The Picks-and-Shovels Rotation: Why the Smartest AI Money Is Moving Downstream to Power, Cooling, and Silicon

The Stack Has Inverted

For most of 2023 and into 2024, the prevailing AI investment thesis centered on software: foundation model rounds, SaaS wrappers, AI-native startups. That logic is colliding with concrete constraints—megawatts, square footage, interconnection queues. The physical layer—power, land, cooling, grid access, silicon supply chains—has become the binding constraint. Sophisticated institutional capital has recognized this shift and is rotating accordingly.

The arithmetic is straightforward. Goldman Sachs projects roughly $7.6 trillion in cumulative AI capital expenditure between 2026 and 2031, spanning compute, data centers, and power. The breakdown: compute at $5.1 trillion, data centers at $2.1 trillion, and power at $358 billion. That power figure appears modest relative to compute, but it is the bottleneck through which every dollar of compute must flow.

What the Energy Data Actually Shows

The IEA's April 2026 update confirmed the demand shock is real and accelerating. Electricity demand from data centers soared 17% in 2025, and AI-focused data centers climbed even faster, well outpacing global electricity demand growth of 3%. More strikingly, electricity consumption from AI-focused data centers surged 50% in 2025 alone.

The forward curve steepens. Data center electricity consumption is set to double by 2030, and power use from those focused on AI will triple. The IEA's base case projects global data center electricity consumption to reach around 945 TWh by 2030, representing just under 3% of total global electricity consumption. A demand signal of this duration and scale gives utility planners, power equipment manufacturers, and cooling suppliers something to underwrite.

The counterargument exists: power consumption per AI task is declining rapidly, with efficiency improving at an unprecedented rate—but more people are using AI, and energy-intensive uses such as AI agents are proliferating. Both trends are real. The latter is currently winning.

Where the Capital Is Actually Going

The capital bifurcation is structural. VC still chases model and application-layer risk. PE and institutional capital is rotating into contracted, cash-flowing physical assets.

Driven by data center investments, the capital expenditure of five large technology companies surged to more than $400 billion in 2025 and is set to increase by a further 75% in 2026. These hyperscaler commitments create the demand signal that makes infrastructure assets contractable. KKR has stated the logic plainly: those who control the moats should reap compounding returns—power, land, grid connections, and permits are structural bottlenecks to building data centers.

The money follows the thesis. KKR, the Kuwait Investment Authority, NVIDIA, and Vistra launched Helix Digital Infrastructure with more than $10 billion in committed capital to deliver data centers, power, and connectivity for hyperscalers; former AWS CEO Adam Selipsky is leading it. This is PE vertically integrating into power generation and data center construction—a direct structural response to interconnection bottlenecks.

Transaction data backs this up. Data center deals hit $61 billion in 2025. North American colocation vacancy fell to approximately 1.4% by year-end 2025, down from 9.8% in 2020. Secure offtake agreements guarantee a tenant will pay for contracted capacity over a period of time, whether they use it all or not—a financing structure far more compatible with PE and infrastructure debt than VC equity models.

The Grid Is the Real Constraint

Every projection above rests on one hard physical limit: grid interconnection. Investment in AI data centers is accelerating faster than power grids were designed to accommodate; in many regions, connecting a new facility to the grid can take 4 to 10 years, while AI data centers are typically planned and built within two to three.

Goldman documents an 11 GW U.S. power shortfall today, widening to 40-plus GW by 2028. U.S. transmission investment remains flat at around $25–30 billion annually despite hyperscaler CapEx surges. Nearly 40% of data center projects in development this year risk significant delays, with insufficient power access cited as a leading reason. A Goldman Sachs analysis from May found that as little as 50% of data center capacity scheduled to come online in the next two years is on track to do so on time.

This is the thesis, not a flaw in it. Scarcity of permitted, grid-connected land is the moat. Whoever controls that access commands pricing power that persists across every AI generation, regardless of which model architecture or chip dominates.

The SMR Signal

One indicator cuts through speculation: the pipeline of conditional offtake agreements between data center operators and small modular reactor nuclear projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts today. Hyperscalers are not signing 20-year nuclear offtakes expecting demand to evaporate. They are signing them because waiting in interconnection queues for utility power is incompatible with their build timelines.

The Risks Are Real

This thesis carries material risks worth holding alongside the bullish case.

First, the $7.6 trillion figure is highly sensitive to assumptions about how the infrastructure itself is built and renewed. Small shifts in GPU replacement cadence, architecture mix, or data center costs move cumulative spend by hundreds of billions.

Second, the grid bottleneck that creates infrastructure scarcity also defers revenue and compresses IRRs on contracted assets. Thirty to fifty percent of planned 2026 AI data center capacity is projected to slip to 2028. Contracted is not synonymous with on-time.

Third, neocloud operators—the primary conduit for infrastructure capital—carry concentrated counterparty risk. Revenue concentrated among Microsoft, Meta, and OpenAI introduces cycle exposure that asset-backed lending structures do not fully mitigate.

Fourth, this shift from silicon efficiency to physical scale is precisely why grid connectivity has become the binding constraint—but it also means policy or permitting changes could unlock supply faster than current models predict, compressing returns for investors who underwrote peak-scarcity scenarios.

The Operator Takeaway

For founders and CTOs: your AI infrastructure cost structure increasingly turns on power access, not compute pricing. Operators with long-term power purchase agreements and grid-connected land will have structural cost advantages over those procuring capacity spot-market as interconnection queues lengthen.

For investors: KKR joining EQT in treating AI compute infrastructure as a long-duration asset class signals that institutional capital has formed a consensus view on physical AI demand extending well beyond a single model cycle. The question is no longer whether the physical layer matters. It is whether you underwrote the right power access, in the right jurisdiction, at the right contract duration, before the queue closed.

As compute, storage, and energy converge, controlling scarce inputs—power, land, and grid access—will define who wins. That was written by a PE firm with billions deployed on precisely this premise. The software narrative was not wrong. It was just early in the stack.