NextFin News - The artificial-intelligence buildout has entered a more expensive phase, and the bottleneck has moved. It is no longer just the supply of chips that constrains the race; it is the price of the electricity to run them and the cost of the debt used to pay for them. As the four largest U.S. technology companies steer roughly $725 billion of capital spending into 2026 — up 77% from last year's record — the industry is learning that every dollar of AI capacity now carries a heavier charge for power and financing than it did a year ago.
The Bill Arrives in Two Parts
The scale of the commitment is difficult to overstate. Amazon, Microsoft, Alphabet and Meta are collectively guiding to about $725 billion of capital expenditure in 2026, according to a tally of company earnings disclosures tracked by a running industry analysis. That is nearly double the roughly $410 billion spent in 2025, and analysts at Evercore and Bank of America already expect combined big-tech capital spending to exceed $1 trillion in 2027. Amazon leads at approximately $200 billion, followed by Microsoft near $190 billion, Alphabet at $175 billion to $185 billion, and Meta in a $115 billion to $135 billion range.
But the bill is arriving in pieces, and each piece has grown more expensive. The first is power. Wholesale electricity at grid nodes near large data-center clusters costs as much as 267% more than it did five years ago, according to an analysis of local pricing points. In the PJM Interconnection, which serves much of the data-center-heavy Mid-Atlantic, capacity prices jumped from $28.92 per megawatt-day in the 2024/25 auction to $269.92 for 2025/26 and $329.17 for 2026/27 — an elevenfold move in two years. An independent market monitor for PJM attributes 63% of the 2025/26 increase to data-center demand, translating into $9.3 billion of costs passed through to ratepayers. The 2026/27 clearing price would have been higher still if the grid operator had not imposed a cap.
The second is financing. The pure-play AI cloud providers — the "neoclouds" that rent GPU capacity to AI labs — have leaned heavily on debt and structured financing to fund their buildouts. CoreWeave, the largest of them, closed a $2.6 billion delayed-draw term loan on August 10, priced at SOFR plus 5.50 percentage points and arranged by JPMorgan and Mitsubishi UFJ. With that deal, the company has secured more than $30 billion of debt and equity capital this year. Its own filing with the Securities and Exchange Commission shows three unnamed customers accounted for 36%, 26% and 10% of second-quarter revenue — 72% combined. That is concentration risk financed at floating rates.
"For now, the economics are holding. But the margin for error is narrow."
Depreciation alone still consumes more than two-thirds of revenue, according to the same analysis of the figures. Global AI sales, excluding China, reached $25 billion for hyperscalers and neoclouds in the first quarter of 2026, exceeding the industry's estimated $21 billion in depreciation costs tied to data centers and chips for the second consecutive quarter. That narrow margin is the story. The AI race has not slowed. It has simply gotten more expensive to run, and the incremental cost is flowing to two groups that were bit players in the first act of the boom: utilities and bondholders.
Why the Bottleneck Moved From Chips to Kilowatts
For the first three years of the generative-AI boom, the binding constraint was obvious: whoever had the most graphics processors won. In 2026, the chips exist. What does not, in enough places, is the electricity to run them and the cooling to keep them from overheating. Microsoft has disclosed an $80 billion backlog of Azure orders it cannot fulfill because of power constraints — demand that is real but unmonetizable until the electrons arrive.
The physics explain why this is a structural shift rather than a cyclical squeeze. An AI-optimized rack draws 30 kilowatts to more than 100 kilowatts, compared with 5 kilowatts to 15 kilowatts for a traditional data-center rack. That density overwhelms local distribution networks built for lighter loads, and grid interconnection queues do not clear on a product cycle's timetable. U.S. data centers already consume about 4.4% of national electricity, and industry projections put combined demand nearly doubling between 2025 and 2028, from 80 gigawatts to 150 gigawatts — equivalent to adding a country with Spain's power needs in three years. Goldman Sachs projects the AI buildout will lift electricity costs 6% between 2026 and 2027 and another 3% by 2028.
This is not a cost that disappears when chip supply catches up. Transmission lines take a decade to permit and build. The constraint is durable, which means the cost is durable. Electricity already accounts for 20% to 30% of a data center's operating expense, and that share rises as racks grow denser. Site selection has shifted from proximity to users to proximity to power.
The takeaway: the scarce factor of production has changed. In the first act, scarcity was semiconductors; in the second, it is electrons and interconnection rights. That is a structural change, and structural changes do not mean-revert on their own.
Where the Cyclical Risk Concentrates: The Financing Structure
If power is the structural leg of the problem, financing is where the cyclical risk lives — and where the industry's second-order vulnerability shows up. The neoclouds illustrate the pattern. They raise debt against GPU purchase contracts, buy chips from Nvidia, rent capacity to AI labs, and in some cases pay Nvidia a share of the cloud revenue. Critics call it circular financing: the same capital flows from the chipmaker to the customer and partially back. Nvidia is both the primary hardware supplier and, in CoreWeave's case, an equity holder. The arrangement worked comfortably while GPU demand was insatiable and rates were contained.
It is working less comfortably now. CoreWeave's August facility was priced at SOFR plus 5.50 percentage points and received non-investment-grade ratings — Ba2 from Moody's and BB+ from Fitch. The deal's structure is telling: it carries an approximate five-year maturity while the underlying customer contracts average about three years, a mismatch the company framed as a way to finance shorter-dated contracts. CoreWeave's chief development officer, Brannin McBee, called the facility "a major unlock" and said lenders are "now comfortable financing shorter-dated contracts." Comfort, in this market, carries a price. Every refinancing is more expensive than the last when the floating reference rate and the credit spread are both moving against the borrower.
