Markets are pricing in a historic AI spending surge. Goldman Sachs says even that is too conservative. Both of those things cannot remain true simultaneously — and one of them is about to be proven wrong.
The apparent paradox at the heart of Goldman Sachs’s latest AI market read is this: after two years of breathless AI coverage, record-breaking GPU orders, and trillion-dollar valuations assigned to companies riding the wave, the investment bank is arguing that the consensus is still undercounting the magnitude of the shift underway. The AI boom, Goldman contends, is bigger than investors think — a claim that deserves scrutiny rather than reflexive endorsement.
The Three Things Worth Knowing
-
The Assumed Story: Markets Have Already Priced In AI Dominance
The conventional wisdom heading into mid-2025 is that AI enthusiasm is already baked into valuations. Nvidia’s market capitalisation has swelled to rival the largest companies on earth. Hyperscaler capital expenditure — the hundreds of billions poured annually into data centres, chips, and networking by companies like Microsoft, Amazon, Google, and Meta — has become a recurring headline. Analysts covering these companies have repeatedly revised their AI revenue forecasts upward, and the broad S&P 500 technology sector has re-rated significantly on the expectation that AI will prove transformative.
The logical inference from all of this, widely held among investors, is that the easy money has been made. The AI trade, in this view, is mature — still valid as a long-term theme, but no longer the source of alpha it was in 2023. Goldman’s report lands as a direct challenge to that assumption.
-
The Overlooked Angle: Infrastructure Demand Is Structurally Undercounted
Goldman’s core argument is not simply that AI will be important — that debate is largely settled. The more pointed claim is that the scale of required investment to deliver on AI’s promise has been systematically underestimated, particularly in physical infrastructure. Data centres, power grids, cooling systems, and the semiconductor supply chains that feed them are all implicated. The bank’s analysis suggests that current capex projections from the major hyperscalers, while large in absolute terms, are likely to be revised upward as model complexity increases and enterprise adoption accelerates beyond early-adopter cohorts.
This is a meaningful distinction for capital allocators. It shifts attention from software-layer AI plays — where valuations are already aggressive — toward the picks-and-shovels layer of the economy: utilities, data centre REITs, networking hardware, and specialised chip manufacturers. It also implies that the energy sector, frequently overlooked in AI coverage, sits at the critical path of the entire buildout. As Amazon engineers have publicly flagged, the internal pressure to accelerate data centre deployment is immense — even when it strains other parts of the organisation.
There is a telling alignment between Goldman’s macro-level infrastructure thesis and the ground-level signals coming from inside the hyperscalers themselves. When engineers at one of the world’s largest cloud providers raise concerns about the pace of a $200 billion data centre push, and a leading investment bank simultaneously argues that even that pace is being underestimated, the two data points together suggest not irrational exuberance but a genuine supply-demand mismatch in AI infrastructure — one that could persist for years rather than quarters.
The broader implication: investors focused exclusively on AI model companies and application-layer startups may be looking at the wrong part of the supply chain. The physical world — land, power, silicon — is where the near-term constraint sits.
-
What This Changes: The Frame for Evaluating AI Investments
If Goldman’s read is correct, it has practical consequences for how investors should evaluate AI-exposed positions. The standard analytical frame has been to assess which AI companies will win the model race — a question that has driven intense focus on frontier model competition between players like Anthropic and OpenAI and their respective enterprise traction. That question remains live, but Goldman’s thesis implies it may be secondary to a more structural question: who controls the infrastructure those models run on?
Power availability, chip supply, and data centre density are not software problems. They are solved on timescales of years, not sprints, and they create durable competitive moats that are harder to dislodge than a model architecture advantage. Nvidia’s position — which Jensen Huang has been communicating with increasing directness — is partly a bet that hardware constraints will persist long enough to cement its ecosystem lock-in. Goldman’s framing lends analytical support to that view, even if the bank stops well short of a stock recommendation.
There is also a timing dimension. Enterprise AI adoption tends to lag infrastructure deployment by twelve to eighteen months as companies work through integration, security review, and workflow redesign. This means the demand curve Goldman is projecting has a compounding character: infrastructure built today supports adoption waves that haven’t yet materialised, which in turn justify further infrastructure investment. The cycle is self-reinforcing — until it isn’t, which is the risk.
The Strongest Counterargument
The most credible pushback to Goldman’s thesis comes from a growing cohort of economists and technology analysts who argue that AI capital expenditure is racing well ahead of demonstrable economic returns — a version of the critique that has appeared in Goldman Sachs’s own research. Earlier analysis from the bank’s own economists questioned whether AI productivity gains would be broad enough and fast enough to justify the scale of investment underway, citing historical parallels with previous technology cycles where infrastructure overbuilding preceded painful corrections.
Critics in this camp, including some prominent voices in enterprise software analysis, point to the gap between AI feature announcements and measurable revenue impact at the business level. If enterprise adoption stalls — because integration is harder than marketed, because regulatory constraints tighten, or because early AI tools underdeliver on efficiency promises — the infrastructure demand projections underpinning Goldman’s bullish read could prove premature rather than conservative. The question of whether most AI spending generates genuine ROI remains genuinely open.
Does this counterargument fatally weaken Goldman’s conclusion? Not entirely. The bank’s thesis is primarily about the scale of the buildout relative to current market pricing, not about whether every dollar of AI capex is optimally deployed. Even in a scenario where enterprise adoption is slower than hoped, the physical infrastructure required to support the models already in production — and the regulatory and geopolitical pressure to onshore AI capacity — creates a structural floor for demand. The correction risk is real; the thesis that the boom is larger than priced does not require every assumption to be right simultaneously.
Where This Ends Up
The most likely outcome is that Goldman’s infrastructure-first framing proves directionally correct over a three-to-five year horizon, even if the path is uneven. Physical AI infrastructure — power, chips, data centres — is genuinely supply-constrained in ways that software is not, and the compounding adoption dynamic means demand has structural momentum. Investors who have concentrated AI exposure in application-layer software may find themselves revisiting the picks-and-shovels layer as the cycle matures. The next generation of AI computing hardware, including novel chip architectures, will likely attract significant capital as current silicon approaches physical limits.
The second-most-likely outcome is a partial correction in AI infrastructure sentiment — not a collapse, but a repricing — if enterprise adoption metrics disappoint over the next two to three quarters. That scenario would tip the balance if large enterprises begin reporting that AI tools have failed to deliver the efficiency gains projected in their business cases, triggering a pause in internal AI budgets that flows back to hyperscaler capex guidance. The condition to watch: enterprise AI renewal and expansion rates in the second half of 2025. If they soften, the bears will have their evidence. If they hold or accelerate, Goldman’s call will look prescient rather than promotional.
