Nvidia’s stock is one of the most-discussed investments on the planet. Hyperscalers are spending at a pace that dwarfs any prior technology buildout. Both of these things are broadly understood. And yet Goldman Sachs’ latest infrastructure forecast suggests that the most rewarding AI trade of the next five years may belong to companies most investors have barely glanced at. Both of those truths cannot be simultaneously priced in.
Market Context
The scale of what Goldman Sachs is projecting deserves careful framing before any thesis is built on top of it. The bank’s Global Institute and Global Investment Research estimate that the five largest hyperscalers — broadly understood to be Amazon, Microsoft, Google, Meta, and Apple, though the precise composition varies by analyst — collectively spent roughly $150 billion annually on capital expenditures in 2023. By 2026, Goldman estimates that figure has reached approximately $800 billion, a 94% increase in a single year. Projections for 2027 stand at $1.2 trillion, rising to $1.4 trillion in 2028.
The cumulative figure through 2031 — covering compute, data centers, and power infrastructure — lands at $7.6 trillion under Goldman’s baseline model, with annual AI-related capital expenditure reaching $1.6 trillion by the end of that window. To appreciate the magnitude: the additional $200 billion Goldman expects hyperscalers to spend in 2028 alone nearly matches what those same companies spent across the entirety of 2024.
This is not a technology spending cycle in the conventional sense. Conventional cycles peak, then correct. What Goldman is describing looks structurally different: a buildout of physical infrastructure at a scale more analogous to electrification or telecommunications network construction than to a typical enterprise software upgrade wave. The International Energy Agency corroborates part of this picture, projecting that global data-center electricity consumption will roughly double from 485 terawatt-hours in 2025 to approximately 950 TWh by 2030, with AI-focused data centers specifically expected to see electricity demand triple over that same period.
Those are not GPU-demand numbers. Those are grid-infrastructure numbers. And that distinction is central to understanding where this market is heading.
Competitive Landscape
The first chapter of the AI infrastructure trade was, by most accounts, dominated by semiconductor companies — most visibly Nvidia, whose H100 and subsequent-generation accelerators became the defining constraint of the AI buildout. That trade made intuitive sense: if everyone needs GPUs to train and serve AI models, buy the GPU maker. The logic worked. Nvidia’s market capitalisation surged accordingly.
But Goldman’s analysis, and the IEA’s supporting data, suggest the bottleneck is migrating. A GPU sitting unpowered in a warehouse, as the source framing puts it, produces exactly zero AI revenue. What turns chips into productive computing capacity is a layered stack of enabling infrastructure, and the companies positioned across that stack span several industries that have historically had little to do with artificial intelligence.
In memory, the principal players named in Goldman’s analysis include Micron Technology, SK Hynix, and Sandisk. AI accelerators require high-bandwidth memory at volumes that are structurally different from conventional data-centre workloads, and Goldman reports that supply constraints have already pushed memory producers’ gross margins to roughly 80% — more than twice their historical average.
In networking and optical connections, Broadcom, Arista Networks, and Coherent are the analysts’ most-cited names. As AI clusters scale to thousands of interconnected processors, the speed and efficiency of data movement between those processors becomes a genuine architectural constraint. Networking chips, optical transceivers, and high-speed switches are not glamorous, but they are non-negotiable.
In power and electrical infrastructure, Eaton, GE Vernova, and Vertiv occupy the positions of greatest current interest. Every new AI cluster requires transformers, power distribution units, cooling systems, and uninterruptible power supply infrastructure before a single inference can run. The IEA notes that supply chains for gas turbines and transformers have already tightened as data-centre projects expand — a supply-constraint signal that precedes, rather than follows, the demand surge Goldman is projecting.
Three Theses on This Market
Thesis One: The GPU Trade Has Further to Run
The conventional view holds that semiconductor companies remain the primary AI beneficiaries because model complexity continues to rise, inference at scale demands ever-more compute, and the most capable chips command the highest margins. Under this thesis, Nvidia and its supply chain remain the core AI trade, with infrastructure plays as secondary beneficiaries. Proponents point to the fact that each new model generation tends to require exponentially more compute, sustaining GPU demand irrespective of infrastructure scaling.
