios

Scenario	Use / rate / delay / price	Equity paid in	Equity returned	Cash multiple	NPV at 12%
Base case	80% / 8% / 0m / 100	$50.0m	$77.0m	1.54×	$-6.3m
12-month delay	80% / 8% / 12m / 100	$50.0m	$44.8m	0.90×	$-24.6m
Underused	50% / 8% / 0m / 100	$50.0m	$11.0m	0.22×	$-43.8m
Costly debt	80% / 13% / 0m / 100	$50.0m	$62.0m	1.24×	$-14.8m
Combined stress	45% / 13% / 12m / 85	$52.8m	$0.0m	0.00×	$-52.5m

USD millions; price index 100 is the baseline. All scenarios end at month 60 with a $15m asset sale. Interim gaps receive assumed sponsor support. These are model cases, not forecasts or default probabilities. Under combined stress the lender also loses $44.12m of principal. The fixed 12% equity hurdle is illustrative.

Inspect the model and its limitations

This is an added teaching model, not a company forecast from the report. All assumptions below are illustrative. No market price, utilisation estimate or company debt term is being implied.

At delivery	$100m spent: $65m equipment, $35m construction. $60m debt and $50m sponsor cash fund the spend plus a $10m reserve.
Operating economics	Full-use annual revenue $56m at price index 100; variable operating cost $12m at full use. Site rent $3m annually throughout. Other costs $2m annually before opening, $3m after; maintenance $2m annually after opening.
Funding	Interest on $60m is paid monthly from delivery, at the chosen fixed all-in rate. Principal is due at month 60. No scheduled amortisation.
End of year five	Assets sell for $15m. Cash pays debt principal first; the residual goes to equity. The terminal date and sale value do not move with a delay.
Interim shortfalls	The sponsor is assumed to add equity whenever cash would fall below zero. This is not assured in reality. No early default is modelled.

Monthly net cash = revenue − variable costs − rent − other costs − maintenance − interest.
Revenue begins after the delay; interest and rent begin immediately.
Equity cash multiple = cash paid to equity at month 60 ÷ all sponsor cash contributed.

Excluded: taxes, working-capital lags, a ramp to steady utilisation, variable-rate resets within the horizon, refinancing fees, covenants, early default, supplier and landlord operating costs, future capex and dilution terms. The residual asset value is fixed, not an estimate of GPU resale prices. Cash is retained rather than distributed before month 60. Equity NPV uses a hypothetical 12% effective annual required return: terminal equity cash is discounted from month 60, while initial and any additional sponsor contributions are discounted from their actual model dates. The 12% hurdle is an assumption, not a market estimate. More complete underwriting could materially change the result.

Timing relationships are illustrative, not to scale. Contract terms and interest capitalisation differ across projects. This model pays interest monthly from delivery; it does not capitalise interest into the asset or debt balance.

See who gets paid ↓

Monthly values in a table

Use the assumptions above to reproduce the baseline. The monthly table is populated when scripting is available.

One project.
Four different payoffs.

The project’s weakness does not hit everyone at the same moment. In this model the hardware is paid for upfront, rent and interest are honoured during the buildout, and equity gets what remains at the end. That explains why a supplier can report a strong period while the customer absorbs a deteriorating return. Future orders can still suffer; they are outside this single-project exercise.

Figure 3

Who receives the cash in your scenario?

Baseline · 80% utilisation · 8% debt cost · no delay

Hardware supplier

$65m

Equipment cash received at delivery.

Property owner

$15m

Contracted rent received across five years.

Lenders

$84m

Interest plus repayment of principal.

Equity sponsor

$77m

Residual cash returned after debt repayment.

These are different types of receipts, not profit margins or shares of one revenue total. Lender receipts include return of principal; equity receipts include return of invested capital. Suppliers and landlords have costs of their own. All contractual interest and rent is assumed paid, with extra sponsor money where needed.

At month 60: $122m cash, plus $15m asset sale, less $60m principal repaid, leaves $77m for equity.

Interim sponsor top-ups, where required, are included in the cash balance and in total equity contributed. Recoveries are simplified; legal seniority and security packages are not modelled.

