Who Will Pay for the AI Buildout?

AI is already useful. Companies are paying for it. OpenAI and Anthropic report fast revenue growth. Cloud providers are expanding capacity. Developers use AI tools every day.

But that does not answer the financial question.

The real question is this:

Will AI generate enough revenue and productivity gains to justify the money now being committed to chips, data centers, electricity, land, debt, and replacement hardware?

That question matters because the spending has moved far ahead of what most companies can currently prove on their income statements. Amazon, Alphabet, Meta, and Microsoft were guiding toward roughly $725 billion of combined capital spending in 2026, up sharply from the year before. Much of that money is tied to AI infrastructure.

At the same time, most companies still struggle to show a clear return from AI adoption. The technology may be real while parts of the investment cycle still become overbuilt, overleveraged, or unprofitable.

That is how many technology bubbles work.

A Useful Technology Can Still Produce Bad Investments

People often treat the AI debate as a choice between two positions:

• AI will change the economy, so the spending must be justified.
• AI companies are losing money, so the technology must be fake.

Both arguments miss the point.

A useful technology does not make every company, project, or valuation attached to it a good investment.

The internet was real. The dot-com bubble was real too.

Fiber-optic networks were useful. Telecom investors still lost huge amounts of money when companies built too much capacity, borrowed too aggressively, and assumed demand would arrive faster than it did.

Railroads transformed the economy. Many railroad investors still went broke.

The value of the technology and the return earned by the investor are two different questions.

For AI, the first question is already being answered. The tools can write code, summarize documents, assist customer-support teams, search internal data, and automate parts of office work.

The second question remains open:

Can those benefits produce enough durable cash flow to support the current buildout?

The Spending Chain Behind the AI Boom

The AI economy is not one market. It is a chain.

1. Chipmakers sell GPUs, memory, and networking equipment.
2. Cloud companies buy that equipment.
3. Data-center developers build facilities to house it.
4. Utilities and power developers add generation and grid capacity.
5. AI model companies rent computing capacity.
6. Software companies build products on top of those models.
7. Businesses and consumers pay for the final products.

The strongest revenue today sits near the top of the chain.

Nvidia sells chips. Cloud providers sell computing capacity. Construction firms build facilities. Utilities sign new agreements. Those companies can report revenue before the final customer has proven that AI creates an adequate return.

That is where the investment risk begins.

A chip sale is real revenue for the chipmaker. It does not prove that the buyer will earn enough money from the chip.

A signed data-center lease may help a developer borrow billions of dollars. It does not prove the tenant will need the capacity for the full life of the facility.

A cloud contract may produce revenue for a hyperscaler. It does not prove that the enterprise customer has improved its own profit.

Each step can look healthy on its own while the full chain remains dependent on future demand.

Adoption Is Growing Faster Than Measurable Returns

Enterprise AI adoption is broad. Measurable financial results are much less common.

A Stanford-led survey of almost 6,000 executives found that about 70% of firms were using AI. More than 80% reported no impact on employment or productivity over the prior three years.

That does not mean those firms lost money. It means most had not yet seen a measurable company-level effect.

The widely cited MIT NANDA research made a similar point, though the headline is often distorted. The report said about 95% of the enterprise AI initiatives it reviewed showed no significant measurable effect on profit-and-loss statements during the study period.

That is different from saying 95% of AI projects were worthless.

A project may save small amounts of time, improve employee satisfaction, or create benefits that are difficult to isolate. Some projects may need years before the return becomes visible. The same research also found better results when companies bought specialized products or worked with outside vendors instead of trying to build everything internally.

Still, the gap is hard to ignore:

• infrastructure spending is already enormous;
• enterprise returns remain early and uneven;
• much of the investment case relies on future gains.

The market is funding the buildout before the average customer has proven the return.

Token Usage Is Not the Same as Economic Value

AI companies often highlight users, tokens, queries, coding sessions, and enterprise seats. Usage matters, but it is not the same as economic value.

A company can spend heavily on AI without changing how it operates. The useful questions are more direct:

• Did it reduce labor hours or another software expense?
• Did it increase revenue, output, or customer retention?
• Did it lower errors?
• Did the gain exceed the cost of the model, integration, security, and training?

Palantir CEO Alex Karp has argued that companies can spend heavily on tokens without connecting AI to real workflows and measurable outcomes. Palantir sells the integration layer it says solves that problem, so his commercial incentive should be kept in mind. The mechanism still matters: buying model access is not the same as improving a business.

