Wealth managers say the AI infrastructure race is early-to-middle innings on spending — but that permitting, power, and monetization risk are changing who benefits
Hyperscaler capital expenditure on AI infrastructure is projected to exceed $690 billion in 2026 alone, according to data compiled by Data Center Watch, while US utilities have committed to a further $1.4 trillion in grid infrastructure spending through 2030. Yet the first quarter of 2026 produced the largest single-period disruption to that buildout on record: at least 75 data center projects worth approximately $130 billion were blocked or delayed between January and March, matching the full-year total for 2025, according to a June 2026 report from Data Center Watch, a tracker maintained by AI intelligence firm 10a Labs.
The number of active opposition groups more than doubled to 833, spanning 49 states. New York’s legislature passed a one-year permitting moratorium on large data centers in June, and the bill awaits Governor Kathy Hochul’s signature. For advisors and investors trying to calibrate their AI infrastructure exposure, the question is no longer simply how much is being spent — it is where the value actually lands.
Three investment professionals with distinct vantage points — a CIO focused on public equities, a CIO managing $11 billion in RIA assets, and a real asset specialist serving ultra-high-net-worth families — offer frameworks that diverge sharply on which part of the buildout deserves the most attention.
The market is questioning monetization, not capacity
Nate Garrison, CFA, Senior Vice President and Chief Investment Officer at World Investment Advisors, argues that the physical buildout itself remains early-to-middle innings. What has changed is the market’s willingness to extend unlimited patience to hyperscalers on AI capital expenditure.
“The question is no longer just, ‘Can they build enough compute?’ It is also becoming, ‘Can they monetize it reasonably fast enough to justify the spending?'” Garrison said.
Garrison draws a sharp distinction between the hyperscalers themselves and the picks-and-shovels suppliers who have been the cycle’s clearest near-term winners: chip manufacturers, memory providers, networking infrastructure, power equipment, cooling systems, and data center hardware. The hyperscalers, he argues, are effectively transferring enormous free cash flow to these suppliers — but that phase will not persist indefinitely. The long-term winners will be the platforms capable of converting compute spending into recurring, subscription-style revenue from enterprise customers.
“Microsoft appears especially well positioned because it has a massive installed base and deep enterprise relationships, giving it a natural distribution advantage. As AI becomes embedded in the daily workflow of corporate America, I think Microsoft has one of the clearest paths to turning AI infrastructure spending into durable recurring revenue,” Garrison said.
Garrison also makes an important distinction between two types of spending slowdown. If hyperscalers reduce capital expenditure because they are becoming more disciplined and focused on capital efficiency, he views that as potentially healthy — and potentially positive for their stocks as investors shift attention to free cash flow generation. If spending slows or stops because AI monetization disappoints, the repricing would be systemic, falling hardest first on the suppliers whose valuations are most directly tied to continued hyperscaler capex. The broader risk, in his view, is not just lower tech valuations — it is the potential deferral of what could be one of the most significant productivity-enhancing technologies of our generation.
Edge computing is the next phase — and the winner is still unknown
Matt Dmytryszyn, CFA, Chief Investment Officer at Composition Wealth, places the buildout at roughly the top of the fourth inning.
His framing draws on a structural shift in what data centers are being built for. The first phase of construction was driven primarily by model training — the computationally intensive process of building foundation AI models. As AI usage progresses and agentic AI adoption accelerates, Dmytryszyn expects demand to shift toward smaller, distributed data centers positioned at the edge, closer to end users. That transition, he argues, represents the next material phase of the buildout and creates a different set of infrastructure and investment requirements.
On the question of who captures the most value, he is deliberately agnostic. A year ago, he notes, Anthropic was viewed as a niche player; its rapid market share gains in 2026 have surprised even close observers of the space. That humility about forecasting individual winners drives him toward companies with multiple structural paths to benefit from the AI transition.
“It’s too early to call who is going to be the winner. The bigger question for us is how the economics will play out between the model provider, the hyperscaler that provides the compute capacity, and the semiconductor and equipment supplier. At this time, Amazon appears well-positioned, given its capacity as well as its own Trainium custom silicon chipsets. Google also appears attractive with full vertical integration stemming from their custom TPU chips to hyperscale compute capacity, and their Gemini models,” Dmytryszyn said.
Dmytryszyn’s preference for companies with multiple avenues to win reflects a broader investment discipline: in a technology cycle where market share is still in flux and the economic structure between model providers, infrastructure operators, and semiconductor suppliers has not yet settled, concentration risk is high.
The real constraint is land, power, and entitlements — not announced capital
Michael Shawn, founder of Peregrine Private Client, brings the most contrarian and operationally grounded perspective of the three. His view, drawn from direct exposure to real asset markets rather than public equity analysis, is that the conversation about data center spending misses what is actually scarce.
The headline figures — 16 gigawatts of data center capacity announced for 2026 — obscure what Shawn estimates is closer to 5 gigawatts actually available on the ground, after accounting for permitting delays, power grid constraints, transformer lead times of up to five years, and turbine order books full into 2030. The $130 billion in Q1 2026 disruptions documented by Data Center Watch, he notes, represents in a single quarter approximately what was blocked in all of the prior year.
“Here’s where I’d push back on consensus. In a permit-constrained market, the winner isn’t whoever spends the most. It’s whoever already controls the land, the interconnect, and the power. Announced capital is not deliverable capacity. Those are different things, and the market keeps pricing them the same. Which means the opposition movement just handed the incumbents a moat they never had to pay for,” Shawn said.
Shawn’s framework — evaluating data center assets the way he would any real asset, assessed through the lens of dirt, power, entitlements, and time — points toward a structural opportunity that he argues public market investors are systematically mispricing. In a permit-constrained environment, existing, fully permitted, interconnected campuses in cooperative jurisdictions carry a different risk-return profile than announced projects still dependent on regulatory approval. The opposition movement, paradoxically, has created a defensible moat for whoever already holds those assets.
Taken together, the three perspectives frame the data center buildout not as a single trade but as a multi-layered set of investment decisions: public equity exposure to hyperscalers and their suppliers, portfolio positioning toward companies with multiple AI monetization pathways, and real asset access to the permitted capacity that has become structurally scarce. Each lens implies a different set of risks, timelines, and client suitability considerations — and each is worth examining separately before treating “AI infrastructure” as a single, undifferentiated allocation.
