The $730B AI Capex Trade: Utilities Are the New Tech Stocks

January and February 2026 earnings calls set the tone for the year. Each of the four largest cloud platforms disclosed capital spending plans well above what analysts had modeled just a quarter earlier.

Amazon told investors it would direct close to $200 billion toward infrastructure in 2026, up from roughly $125 billion the year before. Alphabet followed with guidance in the $175 billion to $185 billion range, nearly double its 2025 total. Microsoft signaled it would push past $120 billion for the fiscal year, and disclosed an Azure order backlog it could not fill because of power constraints rather than chip supply.

None of the four companies described this spending as discretionary. Cloud executives repeatedly used a specific phrase on their calls: capacity constrained, not demand constrained. That distinction matters for how the money gets spent. When demand is the limiting factor, companies pause. When capacity is the limiting factor, they build faster, and building faster means pouring concrete, ordering turbines, and signing power purchase agreements years in advance.

The backlog behind the number

By the close of February, sell-side estimates for combined 2026 hyperscaler capex had moved from the mid-$500 billion range to a cluster around $690 billion to $725 billion, depending on which companies were included in the count.

Alphabet's disclosure carried an additional data point that reinforced the demand story: Google Cloud's backlog of signed but unfulfilled contracts had climbed past $460 billion, a figure that suggests current spending plans are catching up to already-booked business rather than getting ahead of speculative demand. Analysts covering the sector noted that this backlog dynamic is unusual for a capital-intensive industry, where spending typically leads demand rather than trailing it.

Credit analysts at several fixed-income research desks flagged a related shift worth watching: a growing share of hyperscaler capex is now going toward leased rather than owned data center capacity, which reduces upfront cash requirements but shifts long-term cost exposure onto the infrastructure landlords and power suppliers on the other side of those leases.

A theme emerged through March and April: the binding constraint on AI infrastructure had shifted from graphics processors to electricity.

Chip supply, while still tight, has become more predictable as Nvidia and its competitors have scaled production and diversified packaging capacity. Power has not scaled at the same rate. Grid interconnection queues in major U.S. markets now stretch several years, and utilities have warned that new large-load requests are arriving faster than transmission upgrades can be permitted and built.

Goldman Sachs research circulated in the spring estimated a U.S. data center power shortfall exceeding 11 gigawatts at the time, with that gap projected to widen toward 45 gigawatts by 2028 absent a faster buildout of generation and grid capacity.

This is the pivot point for the trade. Every dollar of hyperscaler capex that cannot be deployed because of a missing substation or an unbuilt gas turbine is a dollar that gets rerouted toward the companies that can solve the power problem. Data center operators have responded by contracting directly with power generators, sometimes bypassing the public grid entirely through behind-the-meter arrangements that let a facility draw power straight from an on-site turbine.

Meta's May 2026 earnings call became the clearest signal yet that hyperscaler spending discipline was loosening further, not tightening.

The company raised its full-year 2026 capex guidance to a range of $125 billion to $145 billion, up from a prior range of $115 billion to $135 billion, citing higher component pricing and additional data center construction costs. Mark Zuckerberg framed the increase around building infrastructure for what he called personal superintelligence, a framing that unsettled some investors rather than reassuring them.

Meta shares fell more than 9% the following trading session, a reaction that reflected genuine investor tension. The market wants the AI buildout to continue. It also wants evidence that the spending will eventually convert into proportionate revenue. Nvidia's own numbers offered some support for the bull case: fourth-quarter data center revenue reached roughly $62 billion, up about 75% year over year, with networking revenue growth running even faster.

Even so, the ratio between AI infrastructure spending and AI-attributable revenue remains lopsided across the industry, and that gap is precisely why some analysts argue the more durable trade sits one layer removed from the hyperscalers themselves, in the physical infrastructure that gets paid regardless of which cloud provider or model wins the underlying compute race.

Utilities spent the spring translating data center demand into signed, dollar-denominated contracts, and that is where the capex story starts showing up in company-specific numbers rather than macro estimates.

Constellation Energy and Vistra

Constellation Energy has positioned its nuclear fleet as the most reliable large-scale power source for data center operators seeking firm, carbon-free electricity. The company's agreement to restart a reactor at the former Three Mile Island site to supply Microsoft's future power needs became one of the most cited examples of a hyperscaler locking in a nuclear power source years ahead of a data center's completion. Constellation generated roughly $4.2 billion in operating cash flow during 2025 while continuing to sign long-term contracted demand from data center and industrial customers.

Vistra, an integrated power producer with a diversified generation fleet, has pursued a different angle. Beyond signing its own long-term agreements with hyperscalers, Vistra benefits when localized power shortages push wholesale electricity prices higher, since its merchant generation fleet captures that upside directly. The company also returned capital to shareholders through buybacks even as it expanded its contracted data center pipeline.

