There are two honest ways to talk about the AI infrastructure boom. One starts with demand. Model usage is rising, enterprise budgets are moving from pilots to deployment, and cheaper inference can create work that did not make sense at older prices. The other starts with financing. The largest technology companies are building so much physical infrastructure that their old habit of paying from operating cash flow is no longer the whole story.

The second frame is starting to matter more.

The radar item that caught my eye was a JPMorgan note, reported by Investing.com and Yahoo Finance, arguing that the AI capex cycle looks more economically viable than it did six months ago. The argument is reasonable. JPMorgan points to faster AI-company revenue growth and estimates cumulative AI data-center capex through 2030 around $5.5 trillion, with some external estimates as high as $10 trillion. Its equity analysts reportedly see AI cloud providers, model providers, and neoclouds reaching a combined revenue run rate of about $1.6 trillion by the end of 2026, then rising toward $2.5 trillion to $3 trillion by 2030.

That is the steelman. If the revenue base is already that large, the buildout is not pure faith. The required enterprise-spending shift also looks less absurd than the harshest bubble arguments imply. JPMorgan’s Asia-Pacific survey found average AI spend rising from 4.5 percent of expenses plus capex over the prior year to 5.8 percent over the next year. Applied globally, that points to roughly $1.7 trillion of AI spending. Getting to $2.5 trillion by 2030 would require something like 6.5 percent to 7 percent, according to the same report. That is a stretch, but it is not fantasy if AI starts replacing labor, legacy software, and outsourced services rather than merely joining the software stack.

I buy more of that argument than the simple bubble label allows. A data center that is already leased to a hyperscaler is different from a fiber trench laid into speculative demand. GPUs are not empty dot-com office space. The big buyers have customers, cash flows, distribution, and real usage. The question is how much of the future surplus they keep after they pay for chips, power, land, cooling, construction, leases, and debt service.

That is where the story changes shape.

The Bank for International Settlements put the financing issue cleanly in a January 2026 bulletin. AI investment is surging both in nominal terms and as a share of GDP, and the anticipated investment needs are large enough that firms will have to shift from funding the boom mainly with operating cash flows toward more debt. The BIS also notes that private credit is playing a rapidly increasing role. It estimates annual data-center spending could rise by $100 billion to $225 billion over the next five years, taking data-center spending from about 0.5 percent of GDP today to 0.8 percent to 1.3 percent.

Those percentages look small until you remember they describe one slice of one technology cycle. At that scale, financing terms are no longer background plumbing. They are part of the product economics.

FactSet’s July work gives the clearest operating snapshot I found. Aggregate capex for Alphabet, Amazon, Meta, Microsoft, and Oracle rose from about $95 billion in fiscal 2020 investing cash flows to roughly $490 billion in the twelve months to May 2026. FactSet expects more than $690 billion in fiscal 2026 and more than $900 billion by fiscal 2028. Calendar 2026 guidance points closer to $800 billion when finance leases and customer prepayments are included. It also expects fiscal 2026 free cash flow to approach zero or turn negative for all except Alphabet and Microsoft.

That last sentence is the hinge. These firms entered the cycle as unusually profitable businesses with fortress balance sheets. The AI buildout is turning several of them into something closer to infrastructure companies with software margins still under negotiation.

Goldman Sachs Research says the large technology companies leading the buildout may spend a combined $5.3 trillion from 2025 through 2030. Goldman also argues that private markets will matter more in data-center financing, including infrastructure funds, real estate structures, investment-grade debt, and other private asset channels. Its analysts point out that hyperscaler capex estimates are growing faster than actual data-center construction. That gap matters. Planned capacity is a promise. Built, powered, leased, and utilized capacity is a cash-flowing asset.

The bullish version is straightforward. Hyperscalers borrow because rates are manageable, capacity is scarce, and demand is visible enough to justify locking in supply. External financing is rational when the asset is pre-leased, the tenant is strong, and the alternative is losing share in the most important compute market of the decade. Debt can be the right instrument for an infrastructure asset with contracted revenue.

The bear version is also straightforward. The financing stack is expanding before the revenue stack has proved its long-duration margin. Some obligations sit in leases, project vehicles, private credit deals, and supplier financing rather than plain corporate bonds. That does not make them fake. It changes who has the claim and when the claim bites. A customer can stop experimenting with a model faster than a data-center owner can redeploy a site designed around a specific power envelope, network layout, and tenant requirement.

This is why the fight over whether AI is a bubble often feels badly specified. Bubble relative to which claim?

For Nvidia equity, the bet is mostly about accelerator demand, gross margin, and the durability of the upgrade cycle. For a hyperscaler, the bet is about utilization, cloud pricing, customer retention, and whether AI spend replaces other costs. For a private credit lender, the bet is about collateral, tenant quality, structure, and recovery value if demand arrives later than promised. For a utility customer, the bet may show up as grid investment and higher bills. For an enterprise buyer, the bet is whether AI spend lowers labor or software costs enough to justify a larger budget line.

These are related trades, but they are not the same trade.

The JPMorgan note is useful because it refuses the lazy version of the bear case. If AI revenue really scales toward the reported numbers, a large capex cycle can be economically coherent. The BIS, Goldman, and FactSet work add the part that equity narratives tend to compress. Coherent does not mean self-funding. Coherent does not mean every layer keeps attractive economics. Coherent does not mean the capital markets can absorb the same exposure forever without demanding better terms.

I would watch the financing terms before I watch another total-addressable-market slide. Spreads on AI data-center debt, guarantees from tenants or chip suppliers, lease duration, residual-value assumptions, project-level covenants, prepayment structures, and private-credit participation will say more about the true distribution than another headline capex number. If lenders keep accepting long maturities, moderate spreads, and weak guarantees, the market is saying the buildout’s cash flows look bankable. If structures become shorter, more secured, more tenant-specific, and more expensive, credit is marking down the story even if equity is still applauding growth.

The labor-substitution assumption deserves the same treatment. AI spend becomes durable when it comes out of an existing cost base. If a bank spends more on model inference and less on outsourced document review, the supplier mix changes but the budget has a funding source. If a software company spends more on coding agents while holding headcount flat, the economics can work. If AI spend remains an additive experiment across departments, the revenue line can grow for a while and still disappoint the capital stack behind it.

This is where who pays becomes concrete. Enterprise customers pay through budgets. Cloud providers pay through capex and capacity commitments. Utilities and ratepayers may pay for grid upgrades. Lenders pay upfront and hope the contracted cash flows arrive on schedule. Equity holders pay if dilution, debt service, or lower terminal margins take more of the upside than the growth story assumed.

My prior is that the AI infrastructure cycle is real, overbuilt in places, and underpriced in its second-order claims. The technology can be useful and the financing can still get crowded. Railroads mattered. Telecom mattered. Data centers matter. History is full of useful infrastructure that gave customers more surplus than the people who financed the first wave expected.

The distribution I want is simple. How much AI revenue becomes durable free cash flow after chips, power, leases, debt, and replacement cycles? The answer will not be the same for Nvidia, Microsoft, Oracle, a private credit fund, a utility, and a customer trying to cut support costs. That is the point. The AI capex boom has left the clean world of product demos and entered credit’s jurisdiction.

Sources: Yahoo Finance and Investing.com coverage of JPMorgan’s AI capex analysis; BIS Bulletin No. 120, “Financing the AI boom: from cash flows to debt”; FactSet, “Hyperscalers Tap External Financing as AI Capex Outruns Cash Flow”; Goldman Sachs Research, “Private Markets Are Expected to Have a Growing Role in Data Center Financing.”