In June 2024, Goldman Sachs published a report titled "Too Much Spend, Too Little Benefit?" questioning whether the AI infrastructure investment cycle would ever produce returns commensurate with its cost. In May 2026, Goldman Sachs published a separate analysis naming Nvidia and Micron the biggest winners of the AI earnings cycle and stating that consensus capex estimates for 2027 are too conservative.
That is not a firm changing its mind. It is a firm being right about two entirely different layers of the same stack at the same time. James Covello, Goldman's head of global equity research, is right that hyperscalers are spending at a pace that has not yet produced demonstrable returns. Goldman's equity desk is right that the companies supplying that spending are printing money regardless of whether the returns ever arrive. The two positions are not in tension. They are a precise diagnosis of where in the AI economy value is concentrating and where it is not.
Everything else worth understanding about the AI capex boom follows from that diagnosis.
The Structural Trap the Coverage Keeps Missing
Amazon has committed approximately $200 billion in capital expenditure for 2026. Alphabet between $175 and $185 billion. Meta between $115 and $135 billion. Microsoft tracking toward $120 billion or more. Combined: $725 billion, up 77% from $410 billion last year. Goldman projects the annual figure reaching $1.6 trillion by 2031.
The standard explanation for why these numbers keep growing despite unproven enterprise returns is competitive anxiety, or as Goldman put it in May 2026, "FOMO has proven a stronger incentive than poor stock performance." That framing is accurate but it understates the structural mechanism at work.
Each hyperscaler knows that if it alone reduces capex while the others maintain it, it permanently cedes the data center capacity, model training infrastructure, and enterprise cloud positioning that determines competitive standing in the AI era. A company that underbuilds in 2026 cannot catch up in 2028 by spending more. The lead times on data center construction, power interconnects, and GPU procurement run two to four years. A capacity gap opened now stays open for the better part of a decade.
So the competitive logic facing each hyperscaler individually is: spend heavily and accept compressed near-term returns, or spend conservatively and accept permanent competitive disadvantage. In game theory this is a well-studied problem. When every player in a market faces the same asymmetric payoff structure, the outcome is collective overspending at levels that would not be rational for any single player acting without reference to what the others are doing. The hyperscalers are not irrationally fearful. They are rationally trapped.
Covello's prediction two years ago that stock underperformance would trigger spending discipline turned out to be wrong precisely because he was evaluating individual rationality in a situation where collective rationality is structurally unavailable. Microsoft, Amazon, Alphabet, and Meta cannot coordinate a spending reduction even if they collectively wanted one. Antitrust law prevents it. Competitive dynamics demand it.
The companies supplying them with hardware face no such constraint.
Nvidia's Position Requires Careful Description
In the most recent quarter, Nvidia posted $81.6 billion in revenue, $75.2 billion of which came from data centers, and $58.3 billion in net income. A 71% net profit margin. On an annualized basis, Nvidia made $194 billion in AI chip revenue in 2026.
Every description of this performance tends toward superlatives that obscure the structural explanation. The performance is not the result of superior execution alone, or marketing, or product quality, though all of those are real. It is the result of occupying the single supply-side chokepoint in a market where the buyers are structurally compelled to spend regardless of return, and where switching costs are high enough that the buyers cannot easily route around the chokepoint even when they try.
Nvidia holds approximately 86% of the AI chip market. The custom silicon programs at Google, Amazon, and Microsoft reduce their Nvidia dependency at the margins, but the foundational training and inference workloads that consume the most compute still run predominantly on Nvidia hardware. The CUDA software ecosystem that has accumulated over fifteen years of developer tooling, optimized libraries, and institutional knowledge creates a switching cost that hardware specifications alone cannot overcome. A competitor GPU that matches Nvidia's performance numbers still requires years of ecosystem development before it can absorb enterprise workloads that are already running.
What makes Nvidia's position genuinely durable rather than temporarily dominant is the compounding of the data flywheel. The hyperscalers buying Nvidia GPUs today are training models whose outputs will be used to design future GPU architectures. Nvidia's own AI chip design process runs on Nvidia hardware. The company is embedded in the improvement cycle of the technology it sells.
The Company in Veldhoven That Nobody Discusses
The most important company in the AI infrastructure buildout is not in Silicon Valley or Taiwan. It is in Veldhoven, a mid-sized city in the southern Netherlands, and it makes machines that most people outside the semiconductor industry have never heard of.
ASML manufactures extreme ultraviolet lithography systems, the machines that etch circuit patterns onto silicon wafers at node sizes below seven nanometers. EUV machines cost approximately $200 million each. The process of developing them took decades of physics research and required solving engineering problems at scales measured in atoms. No other company on the planet makes them.
Every leading-edge AI chip that matters, Nvidia's Hopper and Blackwell series, AMD's MI300, Google's TPUs, Apple's custom silicon, requires TSMC to manufacture it, and TSMC cannot manufacture at leading-edge nodes without ASML's EUV machines. The causal chain is direct: Goldman's forecast that AI capex grows to $1.6 trillion by 2031 is simultaneously a forecast about ASML's order book, because there is no path from current chip demand to future chip demand that does not run through Veldhoven.
