NextFin News - Dan Dees, co-head of Goldman Sachs' Global Banking and Markets, has a warning for investors who still think of artificial intelligence as a software story: the real transformation is physical, and it is only beginning. In a Sept. 14, 2026 newsletter appearance, Dees described AI's impact as having a broad "blast radius" that reaches far beyond chips and code into power grids, data centers, credit markets, and the industrial economy — a buildout he sizes in the trillions.
The message reframes the entire AI investment thesis. For the past three years, the market has rewarded the picks-and-shovels layer — chip designers, cloud hyperscalers, and the model builders renting their capacity. Dees argues that phase is giving way to something wider and more capital-intensive: the construction of the physical economy that AI requires to exist at scale.
The Physical Buildout Nobody Priced For
The central tension is this: the market has spent the AI boom pricing semiconductors and cloud hyperscalers, but the bottleneck — and the larger capital opportunity — has migrated to concrete, steel, and electrons.
Goldman Sachs' investment banking division, in its "Powering the AI Era" research, projects hyperscalers will invest $1 trillion in AI by 2027, surpassing the more than $800 billion spent on internet infrastructure across the entire dot-com era. A 2026 follow-up, "Harnessing AI for the Real Economy," pushes the horizon further: hyperscalers alone are projected to invest more than $6 trillion in AI through 2030, with total funding needs of $7 trillion or more between 2026 and 2031 across compute, data centers, and power. Goldman Sachs Global Institute estimates roughly $7.6 trillion of aggregate AI capital expenditure across those three categories over the same period.
The year-by-year breakdown of that $7.6 trillion reveals the escalation curve. Compute spending alone is projected to rise from $494 billion in 2026 to $1.13 trillion by 2031. Data center investment climbs from $232 billion to $436 billion over the same window. Power spending — the smallest line item today — grows from $39 billion in 2026 to $73 billion by 2031, nearly doubling as the grid becomes the binding constraint.
Even the full sum, Dees notes, equals only about 1.2% of today's annual U.S. GDP — a reminder that the buildout is large in absolute terms but affordable in macro context. The question is not whether the capital exists. It is whether the physical inputs can be delivered on time.
"The future of AI will not be forged in code and large language models alone. It will be built with concrete, steel and silicone," Dees said.
Power Is the Binding Constraint
The most consequential detail in the Goldman analysis is not how much money is being spent, but what stands in the way of spending it. The binding constraint on AI progress is no longer capital — it is power.
Global data center power demand is expected to rise 50% by 2027, with 60% of that growth requiring new generating capacity, and to climb 160% by 2030, according to Goldman Sachs Global Investment Research. AI server racks in 2027 will require 50 times more power than cloud-computing equivalents did five years earlier, forcing a shift to liquid cooling and denser facility designs.
Supply is not keeping pace. Data center vacancy sits at a record low of 3%, and near zero in the most sought-after markets. About 60% of data center demand growth by 2027 will need to be met with new power capacity — a problem that cannot be solved with software efficiency gains alone. Utility-sector capital expenditures have more than doubled over the last decade, but the aging U.S. power grid is not yet ready for AI-driven growth.
"The most critical obstacle for unleashing AI's potential is not capital—it's power," the firm's research states.
This is where the second-order effect kicks in. A power-constrained AI buildout does not simply slow the technology — it reroutes capital. Every dollar that would have gone to model training or software development must first be spent on substations, transmission lines, and generation. The marginal product of AI capital falls in the near term because a growing share of each dollar is consumed by infrastructure that produces no intelligence on its own.
The Capital Stack Has to Be Reinvented
Financing a buildout of this scale requires more than the traditional playbook. For cloud data centers in the prior cycle, most costs sat in the "shell" — power, land, physical infrastructure — and financing looked like real estate. AI data centers invert that: the servers and GPUs inside are the dominant cost, and they depreciate faster than the buildings that house them.
That mismatch breaks the old financing model. A lender accustomed to underwriting a 20-year building against 10-year leases now faces a facility whose most valuable components — the GPU clusters — may be obsolete in three to five years. The residual value assumptions that made data center real estate investment trusts a staple of institutional portfolios no longer hold.