Thesis Two: The Infrastructure Layer Is the Durable Trade
Goldman Sachs’ analysis leans toward this thesis. The argument is that as GPU supply normalises — driven by expanding foundry capacity, competing chip architectures from AMD, Google, and custom silicon programmes at every hyperscaler — pricing power in compute hardware could erode faster than pricing power in the physical infrastructure required to operate it. Transformers, cooling systems, and high-bandwidth networking face supply chains measured in years, not quarters. Companies like the broader AI infrastructure investment thesis we’ve previously examined suggest the physical layer may offer more durable margin protection.
Thesis Three: The Energy Sector Is the Overlooked AI Trade
A smaller but growing cohort of analysts argue that neither chips nor immediate electrical infrastructure is the real constraint — it’s power generation itself. If data-centre electricity demand doubles by 2030 as the IEA projects, and if grid interconnection queues remain backlogged (a well-documented condition in most major markets), then the bottleneck that ultimately governs hyperscaler capex deployment is megawatts, not silicon. Under this thesis, power generators, grid operators, and long-duration energy storage companies capture the terminal value of the AI buildout, regardless of which chip architecture dominates.
Evidence For Each
The GPU-first thesis draws support from hyperscaler earnings commentary, where executives consistently describe compute availability — not power or networking — as the binding constraint on AI service expansion. This language has been consistent across multiple quarters.
The infrastructure-layer thesis gains credibility from Goldman’s own margin data. An 80% gross margin in memory — more than double the historical norm — is a signal that supply is genuinely constrained relative to demand, and that pricing power currently sits with suppliers, not buyers. The IEA’s observation that transformer and gas turbine supply chains are already tightening tells a similar story for electrical infrastructure. This dynamic echoes what chip-sector analysts have flagged regarding the risks of assuming compute hardware retains permanent pricing power.
The energy thesis finds support in the IEA’s projection that AI-specific data-centre power demand will triple by 2030, and in reports from grid operators in Virginia, Texas, and parts of Europe indicating that new data-centre interconnection requests are being queued for years. Trump’s characterisation of electricity as “the oil of the next 50 years” — whatever its political context — reflects a real structural shift in energy demand that utility planners are actively modelling.
Our Synthesis
What the source article and the Goldman data together suggest, but do not explicitly state, is that AI infrastructure spending is creating a sequential bottleneck migration rather than a stable hierarchy of winners. In prior technology buildouts — fibre-optic networks in the late 1990s, cloud infrastructure in the 2010s — each phase of the buildout solved one constraint only to surface the next one. Chips were the first AI bottleneck; memory and networking appear to be the current bottleneck as evidenced by margin expansion; power generation and grid interconnection are likely the next. Investors who treat this as a single static trade risk being perpetually one phase behind the actual constraint — buying memory stocks just as chip supply normalises, then rotating to power just as grid interconnection backlogs begin to clear. The more resilient analytical frame is to identify which bottleneck is forming, not which one has already formed.
Financial and Strategic Implications
For incumbents in electrical infrastructure — companies like Eaton and GE Vernova that have existed for decades before the word “hyperscaler” entered the vocabulary — the current environment represents an unusual pricing moment. These are businesses with established manufacturing bases, existing utility relationships, and products (transformers, switchgear, power distribution units) that require years to scale production. Their competitive moat in the near term is not technological; it is physical. You cannot 3D-print a large power transformer in a meaningful timeframe. That constraint is the source of their current pricing power, and it is not easily eroded by new entrants.
For challengers — smaller optical networking companies, specialised cooling-technology firms, and energy storage startups — the Goldman forecast represents a large addressable market, but also a capital-intensive one. Hyperscaler procurement teams are sophisticated buyers who prefer established vendors for mission-critical components. Breaking into that supply chain at scale requires either a differentiated technology or a willingness to accept the risk of building capacity ahead of contracted demand.
For investors, Goldman’s own caution is worth noting: the bank flags that slower capex growth will eventually intersect with rising depreciation expenses, which could reduce AI’s contribution to S&P 500 earnings even as absolute spending levels remain elevated. This is a structurally important hedge. $1.4 trillion of annual hyperscaler spending in 2028 does not automatically translate to $1.4 trillion of earnings-accretive activity. Depreciation clawback is a real drag, and it arrives on a fixed schedule regardless of how AI revenue scales. As we’ve previously noted in our analysis of frontier AI revenue projections, the gap between infrastructure spending and monetisation timelines remains a central risk across the sector.