Fixed-Price Subscriptions Can Hide Cost Problems

Consumer subscriptions create a related problem. Revenue per subscriber is fixed, but the cost to serve each user can vary widely. A heavy user or autonomous coding agent may consume far more compute than a casual user because the system plans, reads files, calls tools, retries, and revises its work.

Providers can lower inference costs, impose limits, route simple tasks to cheaper models, charge heavy users more, or subsidize consumer plans with enterprise revenue.

Public information does not prove that every additional ChatGPT or Claude subscriber loses money. The companies do not disclose enough unit-level data. But flat pricing can still hide large differences in customer profitability.

The Data-Center Financing Chain

This is where the AI bubble debate connects directly to private credit.

A data-center project may be placed inside a special-purpose vehicle, or SPV. The SPV owns the facility, equipment, contracts, and debt.

A simplified structure looks like this:

Investor or lender → data-center SPV → construction and GPUs → customer contract → revenue used to pay operating costs and debt

The project may rely on:

• a long-term lease;
• a compute contract;
• a minimum revenue commitment;
• a debt-service reserve;
• a parent guarantee;
• GPU infrastructure as collateral.

This is normal project finance. Power plants, airports, pipelines, and real estate projects also use SPVs.

The risk comes from the assumptions.

Will construction finish on time? Will the power connection arrive? Will the customer remain solvent? Can the customer cancel? Will the facility find a replacement tenant? How quickly will the chips lose economic value?

The debt may look safe because a large customer signed a contract. But the project still depends on that customer’s future ability and willingness to pay.

CoreWeave Shows How the Structure Works

CoreWeave offers a clear example. One recent $8.5 billion facility was tied to a large Meta agreement and used to finance GPU purchases.

The structure is straightforward:

1. Meta commits to purchase compute.
2. That contract supports a loan.
3. CoreWeave buys GPUs.
4. Customer payments service the debt.

The risk is concentration. Construction delays, a customer exit, weaker pricing, or lower demand could leave the project with debt and specialized hardware that is losing value.

Nvidia is also CoreWeave’s main GPU supplier and an investor. That does not prove fake revenue. It does show that supplier, investor, customer, and borrower relationships can overlap.

When Financing Becomes Circular

Circular financing can occur without fraud:

1. A supplier invests in a customer.
2. The customer buys the supplier’s products.
3. The purchase creates supplier revenue.
4. The contract supports more borrowing and purchases.

The original Nvidia–OpenAI proposal raised this concern. Nvidia discussed investing up to $100 billion in OpenAI while OpenAI purchased large amounts of Nvidia equipment. That proposal stalled and was replaced by a smaller $30 billion investment, so the original structure should not be described as completed.

The episode still raises a fair question:

How much demand comes from independent end customers, and how much comes from companies that are also investors, suppliers, creditors, or partners?

Public disclosures do not provide a clean answer.

Off-Balance-Sheet Does Not Mean Invisible

Large technology companies also use joint ventures, variable-interest entities, and long-term leases to finance data centers. Meta disclosed a maximum exposure of roughly $46 billion to one venture that it did not consolidate because it concluded it was not the primary beneficiary.

That accounting treatment may follow the rules. The commitment can still matter.

Calling every structure “hidden debt” is too aggressive because these obligations are often disclosed. Investors still need to look beyond the headline debt number and include leases, guarantees, venture commitments, and other future payments.

The Telecom Boom Is the Better Historical Comparison

The subprime comparison helps explain SPVs, leverage, institutional exposure, and weak underwriting.

The telecom boom is the better comparison for the investment cycle itself.

During the 1990s, companies built huge amounts of fiber capacity based on expectations of future internet traffic. The demand eventually arrived. It did not arrive fast enough to save many of the companies that financed the buildout.

The Richmond Fed found that AI-related equipment and software investment has followed a path similar to the early telecom boom. The current level of AI-related investment is much larger in absolute terms, and data-center construction has accelerated more quickly than telecom-facility construction did in the 1990s.

The lesson is not that the infrastructure will be useless.

The lesson is:

Productive infrastructure can still be built too early, at the wrong price, with too much debt.

If AI demand catches up after prices fall and weaker developers fail, later users may benefit from cheap capacity built by earlier investors.

That would be good for the technology and painful for the capital providers.

High Rates Do Not Prevent Bubbles

Low interest rates often help bubbles grow, but they are rarely the sole cause. High rates do not make bubbles impossible.