GE Vernova, NextEra and Talen

GE Vernova has emerged as the equipment supplier at the center of the buildout. The company reported that it had already secured more than 90% of its gas turbine production capacity through 2030, an order book so full that customers have begun paying upfront to reserve manufacturing slots. Its total backlog reached roughly $263 billion in the first quarter of fiscal 2026, and quarterly orders jumped more than 70% year over year. GE Vernova's equipment increasingly supports behind-the-meter installations, letting data center developers generate power on-site rather than wait years for a grid interconnection.

NextEra Energy has scaled renewable generation, battery storage, and grid modernization investment to meet the same underlying demand curve, positioning itself as a generation-side beneficiary even without the same nuclear-specific narrative that has lifted Constellation.

Talen Energy has followed a path similar to Constellation's, pairing existing nuclear generation capacity with direct data center power agreements rather than relying solely on wholesale market pricing. That structure gives the company more predictable, contracted revenue over a multi-year horizon, which is precisely the kind of visibility fixed-income and equity investors have rewarded across this group of names since the start of 2026.

What separates this cycle from prior utility investment cycles is the counterparty. Historically, utilities signed long-duration contracts primarily with industrial manufacturers or municipal buyers operating on thin margins. Today's counterparties are some of the highest cash-generating companies in the world, committing to decade-long power purchase agreements as a strategic necessity rather than a cost-cutting exercise, which changes the credit profile of the contracts utilities are now booking.

The industrial side of the trade extends beyond pure-play utilities into the equipment and construction companies that physically build the connections between generation and compute.

Vertiv Holdings sits inside the data center itself, supplying power distribution and thermal management equipment that keeps AI server racks running at the density modern chips require. The company's organic orders grew roughly 81% in a recent period, and its executive chairman has described AI-driven infrastructure build-out as the defining driver of the business.

Quanta Services, an electrical infrastructure and grid construction contractor, has become a direct beneficiary of the transmission and interconnection backlog utilities are racing to clear. As more data center campuses require new substations and transmission lines, contractors capable of executing that work at scale have seen order books extend years into the future.

Bloom Energy, which manufactures on-site fuel cell power systems, has attracted attention for offering fast-to-deploy power that sidesteps both grid interconnection delays and gas turbine lead times, though its generation technology carries a materially different cost and efficiency profile than utility-scale gas or nuclear power.

Taken together, these companies represent the picks-and-shovels layer beneath the picks-and-shovels layer. Nvidia sells the compute. GE Vernova, Vertiv, and Quanta sell the physical means to power and cool that compute, and their order books are being filled by signed contracts, not by speculative demand forecasts.

Investors looking for exposure to this theme without concentrating risk in a handful of single stocks have a few established, diversified routes into the same underlying trend.

• Utilities Select Sector SPDR Fund (XLU) offers broad exposure to regulated and merchant power companies, including several of the names directly benefiting from data center demand, alongside more traditional utility holdings that carry less direct AI exposure.
• Global X U.S. Infrastructure Development ETF (PAVE) holds a mix of industrial, materials, and engineering companies tied to broader infrastructure spending, capturing some of the grid-construction and equipment names that intersect with the data center buildout.
• iShares Global Infrastructure ETF (IGF) extends similar exposure internationally, which matters given that data center construction and grid investment are accelerating outside the U.S. as well, particularly in markets with available land and cooler climates.

More narrowly targeted infrastructure and utility funds focused specifically on power generation, grid modernization, or data center real estate have also launched or expanded in response to investor demand, though these carry higher concentration risk than the broad sector funds above.

Anyone considering a fund vehicle should read the underlying holdings directly, since fund names built around AI or data center themes vary widely in how much of their portfolio is actually tied to power generation versus other technology exposure.

None of this removes the risks embedded in a trade built on capital spending projections rather than realized earnings.

The first risk is capex discipline. Meta's stock reaction in May showed that investors will punish guidance increases if they read them as evidence of a spending race rather than a return-on-investment plan. If one or more hyperscalers signals a slowdown in 2027 guidance, utility and industrial order books built around continued growth could reprice quickly.

The second risk is regulatory and political friction around large-load electricity requests. Data center power demand has already drawn scrutiny in several U.S. states over rate impacts on residential customers, and any material shift in how utility regulators treat large industrial customers could alter the economics behind current contracts.

The third risk is technological. If a more efficient training or inference approach meaningfully reduces the power intensity of AI workloads, the extraordinary capacity buildout could outrun actual demand, echoing the market's brief but sharp reaction to the DeepSeek R1 cost claims that briefly wiped hundreds of billions of dollars off Nvidia's market capitalization earlier in this cycle.

Sri Lankan investors watching this theme from afar have no direct listed exposure to Constellation, Vistra, or GE Vernova on the Colombo Stock Exchange, but the broader signal is relevant locally: global capital is being redirected toward power generation and grid infrastructure at a scale that could eventually influence turbine, transformer, and grid-equipment pricing and availability for power projects in smaller markets as well.

If hyperscaler capex guidance holds through the rest of 2026 and utilities continue converting data center demand into signed, long-duration contracts, the utilities and industrials layer of this trade should keep compounding alongside the chip names that started it. If capex guidance stalls or reverses in 2027, the companies most exposed to near-term power buildout commitments will feel that reversal first and hardest.