The outcome is not equally bad for everyone. Capacity already operating can become more valuable while unfinished sites struggle. Microsoft also reported cutting the time from GPU delivery to operation by almost half in its largest regions over the fiscal year. Better execution is a genuine way out of the problem; we should watch for it rather than assume delays are permanent. ^[2]

The headline reaches the trade
through a mechanism.

Original research snapshot · 28 October 2026, 11:32 HKT · before that day’s Fed decision. Retained analysis, not refreshed market data.

Rates and the inflation constraint

The financing backdrop has become less forgiving. The Fed raised its target range to 3.75–4.00% on September 16. September CPI then rose 0.6% on the month and 3.7% over the year, with gasoline accounting for more than half the monthly increase. The October 28 Fed decision is still ahead at this note’s research cut-off. I would not build an infrastructure thesis around imminent rate relief when the inflation data leave that assumption open to challenge. ^[7]^[8]^[9]

Iran and the oil channel

Iran and the wider regional conflict matter through that inflation channel, not just through the price of diesel for generators. On October 6, the EIA raised its fourth-quarter Brent forecast to $105 a barrel, $14 above its previous forecast. On October 28, Reuters reported Brent around $115 after the Houthis claimed an attack on a Saudi refinery; Saudi confirmation was still absent from that report. The forecast and the intraday price are different observations, but both make an easy-energy assumption harder to defend. ^[10]^[11]

It would be too crude to translate higher oil directly into higher costs for every American data centre. Petroleum supplied less than 1% of US electricity in 2025. The broader risk is pressure on household spending, transport costs and inflation expectations, potentially keeping financing expensive. An AI project can face a higher required return even if its own electricity contract has not changed. ^[12]

The same pressure reaches Asia

Growth and renewed financial pressure can coexist. Writing from Sri Lanka, that distinction is worth keeping in view: the IMF’s September assessment put second-quarter growth at 4.2%, while reporting August inflation of 8% following the global oil shock. Hong Kong offers another reminder. The HKMA raised its Base Rate to 4.25% after the Fed’s September move, although its HIBOR-based calculation was materially lower. Neither a strong growth story nor the currency peg makes local funding conditions uniform. ^[13]^[14]

The midterms and who pays for the grid

The US midterms are also becoming a debate about who pays for the buildout. On September 30, a Senate procedural vote on the Ratepayer Protection Act fell short at 57–43. The proposal concerned recovery of the incremental grid-upgrade costs serving large users. Ahead of the November 3 election, the investment issue is not simply which party likes AI. It is whether power tariffs, permits and cost allocation change the economics or timing of a particular site. ^[15]^[16]

Making developers cover more infrastructure costs could reduce project returns and still improve the chances of getting projects approved. A smaller return on a project that gets built may be preferable to a larger projected return on one that stays blocked. The cost and the approval benefit need to be assessed together.

A friendlier Fed would not automatically solve a borrower’s funding problem. Treasury yields and credit spreads are different parts of the bill. A rate cut offers limited relief if lenders simultaneously demand a much bigger premium to finance the company. The customer can be right about demand and still struggle with the cost of reaching it.

Illustrative scenarios · not market data

Now trace those risks through a project.

The controls below illustrate the mechanisms. They do not estimate the market impact of the historical events above.

Reference rateDown 1 percentage point

↓

Borrower credit spreadUp 2 percentage points

↓

All-in funding costStill up 1 percentage point

Illustrative changes, not market observations.

Watch the transmission

A friendlier Fed can meet a less friendly lender.

A lower reference rate need not lower the company’s total funding cost. If lenders demand a bigger credit premium, the borrower can still face a more expensive bill.

The model uses one all-in rate, held constant through the scenario. It does not simulate a forecast path for the Fed.

Back to the evidence above ↑

Oil is not a proxy for every electricity bill. The route can run through household budgets, inflation expectations and required returns. This is a conditional channel, not an estimated oil-to-yield relationship; the project model has no oil-price variable.

The approval timetable belongs in the financial model. Grid-cost allocation and approvals affect costs and the chance of building at all. Paying more may reduce the projected return and improve delivery prospects. The delay control changes revenue timing only; it excludes extra construction costs and approval probabilities.

