NextFin News - Stonepeak chief executive Michael Dorrell is signaling that the artificial intelligence data-center boom is still being funded at pace even as the physical bottlenecks behind it become harder to ignore. In comments tied to a July 29 television interview, he said banks, private equity and capital markets show “very little sign of slowing down” the buildout, while warning that the pace “cannot go on the way it is going forever.” The tension in that view is the point: AI demand remains strong, financing is still flowing, and the limiting factor is shifting toward the grid, permitting and power delivery.
The question is no longer whether capital will show up. It already has. The question is which parts of the stack will capture the economics once power, land, water and interconnection become the scarce inputs. That is why Dorrell’s remarks matter beyond one interview. They describe a capital cycle moving from software and chips into the much less flexible world of transmission, substations, cooling and utility planning.
Financing Still Flows, but Power Sets the Ceiling
Stonepeak’s own platform activity shows how aggressively capital continues to chase the theme. In April 2025, the firm launched Montera Infrastructure with a $1.5 billion equity commitment to develop and operate hyperscale data centers in North America. Stonepeak said Montera will focus on land with a clear path to near-term power, plan 100-plus megawatt facilities and support cloud computing and AI inferencing workloads. In January 2025, Stonepeak said Digital Edge had raised more than $1.6 billion in new equity and debt capital. The platform then said it had 21 data centers, more than 500 MW of critical IT load in service, under construction or development, and another 300 MW held for future development.
Those are not isolated balance-sheet anecdotes. They show the financing stack broadening around digital infrastructure, with investors now underwriting access to electricity as much as access to demand. That shift is consistent with official demand forecasts. Goldman Sachs Research estimates U.S. data center power demand will rise from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027. It also projects that data centers’ share of U.S. peak summer power demand will rise from 4.1% in 2025 to 5.3% in 2026 and 8.5% in 2027. The U.S. Energy Information Administration said in January that U.S. electricity use is expected to grow 1% in 2026 and 3% in 2027, which would mark four straight years of growth for the first time since 2007 and the strongest four-year run since 2000.
That combination changes the market’s center of gravity. The bottleneck is no longer just capital. Money can move quickly; electrons cannot. If demand for compute keeps compounding faster than grids can expand, then the winning projects will be those that secure power first and finance second. Everyone else is left competing for a queue. In that sense, Dorrell is describing a market where the old advantage of cheap capital is still useful, but no longer sufficient.
This is also why the boom looks structural rather than cyclical. A cyclical surge would be defined by temporary overexuberance in capex, an easy financing window and then a retreat once demand normalizes. Here, the driver is more durable: AI workloads are permanently lifting the floor for power demand, and that forces investors to reprice not just data-center real estate, but the entire chain of generation, transmission, interconnection and cooling that sits underneath it.
The Mechanism Is a Capacity Race, Not a Pure Technology Trade
The critical distinction is that data-center growth behaves like an industrial bottleneck story, not a classic technology euphoric cycle. In a software cycle, supply can often scale with code. In a data-center cycle, supply has to pass through physical constraints: land, substations, gas turbines, transmission lines, switchgear, cooling and local approvals. That creates the second-order effect the market sometimes misses. As AI demand grows, the scarce asset is not the server rack. It is the right to consume power at the right time in the right place.
Stonepeak’s portfolio language is consistent with that mechanism. The firm said Montera is designed around sites with near-term power and that Digital Edge is building across Japan, Korea, India, Malaysia, Indonesia and the Philippines. Stonepeak also said its global digital infrastructure portfolio spans more than 100 facilities and more than 500 MW of capacity, with a pipeline of more than 400 MW under development. That is a portfolio positioned around constrained supply, not just demand growth. The point is not merely to own more buildings. It is to own the scarce inputs that let buildings operate.
The same pattern appears in the broader market. The EIA’s forecast implies that U.S. electricity demand is entering a rare multi-year growth phase, while Goldman Sachs’ estimate implies that data centers alone could account for 8.5% of peak summer power demand by 2027. That matters for two reasons. First, the data-center boom starts competing directly with industrial users and households for capacity. Second, the more power demand concentrates in a few regions with available land and transmission, the more pricing power shifts toward those with early site control, utility relationships and permitting advantage. What looks like a real-estate spread trade on the surface becomes, underneath, a fight over queue position and grid rights.
That is a structural change because the old playbook is losing relevance. Data centers used to be judged mostly as real-estate-like assets with a tech tenant. Now they resemble a hybrid of utility, industrial park and balance-sheet business. The history of earlier digital booms is only partially useful because earlier cycles did not collide with a grid that is already tight in many markets and a regulatory environment that can slow transmission buildout. The old comparison no longer fits cleanly. If anything, the new analogy is closer to an industrial export boom in which the winners are the firms that secure ports, rail and power before demand peaks.
There is a second-order implication here that is easy to miss. A stronger AI buildout does not just lift demand for data-center shells; it can also change the relative value of adjacent assets. Equipment makers that supply switchgear, transformers, cooling systems and backup generation become more strategic. Utility-linked infrastructure, transmission corridors and power development platforms can become more valuable than the operating boxes themselves. The market can still call this “AI infrastructure,” but the economics are increasingly determined by the least glamorous components of the stack.