Customer concentration compounds the interest-rate exposure. If one of CoreWeave's three dominant customers slows its AI spending or renegotiates its contract, the debt service does not slow with it. The same math applies at hyperscaler scale. Alphabet sold $25 billion of bonds in August against an AI order book reported at more than $115 billion — a sign that even the strongest balance sheets are reaching for capital-market funding rather than relying solely on operating cash flow.
The takeaway: power determines how fast capacity can be built; financing determines who survives the wait. The second-order point the market has not fully absorbed is this: the revenue story is priced, but the margin compression is not. Every incremental dollar of AI revenue now carries a heavier power and financing load than the dollar before it.
The Payoff Is Real — and Thinner Than the Headlines Suggest
The strongest evidence that this is not a bubble is the revenue print. AI sales of $25 billion against $21 billion of depreciation means the industry, excluding China, covered its capital-cost charge for a second straight quarter. That separates this cycle from infrastructure booms that produced no revenue at all.
But the margin is the message. Depreciation consumes more than two-thirds of revenue, leaving a thin buffer for everything else: power, which runs 20% to 30% of a data center's operating expense; financing costs that are rising as facilities refinance; and the next generation of chips, which must be purchased before the current generation is fully paid off. The model works only if revenue keeps compounding fast enough to outrun an escalating cost base.
That is a narrower proposition than "AI demand is strong." Demand can be strong and the economics can still deteriorate if the cost of serving it rises faster. This is the trap embedded in the buildout: capex is being spent today to capture revenue that arrives later, while the cost of the capital and the power in between keeps climbing. A company can be right about demand and wrong about returns.
The Counter-Thesis, and What Would Break It
The strongest case against a darker read is simple history. Every midpoint of the AI buildout has been declared a bubble — the 2023 chip shortage, the 2024 inference-cost panic, the 2025 capex scare. Each time, demand absorbed the new capacity. Revenue covered depreciation, and then some. Investor Mark Cuban, speaking on the All-In podcast, predicted "a lot of data centers…are going to be turned into pickleball courts" because AI will become cheaper to run through efficiency gains. He may be right about efficiency without being right about demand: if inference costs fall fast enough, utilization rises and the same fixed infrastructure serves more revenue.
The counter-thesis has a named constituency and a logical foundation. Efficiency gains are real, and AI adoption is still on the early part of the adoption curve. It deserves its weight — at least a third of the argument, by any honest measure.
But it rests on two assumptions the 2026 data does not yet support. First, that efficiency gains outrun the escalation in power and financing costs — when wholesale power near data centers is up as much as 267% in five years and floating-rate debt is repricing at spreads above 5 percentage points. Second, that utilization stays high — when the hyperscalers are adding capacity faster than revenue, by design, to chase a backlog they cannot yet serve. The pickleball prediction requires idle capacity to be temporary. If power delays stretch into 2027 and 2028, temporary becomes the new normal for a portion of the fleet.
The falsifying signal: watch the quarterly ratio of AI sales to depreciation. If that ratio falls back below 1.0 for two consecutive quarters while capex guidance keeps rising, the claim that "the economics are holding" breaks. A second signal sits in credit markets: a sustained widening of neocloud credit spreads above their issuance levels would tell you lenders are repricing the financing risk faster than operators can pass it through.
Who Wins, Who Is Exposed, and What to Watch
The rising price of the AI race redistributes winners and losers without slowing the race itself. The beneficiaries are clear: utilities and independent power producers with data-center load in their interconnection queues; makers of grid equipment and cooling systems; data-center operators that locked in power contracts before the repricing; and hyperscalers with balance sheets strong enough to issue debt at tolerable rates. Power is becoming a moat — capacity secured early is now a structural advantage that competitors cannot quickly replicate.
The exposed are equally clear: neoclouds financed at floating rates with concentrated customer books; regional utilities whose ratepayers absorb capacity-market pass-throughs, a political risk that is already drawing scrutiny from lawmakers; and any AI application company whose unit economics assume inference costs keep falling faster than infrastructure costs rise.
The outlook splits by time horizon, and the horizons point in different directions:
- Short term (6 to 12 months): sentiment stays supported as long as the revenue-versus-depreciation gap holds above 1.0 and bond sales like Alphabet's keep clearing. Volatility will cluster around capex guidance and power-auction results.
- Medium term (1 to 3 years): the field consolidates. Operators that cannot refinance at tolerable rates, or cannot secure power, sell assets or merge into better-capitalized rivals. Margins compress even as revenue grows.
- Long term (3 years and beyond): if the grid catches up and efficiency gains compound, the cost per unit of compute falls and the industry normalizes into a high-volume, lower-margin utility-like business. If grid buildout lags, AI growth itself becomes power-constrained, and the winners are those who own electrons, not just chips.
The AI boom is not ending. It is maturing into something more expensive, more concentrated, and more dependent on infrastructure that cannot be built at the speed of software. The companies that treated power as a line item will learn it was the strategy.
The first act of the AI race was won by whoever had the chips. The second will be won by whoever has the power — and the balance sheet to wait for it.
Explore more exclusive insights at nextfin.ai.
