AI Can Work and Its Investors Can Still Lose

Your AI investment thesis can be right and your shares can still disappoint you. Customers adopt the product. The machines do useful work. Eventually the business earns serious money. Unfortunately, “eventually” occurs after the financing meeting at which someone else receives a large part of the company.

The original investors can still attend the celebration. Depending on the recapitalization, they may be attending as guests.

An October 3, 2026 Reuters analysis by Stephen Eisenhammer examines the gap between AI's potential and the commercial returns needed to fund its infrastructure. It reports economist Stijn Van Nieuwerburgh's estimate of roughly $9 trillion in US investment over 2025 to 2032, with about $3.55 trillion in annual sector revenue needed by 2032 to earn a modeled 10% return.^R1

Those are estimates under assumptions, not a bill that becomes payable if the industry misses a universal revenue target. They establish the size of the commercial task. An individual company's survival depends on its own contracts and cash.

For a founder, the revealing question is who can fund the wait. A company with room to delay a project has a different set of choices from one whose next payment requires a new investor. The second company can remain technologically impressive while becoming financially negotiable.

This article uses evidence available through October 3, 2026. Dollar amounts are US dollars. The financing example is hypothetical; the bankruptcy discussion concerns US companies.

We like to explain investment returns by pointing at how useful a technology will become. It is an agreeable pitch because nobody has to discuss what we paid. A sufficiently large total addressable market can keep that part of the meeting off the agenda for years.

There is evidence that AI can improve real work. In Generative AI at Work, Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that access to AI assistance increased issues resolved per hour by 15% on average in their studied customer-support setting, with substantial variation across workers. That is evidence of productivity in a particular deployment, rather than an economy-wide growth forecast.^A1

But the benefit can land in several places. A customer may keep the saving. Competition may force the application provider to lower its price. The cloud operator may capture a margin. A chip supplier may earn a return before the downstream business does. A shareholder owns a claim on one of those businesses, at a particular purchase price, with particular claims ahead of it.

Counting every layer's revenue as a separate pool of final customer spending would exaggerate the money entering the system. The application company's cloud bill is revenue to the cloud provider, but it is also a cost the application company must recover.

Cheaper capability makes the distinction sharper. Stanford's 2025 AI Index reported that the inference price for performance at a GPT-3.5-level benchmark fell more than 280-fold between November 2022 and October 2024. That measures a historical price change at a defined performance threshold. It does not say every AI service gets cheaper at that rate, or that any particular GPU has become worthless.^S1

The investor consequence is conditional. Lower prices can unlock enough demand to increase total profits. They can also benefit users while squeezing providers whose equipment and financing were priced for fatter margins. The relevant variables are paid usage and contribution per unit after delivering the service, including the capital needed to keep delivering it.

If your bull case celebrates falling customer prices while holding your own selling price constant, the spreadsheet has appointed you the only capitalist exempt from competition.

Founders buying AI services may welcome the price decline. Founders financing infrastructure need to ask whether efficiency gains belong to them long enough to repay their investment. Both can be enthusiastic about AI and arrive at different answers about the same asset.

A board can spend an entire meeting admiring revenue growth without reaching the cash available for creditors. This usually produces excellent minutes. The lender remains unaccountably interested in the bank account.

CoreWeave's results show why the distinctions deserve attention. For the quarter ended June 30, 2026, it reported $2.575 billion in revenue, $1.510 billion in adjusted EBITDA and $640 million in net interest expense. It also reported approximately $104 billion of revenue backlog, subject to delivery and service-availability requirements.^C1

The three figures below are different accounting measures for the same quarter. Net interest expense is not identical to cash interest paid. Adjusted EBITDA excludes costs that still matter to shareholders; subtracting interest from it does not produce free cash flow. Strong reported demand can coexist with a substantial financing burden.

A backlog is valuable evidence when enforceable contracts connect future capacity to customers willing and able to pay. Its usefulness for financing depends on the route from that commitment to collected money. Check acceptance milestones, cancellation provisions, customer credit, delivery costs and payment dates. A signed contract that requires you to finish an expensive facility first may improve your financing prospects while increasing the amount you must finance.

For a project, build a cash schedule that deducts operating costs and working-capital needs, then places required investment and debt payments on their actual dates. Include lease payments once, in the appropriate category. Separate cash that is restricted to one project from cash the parent can move elsewhere. Reconcile the result to the company's cash-flow statement rather than inventing a preferred version of earnings.