The strategic implication for hyperscalers themselves is that their capital allocation decisions are now significant enough to move commodity markets, influence grid planning at the national level, and reshape manufacturing capacity in industries that have historically served the utility sector. That is a qualitatively different kind of market power than operating a cloud platform.
The Strongest Counterargument
The most credible objection to Goldman’s infrastructure thesis comes not from sceptics of AI broadly, but from analysts who argue that hyperscaler capex forecasts have historically been revised downward when economic conditions tighten or when the return on prior infrastructure investment disappoints. The pattern is well-documented: in the early 2000s, telecommunications companies made infrastructure commitments of comparable ambition relative to their balance sheets, and the resulting overcapacity took nearly a decade to absorb.
The counterargument, made explicitly by some technology equity analysts, is that hyperscalers are not committing to this spending because AI revenue is already scaling proportionately — they are committing because they fear competitive disadvantage if they do not. That distinction matters. Fear-of-missing-out-driven capex is not the same as demand-pull capex, and the former has historically been more prone to abrupt revision. If a single major hyperscaler — under pressure from shareholders during an earnings cycle when AI revenue disappoints — announces a capex reduction, the signalling effect on the infrastructure supply chain could be severe.
Does this weaken Goldman’s conclusion? Partly. It argues for greater caution around companies with concentrated hyperscaler exposure and limited revenue diversification. It does not undermine the structural observation that AI infrastructure requirements are real and large; it does counsel against assuming the Goldman trajectory is smooth or guaranteed. The infrastructure thesis is sound; the timing and linearity of the capex path is where the risk lives. This concern aligns with the broader risk landscape we’ve covered in AI market risk analysis — structural opportunity and structural risk are not mutually exclusive.
Risk Factors
Beyond the capex-revision risk described above, several other factors could materially alter the trajectory Goldman projects.
First, algorithmic efficiency gains. If model architectures continue to improve in compute efficiency — as they did meaningfully between GPT-3 and GPT-4 era models — the compute requirements per unit of AI output could fall faster than Goldman’s baseline assumes, compressing demand for infrastructure at the margin.
Second, geopolitical disruption. AI infrastructure supply chains run through Taiwan (advanced semiconductors), Japan and South Korea (memory), and multiple jurisdictions for rare earth materials. Any escalation in US-China trade tensions — already evidenced in recent chip export controls — could interrupt supply chains in ways that slow deployment even as demand remains strong. The competitive dynamics around China’s own AI chip development add a further variable to this calculus.
Third, regulatory and permitting risk. Data-centre construction and grid interconnection are subject to local permitting, environmental review, and utility commission approval processes that have proven difficult to accelerate. If interconnection queues lengthen further, hyperscaler spending commitments could outpace their ability to deploy capacity.
Fourth, concentration risk at the demand side. Five hyperscalers accounting for the overwhelming majority of $7.6 trillion in cumulative spending represents extraordinary customer concentration for any supplier. A shift in a single company’s procurement strategy — switching vendors, developing in-house alternatives, or reducing growth targets — can ripple through multiple tiers of the supply chain simultaneously.
Where This Ends Up
The most likely outcome is that Goldman’s broad directional thesis proves correct — AI infrastructure spending does reach into the trillions annually by the late 2020s — but the distribution of winners is messier than the current consensus assumes. Memory and power infrastructure companies capture outsized margins in the near term, attracting new capacity investment that gradually normalises those margins by 2028 or 2029. The pattern mirrors every prior infrastructure buildout: early constraint creates exceptional returns, exceptional returns attract capital, capital erodes the constraint and compresses the returns. The companies that sustain advantage will be those with either genuine technological differentiation or irreplaceable physical assets — long-lead-time transformer manufacturing capacity, for instance, or proprietary cooling architectures.
The second-most-likely outcome is a partial capex revision — not a collapse, but a meaningful deceleration — if AI revenue monetisation disappoints in 2026 or 2027. The condition that tips the balance toward this scenario is straightforward: if two or more major hyperscalers lower capex guidance in the same earnings cycle, the infrastructure supply chain would face simultaneous demand signal revision across its largest customers. That would create significant volatility in the very stocks that have re-rated upward on the Goldman thesis. The structural AI demand is real. The path from here to $1.4 trillion of annual spending is not guaranteed to be linear.