Dees argues the full capital stack will be needed: equity, public and private debt, sovereign capital, and joint-venture structures, some of which have yet to be invented. "Most of the capital that will define the AI economy has not yet been deployed, most of the infrastructure has not yet been built, and most of the M&A that will shape the competitive landscape has not yet been executed," he said.
The debt market is already moving. Nearly $500 billion of AI-related debt issuance has come to market in 2026, according to Goldman Sachs Research, even as cash-rich technology companies turn to bonds to fund buildouts. Credit strategists at the firm have flagged the rise of AI debt as a structural shift in credit markets, with implications for institutional investors and for the pricing of risk in a sector where revenue visibility remains uneven. Data center CMBS and ABS issuance has begun to scale alongside, drawing on structures familiar from real estate finance but applied to assets with technology-cycle risk.
Beyond the Hyperscalers: The Neocloud Rise
The "blast radius" framing captures a second-order development the market is still digesting: AI infrastructure is decentralizing. Outside the traditional cloud hyperscalers, a new class of "neoclouds" is emerging — companies spending billions on GPUs and data centers to serve specialized AI workloads. These players concentrate investment in the same power-constrained markets, intensifying competition for electricity, land, and interconnection queues.
The implication is that the AI trade is broadening rather than narrowing. Software companies are beginning to reclaim a monetization narrative, and industrial firms are becoming AI plays by virtue of their exposure to the buildout. The beneficiaries are no longer confined to the chip designers at the top of the stack.
Mark Sorrell, global head of the Industrials Group in Goldman Sachs' Investment Banking division, put the convergence in sharper terms: the transition is "unfolding in real time—building the infrastructure required to advance AI, while AI is rewriting the rules of the global economy."
Data Center Diplomacy: Infrastructure as Geopolitics
There is a third dimension to the blast radius that reaches beyond finance: geopolitics. Goldman's research identifies "data center diplomacy" as an emerging tool of statecraft — the use of access to AI infrastructure, power, and advanced chips as leverage between nations.
Because AI data centers require concentrated power and specialized cooling, they cannot be built just anywhere. Countries with abundant, cheap electricity and stable grids gain bargaining power; countries dependent on imported energy or foreign-built infrastructure face new vulnerabilities. Export controls on advanced semiconductors have already turned chip access into a diplomatic instrument. The next layer — who can host the compute, and who controls the power behind it — extends that logic into energy policy and real estate.
For investors, this adds a political risk premium to any asset tied to the AI buildout. A data center project permitted today can be delayed tomorrow by a change in energy policy, an interconnection dispute, or a shift in export rules. The companies best positioned are those with diversified power exposure and sovereign relationships that can survive policy transitions.
Cyclical Capex Boom or Structural Regime Shift?
Here is the judgment call that determines how investors should read the next five years. Is this a cyclical capital expenditure super-cycle that will mean-revert once capacity catches up with demand — or a structural regime shift in how the economy is powered and financed?
The evidence points to structural. Three comparisons anchor the call. First, the dot-com era's $800 billion-plus internet infrastructure spend was concentrated in fiber and network gear; today's AI buildout must also rebuild the power grid beneath it, a layer of capital intensity with no historical parallel outside wartime mobilization or national electrification. Second, the power intensity — 50 times the rack density of five years ago — is not an efficiency problem that chip advances will solve away; it is a physics problem. Third, the timeline extends to 2031 and beyond, with the productivity payoff arriving only after the infrastructure is in place, not before.
Goldman's own historical analogies make the point. In 19th-century America, railroads turned localized markets into a national economy, with railway stocks reaching an 80% share of the U.S. stock market and 215,000 miles of track laid by 1900. In the 1920s, electrification drew $295 billion from utilities and added 50 gigawatts of grid capacity in a single decade. AI, in this telling, is the third great infrastructure inflection — and like its predecessors, it is defined less by the technology itself than by the capital required to scale it.