A long-term investment can have attractive expected economics and still encounter a cash shortfall next quarter. Refinancing can bridge that mismatch when lenders remain willing and the terms leave enough value for equity. The risk is having to negotiate while your alternatives are deteriorating.

The distinction also works in the company's favour. Heavy expenditure today may build capacity that produces cash later. A period of negative free cash flow does not establish that the investment destroys value. The underwriting job is to test the later cash flows and finance the intervening payments.

“Fully funded” deserves a date and a list of conditions. Without them, it can mean everyone was available for the announcement photograph.

Here is a deliberately simplified infrastructure project, not a forecast for CoreWeave or any other company. It costs $100 million, financed with $60 million of debt and $40 million of equity. Assume the existing share register gives founders 30% and early investors 70%, with no options or other securities.

The founders have promised themselves they will tolerate short-term volatility. They have not yet established whether the company can afford the same personality trait.

For one modeled year, the project expects $20 million of operating cash after operating costs, working-capital changes and cash taxes, but before debt service and necessary reinvestment. It owes $12 million of scheduled interest and principal and needs $5 million to keep the operation competitive. That leaves $3 million.

Now delay customer deployment enough to cut operating cash to $10 million. Hold those obligations constant:

The machines still work. Customers may still arrive. The company needs $7 million for this year, plus whatever cushion the rest of its plan requires. Assume it has no spare unrestricted cash or available committed credit and cannot safely cut those payments.

A new investor offers $20 million at a $30 million pre-money equity valuation. The resulting post-money value is $50 million. The investor receives 40%; existing holders retain 60%. The raise covers the modeled $7 million gap and leaves $13 million for later needs. It does not establish that every future shortfall is funded.

Assume all shares have identical economic rights, existing holders do not invest, and there are no fees, preference changes, anti-dilution adjustments or option-pool increases. Founders move from 30% to 18%. Early investors move from 70% to 42%. The newcomer owns the remaining 40%. These are ownership percentages, not a conclusion about board control.

Suppose the business later generates $60 million of distributable equity proceeds after creditors have been paid, with no intervening distributions or further share changes. The original holders collectively receive $36 million, against the original $40 million of equity funding. They lose $4 million in nominal terms despite the productive business. The new investor receives $24 million on its $20 million investment. Whether that is an adequate annualized return depends on how long it waited.

The rescue investor's entry price explains the difference. It purchased exposure after the delay had weakened the old holders' negotiating position. No one had to prove the technology fraudulent. No one even had to abandon the forecast of eventual commercial success.

A shareholder update could truthfully say the long-term vision remains unchanged. Updating the percentages beneath it requires a less inspirational email.

This is a tradeoff, not a morality play. The new investor supplies money that might preserve some value for everyone. If adoption slips again, it can lose money too. Existing holders may rationally accept dilution rather than gamble the whole business on a financing window reopening. What they should avoid is treating the rescue price as irrelevant because the technology still works.

The simple example used identical shares. Actual negotiations can ask for more than a percentage. An investor willing to extend the company's runway may also want priority on repayment, limits on additional borrowing or approval over future transactions. “We believe in the team” is compatible with requiring the team to ask permission before doing anything expensive.

The following are possible negotiated terms, not inevitable outcomes or descriptions of the AI companies discussed here:

A liquidation preference can give new money the first claim on equity proceeds in a sale. That means a percentage-only cap table may overstate what older holders receive in a mediocre exit. Model the distribution under the actual documents, including senior debt and transaction costs. The waterfall is where a large headline valuation becomes individual proceeds.

Governance can shift separately. A board seat or veto over a financing affects the next negotiation even if the investor holds a minority economic stake. Conversely, a large economic interest does not automatically supply every control right. Ask which decisions require consent, when those rights terminate, and whether an emergency financing makes them more consequential.

Bankruptcy is the more severe route. The SEC explains that a US company can reorganize under Chapter 11 while its existing common shares are cancelled. Common holders rank behind creditors for distributions. A successful operating business after reorganization therefore need not deliver a recovery to its former shareholders.^B1

That outcome depends on value, claims and the approved plan. A missed payment does not automatically award the company to one lender, and lending against collateral does not guarantee full recovery.