This has not gone unnoticed by governments. The United States has pressured the Netherlands to restrict ASML's sales to China, and those restrictions are now in place. China is spending tens of billions of dollars attempting to develop domestic EUV capability, with limited progress. The entire AI infrastructure buildout that the hyperscalers are funding is, at its physical foundation, dependent on a single Dutch company that has become one of the highest-stakes geopolitical assets in the world.
When analysts describe the AI spending cycle in terms of cloud revenue and model performance benchmarks, they are describing the surface. At the base of the stack, the physical constraint on how much AI infrastructure can be built is the production capacity of a company in the Netherlands that can manufacture perhaps 60 EUV machines per year.
TSMC and the Economics of the Only Bridge
If ASML is the machine builder, TSMC is the factory. Taiwan Semiconductor Manufacturing Company builds virtually all of the chips that matter at leading-edge nodes: Nvidia's, AMD's, Apple's, and the custom AI silicon of the major hyperscalers.
The economics of leading-edge semiconductor fabrication are structurally monopolistic. Building a new fabrication facility at five-nanometer or below costs $20 to $30 billion and takes four to six years. The yield improvements that make production economical require years of accumulated process knowledge that cannot be transferred or replicated quickly. TSMC has spent decades optimizing its manufacturing processes in ways that competitors cannot close by simply building a facility and hiring engineers.
Intel's attempts to re-enter the foundry business at leading-edge nodes have been slower and more expensive than planned. Samsung's foundry business trails TSMC on yield and customer relationships. The CHIPS Act investments in US domestic semiconductor manufacturing will eventually produce capacity, but the timeline runs to the end of this decade at optimistic projections.
In the interim, the $725 billion in hyperscaler capex flows through TSMC's fabrication capacity before it reaches any data center. The bottleneck is not money. It is physical manufacturing throughput at a factory in Taiwan that is simultaneously a strategic national asset, a commercial monopoly, and the infrastructure substrate of the AI buildout.
Goldman's Two Correct Positions and the Question They Leave Open
Covello's core argument is that enterprise adoption of AI has not materialized at the scale or pace required to justify the infrastructure investment. Since ChatGPT launched, Nvidia's net income has grown approximately 20 times. The hyperscalers have seen far more modest returns relative to their capital deployment. Model companies have been losing money. Enterprises buying AI tools have not produced the measurable productivity gains that would make the investment self-justifying at an aggregate level.
His June 2026 assessment, "At some point you've got to make money," is directed at the layer of the stack above Nvidia: the companies spending hundreds of billions in anticipation of enterprise revenue that has not fully materialized.
Goldman's equity desk is making a different and compatible claim: the infrastructure has to be built regardless of whether the near-term returns materialize, the companies supplying the infrastructure are doing so from structurally protected positions, and the buyers cannot stop spending even if the returns disappoint. That is a supply-side argument, not a demand-side argument.
The open question that neither position resolves is the railroad problem. The US railroad buildout of the 1880s produced enormous overcapacity, mass corporate bankruptcies, and widespread investor losses. It also produced the physical infrastructure that enabled the American industrial economy for the following century. The railroads that went bankrupt were not wrong that rail infrastructure was valuable. They were wrong about the timeline and the competitive structure of who would capture that value once the infrastructure existed.
The hyperscalers are building the compute infrastructure of the AI economy. Whether they are the companies that ultimately capture the economic value of that infrastructure, or whether that value accrues to application-layer companies built on top of what they construct, is a question that the current capex numbers do not answer. The companies certain to profit are the ones supplying the construction regardless of who benefits from the buildings.
The Enterprise Buyer Who Has Not Yet Arrived
The hyperscaler capex plans are underwritten by an assumption: that enterprises will absorb the AI compute capacity being built, through cloud AI services, at prices high enough to justify the infrastructure investment.
That enterprise demand has been slower to materialize than the spending plans assume. Enterprise AI deployments are running into the same friction that Covello identifies at the macro level: unproven ROI, organizational barriers to adoption, data governance gaps, and the institutional inertia of large organizations changing how work gets done. The gap between AI capability and enterprise ability to extract value from that capability is real and has not closed at the pace the infrastructure investment requires.
This does not make the infrastructure investment wrong. It makes the timeline uncertain. The hyperscalers can sustain the spending through their existing cash flows for longer than most companies could, and the competitive structure discussed above means none of them will voluntarily slow down first. The question is whether enterprise adoption eventually justifies the build, not whether it does so on the original schedule.
For enterprises watching this from outside the spending circle, the practical implication is that the AI infrastructure being built is increasingly available, increasingly capable, and increasingly accessible through existing cloud relationships. The constraint on enterprise AI value creation is organizational and architectural, not computational. The compute is there or will be shortly. The clean data pipelines, governance frameworks, and institutional knowledge of where AI produces reliable outputs are what most enterprises are still missing.
The companies that figure out the organizational layer will extract enormous value from infrastructure that someone else paid to build. That is, historically, where application-layer value comes from in platform transition cycles. The hyperscalers building the railroad deserve credit for constructing it. The companies that figure out what to carry on it are the ones worth watching.