There is also a compositional difference that matters. The dot-com buildout was largely privately financed through equity markets at peak euphoria. The AI buildout is being funded by cash-generative incumbents with investment-grade balance sheets, supplemented by debt and, increasingly, sovereign capital. That funding mix makes the cycle more resilient to a sentiment shock but more exposed to a policy shock.
"The AI era is driving industrial transformation unlike any in modern history—faster, broader, and more capital-intensive, creating new infrastructure and industry simultaneously," Dees said.
The Counter-Thesis: A Dot-Com Echo?
The strongest argument against the structural read is the one every bull has to answer: the dot-com bust followed an infrastructure overbuild, and AI capex could meet the same fate if monetization lags. In a Goldman Sachs Exchanges discussion on the AI investment boom, the firm's analysts noted that enterprise adoption has been slower than expected roughly three and a half years into the cycle, and revenue per dollar of AI infrastructure spend remains unproven at scale.
The mechanism of a bust is easy to trace. If AI-generated revenue fails to cover the cost of the compute that produced it, hyperscalers face a choice: keep spending and accept negative returns, or cut capex and strand the assets already ordered. A coordinated capex cut would ripple backward through the stack — GPU orders cancelled, data center leases renegotiated, power purchase agreements repriced. The 2001 telecom crash, in which dark fiber traded for cents on the dollar, is the template bears point to.
This is a real risk, but it attacks the timing of returns, not the direction of capital. Even in a downside scenario where AI revenue disappoints, the power and data center capacity must be built before the disappointment becomes visible — the capital is front-loaded, the payoff back-loaded. That asymmetry is what separates this cycle from dot-com: the overbuild risk sits in servers and software, but the grid and facility buildout has value independent of any single AI application. Fiber optic cable laid in the wrong place was worthless; electricity transmission built for AI data centers can serve cloud computing, electrification, and industrial load regardless of which AI model wins.
The falsifying signal is specific and observable: if hyperscaler capital expenditure guidance is cut by more than 15% year over year for two consecutive quarters, or if AI-related credit defaults rise above 5% of the sector's debt stock, the structural thesis breaks and the buildout should be read as a cyclical boom heading into bust. Until then, the base case is continued escalation.
What to Watch: Three Horizons
Short term (6-12 months): Watch utility capital expenditure announcements and data center power procurement deals. Any acceleration in grid investment is the first read-through that the power constraint is being taken seriously. Equity markets will likely remain focused on hyperscaler earnings guidance — the single most important number for the entire stack. A guidance cut from any of the largest cloud builders would transmit immediately across semiconductors, data center REITs, and independent power producers.
Medium term (1-3 years): The credit market becomes the story. As AI-related debt issuance climbs past the 2026 run rate of nearly $500 billion, spreads and covenant structures will reveal how much risk lenders are willing to underwrite. Data center REITs, independent power producers, and cooling-technology suppliers are the direct beneficiaries; any company with a long-dated, contracted power supply becomes an asset. Watch for the first wave of AI-data-center securitizations and how they are rated — that will set the pricing benchmark for the rest of the cycle.
Long term (3-5 years and beyond): The question shifts from "who builds it" to "who owns it." Sovereign capital and joint ventures will reshape ownership of critical infrastructure, and data-center diplomacy — the use of infrastructure access as geopolitical leverage — becomes a policy tool. The companies that control power and sites, not just models, capture the durable rents. Efficiency breakthroughs in chips and energy could stretch each dollar of infrastructure further, allowing more buildout with less capital — the most likely source of upside surprise.
Three scenarios frame the path. The base case: capex continues to escalate through 2031, power becomes the scarcest input, and the AI trade broadens from a narrow tech rally into a multi-sector capital formation cycle. The upside case: efficiency breakthroughs reduce the capital intensity per unit of compute, letting the buildout go further with less funding while power constraints ease. The downside case: enterprise monetization fails to materialize, hyperscaler guidance cuts cascade, and the debt market reprices AI risk sharply, stranding the most speculative neocloud capacity first.
The bottom line: AI's blast radius is widest where the market is least looking — in the grid, the grid operators, and the balance sheets funding the build. Investors betting on code alone are betting on half the story.
Explore more exclusive insights at nextfin.ai.
