I've spent the last several months tracking where the most disciplined pools of capital are flowing in the AI ecosystem, and what I've found isn't what most people expect. The popular narrative focuses on model performance, chatbot features, and which foundation model will "win." But the smartest institutional money isn't debating Claude vs. GPT-4o. It's buying land in Texas, signing 20-year nuclear power contracts, and quietly acquiring transformer manufacturers.
We are living through the most concentrated infrastructure investment cycle in the history of computing, and the window to position intelligently before 2027 is closing faster than most investors realize.
The Scale That Should Stop You Cold
Let me begin with a number that I genuinely believe most people haven't fully processed: $725 billion. That is the confirmed hyperscaler capital expenditure figure for 2026 alone, a figure derived from the Q1 2026 earnings calls of the five largest US cloud and AI infrastructure providers: Microsoft, Alphabet, Amazon, Meta, and Oracle. According to Dell'Oro Group, global data center capital expenditure surged 57% in 2025, with the top four US cloud service providers, Amazon, Google, Meta, and Microsoft, increasing their individual data center capex by 76% that same year. And analysts now project that full-year 2026 data center capex will surpass $1 trillion for the first time in history.
To zoom out further: Goldman Sachs projects approximately $7.6 trillion in AI infrastructure capital expenditure between 2026 and 2031 across compute, data centers, and power. McKinsey's baseline model puts the total build-out cost to meet projected AI demand by 2030 at $5.2 trillion, with an accelerated scenario reaching $7.9 trillion. These are not venture-stage estimates. These are projections backed by signed contracts, shovel-ready construction, and utility agreements.
This is not a software cycle. It is an industrialization cycle. And understanding that distinction is the first thing any serious investor needs to internalize.
Layer One: The Compute Stack, And Why the Real Story Isn't Nvidia
Yes, Nvidia dominates. Its Q1 FY2027 data center revenue hit $81.6 billion, up 85% year over year, and the company accounts for roughly 75% of total AI compute spend, according to Goldman Sachs assumptions. Nvidia's Blackwell GPU architecture has become the de facto standard for frontier AI training, and its GB200 racks are already being deployed in Stargate's flagship Abilene, Texas campus.
But here's what I think sophisticated capital is actually noticing: only about 25% of hyperscaler capex flows to chips. The other 75%, roughly $540 billion in 2026 alone, goes to the physical skeleton that surrounds those chips: data centers, power systems, cooling equipment, networking hardware, fiber, and land.
The GPU is the brain. The data center is the body. And the body, as any infrastructure investor will tell you, costs orders of magnitude more to build.
Smart capital is rotating, not away from semiconductors, but toward the overlooked physical stack underneath them. And that stack has five distinct layers, each representing a concentrated capital opportunity through 2027.
Layer Two: Data Center Real Estate, The Modern Oil Field
I think of today's hyperscale data center campuses the way an earlier generation of investors might have thought about oil fields: expensive to build, slow to develop, impossible to replicate once established, and structurally advantaged by scarcity.
BloombergNEF reported that over 23 gigawatts of data center capacity was under active construction globally as of late 2025, with roughly three-quarters of that in the United States. US data center construction spending reached a monthly rate of $45.1 billion by December 2025, up 85% from two years prior. Between August 2025 and February 2026, analyst expectations for 2027 spending by the 14 largest public data center developers climbed by 56%.
The US will account for approximately 60% of incremental global demand through 2030, with data center power capacity projected to expand 197% between now and then, reaching 95 gigawatts. What does this mean in practice? It means that anyone with existing land near reliable power infrastructure, in the right geographies, is sitting on what amounts to digital-era mineral rights.
Colocation REITs, infrastructure funds, and private credit vehicles financing this build-out are, in my view, among the most rational long-duration plays in the current environment.
Layer Three: Power Infrastructure, The Binding Constraint
Here is where the investment thesis becomes genuinely structural: power is the binding constraint of the entire AI build-out, and it will remain so through at least 2030.
A single H100 GPU draws 700 watts continuously. A cluster of 100,000 H100s, a scale that multiple companies now operate, draws 70 megawatts. Data centers running inference at scale require hundreds of megawatts. Microsoft's partnership with OpenAI alone requires the equivalent of a medium-sized power station running nonstop.