Why This Time Is Different, and Why It Still May Not Be Linear
Is this cyclical or structural? The right answer is both, but at different horizons. The short-term cycle is clearly cyclical: financing can overshoot, project announcements can outrun actual deliveries, and local conditions can create sharp bursts and pauses. Capital markets are procyclical by nature, and any asset class tied to scarce power can get overbid when the theme is hot. But the medium- to long-term driver is structural because the underlying demand for compute, storage and inference is being recast as a persistent demand for electricity and related infrastructure.
Three historical comparisons support that call. First, earlier cloud and colocation cycles expanded quickly but were not yet constrained by today’s combination of AI intensity and grid congestion. Second, prior industrial buildouts often depended on cheap power and abundant land, conditions that are now scarcer in many data-center hubs. Third, the current wave is occurring as U.S. electricity demand is projected by the EIA to post its strongest four-year run since 2000, which means the backdrop is not a one-off temporary spike. A cyclical correction can still happen in project valuations. A structural shift in the underlying demand curve can still survive that correction.
That distinction also explains why Dorrell’s warning is not bearish on AI itself. It is bearish on the assumption that all parts of the AI stack can scale at the same speed. The bottleneck theory says the system will keep growing, but not evenly. Bottlenecks do not kill demand; they ration its expression. They can move profits from one layer of the stack to another, from the operator of a building to the owner of a substation, from the chip supplier to the landholder, from the tenant to the utility partner. That is the real propagation chain.
What, then, is the market already pricing? On one level, it is pricing abundant capital and strong demand, which is why Stonepeak and its peers keep raising and deploying money. On another level, it is only partially pricing the consequence of that demand colliding with hard physical limits. If the grid cannot expand fast enough, then the valuation of projects should increasingly depend on power certainty rather than demand optimism. That is the second-order question the market is only beginning to ask.
The strongest counter-thesis is that this is still early innings in a long secular buildout, and that utilities, developers and capital providers will adapt fast enough to absorb the demand. On that view, current bottlenecks are not warnings; they are opportunities to earn higher returns by solving the hard parts of the chain. Stonepeak’s own activity lends weight to that case, since the firm keeps investing behind AI-ready capacity and power-linked assets. It is reasonable to argue that a better-capitalized industry can extend the cycle longer than skeptics expect.
But that counter-thesis has to clear a hard test. It must explain away the pace of demand growth and the persistence of power friction. If U.S. data-center power demand does not stay on the path implied by the EIA and Goldman Sachs, or if interconnection queues, power pricing and permitting timelines improve enough to clear the current pipeline without major cost inflation, then the structural bottleneck thesis weakens. Until then, the market is still pricing a buildout that depends on more megawatts, not just more money. That is the threshold that matters.
Who Benefits, Who Gets Exposed
If the bottleneck is power, the beneficiaries shift toward operators that already control land, transmission access and local relationships. Developers with deep experience in substations, cooling design and utility approvals gain value. Infrastructure investors that can provide capital while sharing execution risk also become more important. Stonepeak is effectively describing that strategy when it says it wants to build around power-secured sites and when it keeps expanding into digital infrastructure through platforms such as Montera and Digital Edge. The emphasis moves from pure exposure to AI growth toward control of the infrastructure that lets AI growth happen.
The exposed side is equally clear. Speculative developers without power certainty face delays and margin compression. Projects that rely on optimistic interconnection schedules can see returns reset if permitting slips or if local opposition raises the cost of execution. Communities and utilities are also exposed because a faster buildout can pressure grids, raise local congestion and force political scrutiny. Dorrell’s warning that the current pace cannot continue forever is, in practical terms, a warning about duration risk: the market can keep financing the boom, but not every project can make it through the physical queue. The faster capital moves ahead of infrastructure, the more sorting the market will eventually have to do.
“You don’t need to be a finance person to see this cannot go on the way it is going forever, but I don’t see a pull-back at the moment,” Dorrell said.
The short-term outlook is therefore still constructive for platforms that can secure the scarce inputs first. Scarcity itself is a source of value when everyone is chasing the same grid connection. The medium-term outlook is more selective: returns should increasingly favor firms that control the hard bottlenecks rather than firms with the broadest exposure to AI demand. The long-term outlook is the most important. If the current pattern persists, the sector may look less like a software trade and more like a regulated utility-plus-infrastructure complex with higher capital intensity, longer lead times and lower margin for error.
The base case is continued capital deployment, slower project conversion and more emphasis on power-secured assets. The upside case is that utilities, equipment vendors and specialized infrastructure platforms accelerate fast enough to unlock a larger buildout without a major rise in financing stress. The downside case is a financing air pocket: the market keeps announcing projects, but interconnection delays, local political resistance and higher power costs force a reset in expected returns. Each scenario is rooted in the same mechanism, which is why the near-term number to watch is not only how much capital is raised but how much capacity actually reaches service.
The next signals to watch are straightforward: electricity-demand forecasts, interconnection timelines, permitting friction and the size of new data-center financings. If those variables improve together, the boom can keep stretching. If they diverge, the market will learn that the binding constraint was never enthusiasm. It was electricity. In that sense, the AI trade is not ending; it is narrowing to the assets that can survive the queue.
Explore more exclusive insights at nextfin.ai.