Bubbles usually form through some combination of:

• abundant financing;
• leverage;
• optimistic forecasts;
• weak constraints;
• delayed accountability;
• a believable story about future profits.

The AI boom has several of those features even with higher rates.

The largest hyperscalers generate enormous operating cash flow. Private funds still need assets that offer higher yields. Long-term customer contracts make projects look bond-like. Supplier commitments support financing. Investors fear missing the next major technology cycle.

The narrative is believable because AI is useful.

That may make the boom stronger, not safer.

A weak story cannot attract capital for long. The most powerful bubbles are often built around a real change that investors extend too far.

How the AI Buildout Could Break

There is no single required trigger.

A bust could begin with weak enterprise renewals, low utilization, falling compute prices, a hyperscaler cutting spending, construction delays, or a major customer missing payments. Any of those events could hit chip orders, developers, utilities, neoclouds, and lenders.

The cycle does not need one dramatic failure. A slowdown in new commitments may be enough to expose projects that depend on continuous financing and optimistic demand forecasts.

Why an AI Bust Would Be Deflationary First

An AI capital-spending bust would probably push the economy toward deflation before any policy response arrived.

It would reduce:

• construction;
• equipment orders;
• semiconductor demand;
• hiring;
• private-credit creation;
• utility investment;
• commercial real estate activity;
• stock-market wealth.

Prices could fall for GPUs, data-center capacity, technology shares, private-company stakes, and related property.

The feedback loop would look familiar:

Lower demand → weaker revenue → tighter financing → canceled projects → layoffs → lower demand

That is the same basic credit mechanism seen in other investment busts and market crashes.

The later response could be inflationary. The Federal Reserve might cut rates, which works differently from balance-sheet policy. Governments might add subsidies, guarantees, or fiscal support. Utilities may pass stranded infrastructure costs to ratepayers.

The sequence could become:

AI investment bust → credit and asset deflation → policy response → later reflation

An oil shock or other supply problem could complicate the timing. Consumer prices may stay high while asset prices and business investment fall.

That would feel stagflationary before the deflationary pressure became dominant.

The Strongest Case Against the Bubble Thesis

The bearish case has real weaknesses.

AI revenue is growing quickly. OpenAI and Anthropic both report large increases in annualized revenue. Cloud divisions are expanding. Enterprise adoption is broad. Falling model costs may create more demand. New products may appear that are difficult to imagine today.

The largest hyperscalers also have strong balance sheets. They can fund much of the buildout from operating cash flow rather than fragile short-term borrowing.

Some apparent overcapacity may prove temporary. Data-center projects take years to complete. Comparing today’s demand with every announced project can exaggerate the amount of supply arriving soon.

The market may also be underestimating how quickly usage grows once costs fall.

Those points matter. They are why the conclusion should remain conditional.

The current evidence does not prove that the entire AI buildout is a bubble.

It does show that the return has not yet been demonstrated at the same scale as the spending.

What Would Prove the Spending Was Justified?

The bullish case gets stronger if we see:

• AI revenue growing faster than AI infrastructure spending;
• measurable enterprise profit gains across a wider range of companies;
• independent customers becoming a larger share of compute demand;
• high utilization after new capacity becomes operational;
• falling cost per useful task;
• stronger free cash flow after capital spending;
• low default rates among data-center borrowers;
• projects refinancing without new guarantees or subsidies;
• durable demand after introductory pricing and pilot programs end.

The bearish case gets stronger if we see:

• hyperscalers cutting capex guidance;
• large project cancellations;
• weak utilization;
• falling compute prices without enough volume growth;
• customer concentration increasing;
• more supplier-backed purchases;
• rising credit spreads on data-center debt;
• covenant breaches;
• loan restructurings;
• GPU collateral values falling quickly;
• enterprise AI spending failing to produce measurable returns.

What to Watch Next

The AI buildout is too large to judge through one earnings call or one failed project.

Watch the relationship between five things:

1. Capital spending
2. AI revenue
3. Enterprise returns
4. Compute utilization
5. Credit performance

The technology can succeed while investors lose money.

The infrastructure can prove useful while developers fail.

Revenue can grow quickly and still fall short of the assumptions inside the financing.

The question is not simply if AI works.

The question is who pays for the buildout, how long they keep paying, and if the cash flow arrives before the debt and depreciation do.

See how AI capital spending, credit stress, and Federal Reserve policy are currently interacting on the Macro Board Watch dashboard.

This article is educational and is not investment advice.