The International Energy Agency projects global electricity demand from data centers will more than double by 2030, reaching approximately 945 terawatt-hours. In the United States, data centers currently represent about 6% of total electricity demand, a figure set to nearly double to 11% by 2030. The US power grid was not built for this. In many regions, utility interconnection queues have backlogs measured in years, not months.
This is why the smartest hyperscalers are no longer waiting for grid access. They are building their own power generation capacity, a shift that represents one of the most consequential structural changes in the energy sector in decades. By early 2026, roughly 30% of new data center capacity is being designed to operate at least partially independent of grid infrastructure, up from near zero a year earlier.
For investors, this means the AI power trade is not about "data center stocks" in isolation. It's about utilities with generation near load centers, grid software companies, transformer manufacturers, power delivery specialists, and cooling system engineers. Companies like Eaton Corporation, Vertiv, Schneider Electric, Quanta Services, and Constellation Energy are finding themselves in unexpected demand as the physical requirements of AI compute crystallize.
Layer Four: The Nuclear Bet, Where the Biggest Commitments Are Being Made
The nuclear-for-AI thesis has moved from theoretical to executed. By May 2026, every major US hyperscaler had signed at least one nuclear power deal for AI data center capacity, a total of 13 signed deals across the sector committing more than 9.7 gigawatts of nuclear capacity.
The specifics are worth dwelling on:
Microsoft signed a 20-year contract with Constellation Energy for 835 MW from Three Mile Island's Unit 1 restart, projected to come online in 2028, backed by a $16 billion commitment.
Meta issued an RFP for 1–4 GW of new nuclear capacity and has signed agreements totaling up to 6.6 GW across TerraPower's Natrium reactor, Oklo's Aurora reactor, Vistra, and Constellation.
Amazon invested $700 million in X-energy for up to 12 Xe-100 small modular reactors and is converting the Susquehanna nuclear plant area into a $20+ billion AI campus.
Google signed what appears to be the first US corporate agreement for a fleet of small modular reactors through Kairos Power, committing to 500 MW of KP-FHR capacity, with the first reactor targeted for 2030.
Microsoft has surpassed Amazon as the world's largest corporate clean energy buyer, with roughly 40 GW contracted as of late 2025. Together, Amazon, Meta, Google, and Microsoft accounted for approximately half of all global clean energy purchase agreements signed in 2025.
This is not corporate sustainability theater. This is infrastructure survival. The first nuclear electrons dedicated to an AI data center will arrive in 2027 from the Crane Clean Energy Center (the Three Mile Island restart), a date that I believe marks a new chapter in the AI infrastructure story, one where compute power is no longer borrowed from legacy grids but purpose-built at the source.
The Stargate Phenomenon, A Project Without Historical Precedent
No analysis of the AI infrastructure gold rush would be complete without confronting the Stargate Project, the single largest infrastructure commitment in the history of computing.
Announced in January 2025 by OpenAI, SoftBank, Oracle, and MGX, Stargate is a four-year, $500 billion initiative to build AI data center capacity specifically for OpenAI's operations. The initial $100 billion was deployed immediately, and construction at the flagship campus in Abilene, Texas, 10 buildings, each spanning 500,000 square feet, is already underway. An additional 4.5 GW of capacity is being developed with Oracle, bringing total planned Stargate capacity to nearly 7 GW, en route to a 10 GW target.
Stargate's supply chain implications ripple through the entire ecosystem. According to one analysis, Stargate controls roughly 40% of global DRAM production and will account for approximately 20% of US power grid expansion investment through its build-out. For cloud developers relying on commodity compute access, the concentration of capacity in Stargate creates a structural dependency that will shape pricing and availability for at least the next five years.
The project has also not been without turbulence, structural disagreements between OpenAI, Oracle, and SoftBank over responsibilities and financing have required renegotiation. But the physical infrastructure is being built regardless, and the momentum is unmistakable.
Layer Five: Cooling and Networking, The Unglamorous Fortunes
I want to highlight two infrastructure categories that I believe remain systematically underappreciated by generalist investors: liquid cooling systems and AI networking hardware.
