This article is for informational purposes only and is not intended as investment advice, an offer, or a recommendation to invest in any security. All investments carry risks, including the potential loss of principal.
The headlines about artificial intelligence focus almost entirely on the models — which one reasons best, which one writes code, which lab is ahead this quarter. Behind those headlines, a far larger and more durable story is unfolding: the physical buildout required to run AI at all. Data centers, accelerator chips, networking, cooling, and the electrical power to feed it represent the largest concentrated capital expenditure cycle the technology industry has ever undertaken. For investors, the infrastructure layer may be where some of the most tangible exposure to AI actually lives.
The Scale of the Spend
The numbers are difficult to absorb at first reading. The five largest hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — collectively guided toward roughly $700 billion in capital expenditure for 2026 on their AI data-center buildouts, a step-change increase over the prior year. Industry analysis estimates that roughly 75% of that aggregate spend — on the order of $450 billion — is directed specifically at AI infrastructure rather than traditional cloud.
Zoom out further and the figures compound. Goldman Sachs has modeled roughly $7.6 trillion of AI capital expenditure between 2026 and 2031 across compute, data centers, and power, with annual spend rising from an estimated $765 billion in 2026 toward $1.6 trillion by 2031. McKinsey has separately estimated that companies could invest nearly $7 trillion in global data-center infrastructure by 2030 — a figure roughly comparable to the combined GDP of Japan and Germany. The “$3 trillion” in this article’s title is, if anything, a conservative way to frame a multi-year buildout whose upper estimates run far higher.
Individual company commitments illustrate the concentration. OpenAI alone signaled intent to spend on the order of $1.4 trillion on data centers over time, anchored by the Stargate project’s 10-gigawatt ambition. This is not a normal capital cycle. It is a generational reallocation of capital toward a single technological bet.
The Layers of the Buildout
“AI infrastructure” is not one market. It is a stack of distinct layers, each with its own economics, competitive dynamics, and investability. Understanding the layers is the difference between a vague thematic bet and a considered allocation.
Accelerators and Silicon
At the base sits the accelerator — the processor purpose-built for the parallel computation AI workloads demand. Nvidia captures an estimated 90% of AI accelerator spend, a dominance with few parallels in industrial history. Around it sits an ecosystem: memory makers such as Micron, networking and interconnect providers, and a handful of challengers — including custom in-house silicon from the hyperscalers themselves and wafer-scale alternatives such as Cerebras — attempting to carve out share. This is the layer where the spending is most visible and the incumbent most entrenched.
Data Centers and the Neoclouds
Above the silicon sits the physical facility. Beyond the hyperscalers building their own, a new category of “neoclouds” — CoreWeave, Nebius, IREN, Lambda, and Crusoe among them — has emerged to build and rent GPU capacity at scale. U.S. data-center construction reached a monthly rate exceeding $45 billion by the end of 2025, up sharply from two years prior. Increasingly, hyperscalers are leasing rather than building, which shifts capital and risk toward this operator layer and creates investable businesses whose entire model is renting AI compute.
Power and the Grid
The most underappreciated bottleneck is electricity. AI data centers consume power on a scale that is straining grids, and Goldman Sachs has documented a U.S. data-center power-capacity shortfall exceeding 11 gigawatts, potentially widening toward 45 gigawatts by 2028. This has pulled an entirely separate set of sectors into the AI thesis: independent power producers, nuclear (including the revival of dormant plants), natural gas, transformers and grid equipment, and industrial cooling. For investors wary of buying chips at peak valuations, the power layer offers a less crowded form of exposure to the same demand.
How the Buildout Is Being Financed — and the Risk That Carries
Historically, hyperscalers funded capital expenditure from free cash flow. The scale of the AI buildout has changed that. Companies are increasingly turning to debt markets to bridge the gap between AI capex budgets and internal cash generation, and analysts at firms including Morgan Stanley and JP Morgan have suggested the technology sector may need to issue on the order of $1.5 trillion in new debt over the coming years to finance construction — alongside novel structures such as GPU-backed leasing and project finance.
This financing shift is the heart of the bear case, and investors should weigh it honestly. A buildout funded by cash flow can be paused if demand softens. A buildout funded increasingly by debt carries fixed obligations that persist whether or not the AI revenue materializes on schedule. The central open question is whether AI revenue will grow fast enough to justify infrastructure spending of this magnitude. If it does, the buildout is the foundation of a durable platform. If revenue disappoints while debt-funded capacity keeps coming online, the same infrastructure becomes a source of strain. In our view, that asymmetry — not the technology itself — is the risk that deserves the most scrutiny.
Frequently Asked Questions
How much is being spent on AI infrastructure?
The five largest hyperscalers guided toward roughly $700 billion in capital expenditure for 2026, with an estimated 75% directed at AI infrastructure. Longer-term models from Goldman Sachs project roughly $7.6 trillion of AI capital expenditure between 2026 and 2031, and McKinsey has estimated nearly $7 trillion in global data-center investment by 2030.
What are the main layers of AI infrastructure to invest in?
The buildout sorts into layers: accelerator silicon (dominated by Nvidia, plus memory and networking), data centers and the “neocloud” operators that rent GPU capacity, and the power and grid equipment needed to run it all. Each layer has distinct economics, and the power layer in particular has become a significant and somewhat less crowded form of exposure.
What is a neocloud?
A neocloud is a company that builds and rents out large-scale GPU computing capacity for AI workloads, distinct from the traditional hyperscalers. Examples frequently cited include CoreWeave, Nebius, IREN, Lambda, and Crusoe. As hyperscalers increasingly lease rather than build, these operators have become a notable part of the infrastructure economy.
What is the biggest risk to the AI infrastructure buildout?
A key risk is financing. Hyperscalers and others are increasingly funding the buildout with debt rather than cash flow, creating fixed obligations that persist regardless of whether AI revenue grows as projected. If revenue disappoints while debt-financed capacity continues coming online, the buildout could become a source of financial strain rather than a durable platform. Power-supply shortfalls are an additional constraint.
Positioning Around the Buildout
The AI infrastructure buildout is, in our view, one of the more concrete ways to think about exposure to artificial intelligence — it is steel, silicon, and electricity, with capital expenditure figures that are disclosed and measurable rather than speculative. It is also concentrated, increasingly debt-financed, and dependent on AI revenue arriving on a schedule that no one can guarantee. For accredited investors, the useful work is in deciding which layer of the stack to access, at what valuation, and how that exposure interacts with positions they may already hold in AI through public equities, private companies, or pre-IPO secondaries. At TSG Invest, we track the infrastructure layer alongside the broader AI sector and help investors think through where, and whether, this kind of exposure fits. For the wider picture, see our artificial intelligence sector page.
TSG Invest is a multi-entity financial services firm offering private wealth management, alternative investment access, and institutional-grade portfolio strategies for accredited investors. Securities offered through TSG Capital Advisors LLC. Advisory services offered through TSG Alpha Partners LLC, an SEC-registered investment adviser.
This content is for informational purposes only and does not constitute investment advice. All investing involves risk, including the potential loss of principal. Past performance does not guarantee future results.