The earnings doubled.
The share price didn’t.

Price is the final test, even when the business delivers. As a simple illustration, earnings per share can double over five years while the multiple falls from 40 times to 20 times, leaving the share price unchanged before dividends. That is not a forecast for any company. It is a reminder to distinguish business growth from the return earned by the buyer of the shares.

Figure 4

What did the entry price already assume?

Earnings per share after five years2.0×

Price / earnings paid at entry40×

Price / earnings at the end20×

Earnings per share double, but a multiple falling from 40× to 20× leaves the share price unchanged.

Illustrative arithmetic, separate from the project model. Share-price factor = earnings-per-share factor × exit P/E ÷ entry P/E. Dividends, fees, taxes and dilution are excluded.

Efficiency matters on both sides of the trade. More useful AI can generate more paid tasks, while better hardware and software reduce the compute needed for each one. The hardware outlook depends on which grows faster. Cheaper inference may expand the market and improve customers’ returns, rather than destroy the investment case. We still need to identify which suppliers benefit from that expansion.

And more usage does not always mean more hardware

Usage and efficiency push in different directions. Double the number of tasks while halving compute per task and total compute is unchanged. That arithmetic does not tell us how much demand lower costs will create, but it shows why adoption alone is not enough for the hardware forecast.

Number of paid tasks2.0×

Compute needed per task50%

Total compute: 1.0× the starting level.

Illustrative product of tasks × compute per task. This does not include changing workloads, service quality, capacity utilisation or hardware mix.

Put the number
back in context.

Different disclosures answer different questions. Paid seats tell us something about adoption, backlog about commitments, and cash generation about financing capacity. The company cases below retain the original reasoning, including evidence that cuts against a simple bullish or bearish reading.

Original company research · 28 October 2026 cut-off. Different reporting periods; not a valuation ranking or a live data feed.

Microsoft FY2026 Q4 · July disclosure in the original note

Paid adoption and financing capacity

30m+ paid seats

Microsoft provides a useful counterweight to the idea that all this spending rests on speculative demand. In July it reported more than 30 million paid Microsoft 365 Copilot seats and $19.6 billion of quarterly free cash flow. Those are different measures: the first supports paid adoption; the second shows the wider company’s financing capacity. Neither proves that every AI project earns an attractive return, but a customer generating cash has more options than one waiting for the next funding round. ^[2]

Source [2] ↗

CoreWeave Q2 2026 · September and October observations

Demand can strengthen while funding gets harder

~$104bn backlog

CoreWeave shows why the distinction matters. Its second-quarter results included roughly $104 billion of revenue backlog and $640 million of net interest expense against $2.575 billion of revenue. The backlog is subject to delivery and service-availability requirements. A large order book is valuable, but interest does not wait for every contracted cluster to become productive. ^[3]

There is evidence on both sides. CoreWeave said in September that it was signing short-dated compute contracts at higher prices. By October 15, Reuters was reporting yields around 16% on its 2031 bonds. These are dated observations, not current quotes. Still, they illustrate the question well: customers can want more compute while creditors become less comfortable financing the route to it. ^[4]^[5]

Source [3] ↗ Source [4] ↗ Source [5] ↗

Nvidia FY2027 Q2 · July-quarter filing

Supply secured ahead of customer cash

$279bn commitments

Nvidia’s supply commitments show how far ahead the industry is underwriting demand. Its July-quarter filing showed supply-and-capacity commitments rising from $119 billion to $279 billion, primarily for memory and manufacturing facilities. Some supplier agreements can be adjusted before firm orders are placed, so treating the entire sum as irrevocable debt would be wrong. The scale still makes the allocation of risk across those contracts important. ^[6]

Source [6] ↗

Broadcom FY2026 Q3 · original reporting period

Growth can reach more than one architecture

$16.7bn AI semiconductor revenue

The semiconductor opportunity is broadening, but that is not the same as every supplier keeping its economics. Broadcom reported $16.7 billion of AI semiconductor revenue in its fiscal third quarter, up 221% year on year. That demonstrates another supplier capturing substantial spending; it does not by itself prove Nvidia is losing share. Custom silicon, networking and accelerators can grow together while competition changes the returns on the next dollar invested. ^[17]