Modern AI training clusters generate extraordinary heat densities that air cooling simply cannot manage at scale. Direct-to-chip liquid cooling, immersion cooling, and rear-door heat exchangers are now engineering requirements rather than options. Companies like Vertiv, Schneider Electric, and a new cohort of specialized thermal management startups are benefiting from demand that is both structural and sticky, once a data center is designed around a specific cooling architecture, it rarely switches.
Similarly, AI compute clusters require network bandwidth that is orders of magnitude beyond what traditional data center networking provides. The interconnect between GPU nodes, what allows a 100,000-chip cluster to function as a coherent training system, is a specialized, high-margin engineering challenge. Fiber optic transceivers, optical modules, and high-speed InfiniBand and Ethernet switching equipment have entered extended lead time territory. Optical component manufacturers and network equipment providers serving this market represent a capital allocation that is, in my view, both rational and durable.
The Global Competition for AI Infrastructure Supremacy
What makes this gold rush particularly consequential, and what I believe separates it from prior technology investment cycles, is its explicitly geopolitical character.
The five largest US cloud providers have committed $660–690 billion in capital expenditure for 2026 alone. But the rest of the world is not standing still. Saudi Arabia announced more than $15 billion in AI investments at LEAP 2025, including a $10 billion partnership between PIF and Google Cloud. The UAE is developing what it describes as the largest AI campus outside the United States, a 26 square kilometer facility in Abu Dhabi with 5 GW of planned capacity under its HUMAIN initiative.
The European Union has unveiled a €200 billion AI Continent Action Plan, with €50 billion in public funding and €150 billion from private sources, establishing 13 AI Factories across 17 member states.
The geopolitical framing matters for capital allocation in a specific way: it means governments are willing to subsidize, accelerate permitting, and provide offtake guarantees to attract AI infrastructure investment. For infrastructure funds and private credit vehicles, this creates a risk-adjusted return environment that hasn't existed in the technology sector before.
The Honest Risks: Are We Building a Bubble?
I would be doing a disservice if I didn't engage seriously with the bear case. Three of the four major hyperscalers lost market value following their most recent earnings calls, with investors expressing concern about the ratio of capital commitment to demonstrable revenue returns.
Harvard economist Jason Furman's observation that data center investment effectively drove all US GDP growth in the first half of 2025 is remarkable, but it also raises the question of what happens if AI revenue growth disappoints relative to infrastructure cost. Goldman Sachs has flagged that consensus capex estimates have been too low for two consecutive years, but has also noted that "the timing of an eventual slowdown in capex growth poses a risk to these companies' valuations."
There are also real physical constraints: US grid interconnection backlogs, shortages of high-voltage power transformers, and a skilled labor shortage in data center construction are genuine impediments that capital alone cannot immediately resolve.
My view is that these are cycle risks, not thesis risks. The underlying demand for AI compute is structural and accelerating. The infrastructure gap is real. But investors who conflate the infrastructure investment cycle with the software monetization cycle, and expect them to peak simultaneously, will face timing disappointments.
Where I See the Smartest Capital Concentrating Before 2027
Having mapped this landscape as carefully as I can, here is where I believe the most disciplined capital is actually concentrating:
Power infrastructure adjacent to existing transmission: Land and interconnection agreements near load centers are scarcer than GPUs, and harder to replicate once committed.
Nuclear energy development: Both large-scale reactor restarts and small modular reactor developers are attracting the longest-duration capital in the space, with 20-year PPAs providing revenue visibility that almost no other technology investment offers.
Private credit in data center construction: With hyperscalers issuing debt to fund their build-outs, and construction timelines stretching across multiple years, secured lending against hard infrastructure assets offers risk-adjusted returns without full equity exposure to AI monetization uncertainty.
Thermal management and optical networking: These "picks and shovels" plays within the physical stack are supply-constrained, margin-rich, and strategically necessary for every player in the ecosystem.
International AI campuses in markets with government support: The Gulf states and select European markets with public co-investment offer returns that are partially de-risked by sovereign capital.
The gold rush metaphor is apt in one important respect: the people who got reliably wealthy in 1849 California weren't always the miners. They were the people who sold the shovels, the denim, and the picks. Before 2027, I believe the same principle applies to AI, and the physical infrastructure undergirding this revolution is the most durable, structural version of the "shovel" business I've seen in my lifetime.