For employees, the ownership distinction can be especially painful: continued work and a surviving product do not ensure surviving equity value. Management should explain the effect on their securities directly. A triumphant announcement that “the company has emerged stronger” is rather selective if the employees' old shares have emerged nowhere.

“Patient capital” sounds like a temperament. In practice, patience needs funding of its own. The investor who says it can wait ten years may be describing its philosophy, while its credit facility is describing a different week.

A profitable parent with diversified cash generation can support a delayed project from elsewhere in its business. A thinly capitalized project company may depend on one customer's acceptance of one facility. Both can own useful infrastructure; their ability to refuse unattractive financing differs.

Van Nieuwerburgh's Brookings conference draft describes funding moving through leases, joint ventures, project debt and other structures. The mechanisms broaden financing capacity while creating exposure to concentrated tenants, execution delays and uncertain asset values. Separating an asset from a parent balance sheet changes who bears the risk; it does not make customer payments or collateral quality irrelevant.^P1

Meta's Hyperion arrangement makes the contractual detail concrete. In October 2025, Meta announced a venture in which Blue Owl-managed funds would own 80% and Meta 20%, funding approximately $27 billion of buildings and long-lived infrastructure. That development figure concerns the campus infrastructure, not an all-in budget for its chips.^M1

Meta described initial four-year operating leases with extension options, alongside a conditional, capped residual-value guarantee for the first 16 years of operations. Certain lease termination or non-renewal outcomes could require a payment based on the campus's then-current value. This was a development financing, not evidence of a rescue or a blanket guarantee of every creditor's loss.^M1

Read the risk allocation beneath the ownership percentage. Who owes rent? Who funds completion overruns? Who absorbs a weak resale value? Which obligations survive a decision to stop using the site?

A company can buy flexibility by agreeing to pay for particular consequences of exercising it. The exit option is real; so is the part of the contract that notices you used it.

Hardware also complicates the comforting historical analogy that infrastructure will eventually find a use. A useful site with power access and an ageing accelerator fleet are different assets. Model their replacement needs and residual values separately. The next owner may need substantial fresh capital before it can serve the next generation of demand.

An investor entering at a lower price, with funded reserves and a claim ahead of common, may be better positioned to wait. It still needs adequate collections and recoverable assets. Seniority improves the order of payment; it cannot manufacture money that the business and collateral do not contain.

The strongest counterargument deserves a proper place in the model. Adoption could outpace construction. Better utilization, lower delivery costs and valuable applications could produce durable margins. Long-term customer commitments could support financing on terms that preserve attractive shareholder returns. A company can also reject low-return expansion instead of treating every proposed campus as compulsory.

None of that requires believing that every announced project will succeed. It requires evidence that the particular company can collect enough cash, fund the wait and retain enough of the upside at the price being paid.

Founders already know how to present the market opportunity. The less glamorous discipline is making the financing assumptions equally visible. Put the following on the next board agenda:

1. Date the obligations. Show interest, amortization, maturity payments, leases, purchase commitments and unavoidable investment against collectible customer cash. Mark the earliest shortfall, even if the year as a whole looks solvent.
2. Separate committed funding from hoped-for funding. Identify unrestricted cash, drawable facilities and conditions still outstanding. A supportive investor's email belongs in a different column from funded cash.
3. Stress the commercial ramp. Test slower deployment, lower realized prices and a customer's inability to pay. Recheck replacement spending rather than freezing it to make the downside recover politely.
4. Price the next financing. Model an equity raise at a lower valuation, a more expensive refinancing and a case in which neither is available. Show how enterprise value translates into distributions to each security class.
5. Name the rights that move. Record proposed collateral, priority, consent rights and restrictions. Make counsel review the actual triggers before the board decides it can live with them.

A downside case that assumes an effortless new round has hired optimism as the restructuring adviser.

If you are buying infrastructure rather than building it, apply the same reasoning to your supplier. A low price is attractive until the contract depends on capacity the supplier has not yet financed. Ask what happens to delivery, prepaid money and service continuity if that financing changes hands.

The actionable deadline is the first date at which someone else's consent becomes necessary to keep the plan funded. Work backwards from it. Start the financing process while there are competing offers, consider smaller commitments, and compare accepting dilution now with negotiating after cash has already become scarce.

Before approving the next expansion, require a delayed-adoption case that names the next cheque writer, the amount needed and the ownership or rights they are assumed to receive. If that page is blank, the return forecast is missing a transaction.