Source [17] ↗

SK hynix · 28 October 2026 disclosure in the original note

Visibility is not a fixed margin

>50% of 2027–28 capacity

Memory gives us a more immediate test of how much protection a contract actually provides. On October 28, SK hynix said long-term agreements covered more than half its 2027–28 production capacity, with prices responding to market conditions within agreed ranges. That improves visibility, but it is not a promise of permanently fixed margins. The company also plans significantly higher investment next year. I would examine the contract protection and supply response before assuming today’s scarcity earns tomorrow’s multiple. ^[18]

Source [18] ↗

Keep the evidence separate from the simulator. These are the original note’s company disclosures and interpretations, not inputs to the hypothetical model. They have not been independently re-verified or refreshed in this editorial update.

The useful question is
what would change the view.

I would become more constructive if paid demand stays healthy, delayed capacity starts generating cash and funding needs stop rising faster than the business can support. Stronger providers could take customers from weaker ones, operating sites could earn better returns and lower inference costs could bring in new users. A financing shakeout does not have to end the adoption cycle.

I would worry more if well-funded customers begin cutting orders because utilisation, pricing or customer returns disappoint. That is harder to dismiss as a temporary credit problem. The warning would be weakness spreading across customers, not one refinancing headline or a bad week in semiconductor shares. I would want evidence before calling the selling mainly forced, or the bad news already priced.

Is this a funding problem or a demand problem?

More constructive

Paying customers stay, delayed capacity earns cash and financing stress remains concentrated among weaker providers.

More concerning

Well-funded customers cancel or cut orders because utilisation, pricing or customer returns disappoint.

Who is actually left with the cash?

Evidence to inspect

Collections, obligations, replacement spending and debt maturities—not just revenue growth or contracted capacity.

Avoid assuming

That an equipment supplier, a landlord, a lender and a shareholder all experience the same boom in the same way.

How much success did we pay for?

Work to do

Rebuild the valuation using sustainable margins, reinvestment and a credible financing plan. Leave room for a less friendly outcome.

What is not enough

A strong quarter, an attractive narrative or a fall in the share price without a view on what earnings and cash flow can support.

My preference is for businesses with paying customers, credible funding and an advantage that survives less favourable conditions. I would then ask what the valuation requires me to believe about margins, replacement spending and dilution. The same order can help one company report a strong quarter while leaving another with years of obligations. That is where I would spend the research time.

We can be right about the technology and still pay the wrong price for the wrong part of it.

Who gets paid for the AI boom?

Executive summary

A chip sale is someone else’s cash requirement.

Two kinds of money move through the same industry

Customer cash has to reach the system.

The buildout is paid for before it earns.

Separate contracts can share one dependency.

Customer cash has to reach the system.

Compute operator

Read the contract, not just the order book.

The same demand story. Different funding.

Cash generation buys more options.

Move the revenue. The bills don’t wait.

Build once. Earn over time.

One project. Four different payoffs.

Who receives the cash in your scenario?

The headline reaches the trade through a mechanism.

Rates and the inflation constraint

Iran and the oil channel

The same pressure reaches Asia

The midterms and who pays for the grid

Now trace those risks through a project.

A friendlier Fed can meet a less friendly lender.

The earnings doubled. The share price didn’t.

What did the entry price already assume?

Put the number back in context.

Paid adoption and financing capacity

Demand can strengthen while funding gets harder

Supply secured ahead of customer cash

Growth can reach more than one architecture

Visibility is not a fixed margin

The useful question is what would change the view.

Is this a funding problem or a demand problem?

Who is actually left with the cash?

How much success did we pay for?

Source notes

Who gets paid
for the AI boom?

A chip sale is someone else’s
cash requirement.

Read the contract,
not just the order book.

Move the revenue.
The bills don’t wait.

One project.
Four different payoffs.

The headline reaches the trade
through a mechanism.

The earnings doubled.
The share price didn’t.

Put the number
back in context.

The useful question is
what would change the view.