The biggest investment opportunity in artificial intelligence isn't happening where most people think it is. While OpenAI and Anthropic captured $217 billion in venture funding in the first half of 2026, the five largest hyperscalers committed roughly $700 billion to $725 billion in capital expenditure to build the infrastructure these models actually run on. That second number reveals where the real money, contracts, and hard assets are forming in the AI economy.

Why Is Infrastructure Investment Growing Faster Than Model Investment?

The model layer has become crowded and expensive. Artificial intelligence captured 86% of U.S. venture dollars in the first half of 2026, but deals worth $100 million or more accounted for 87.5% of all capital deployed, according to the PitchBook-NVCA Venture Monitor. For many investors, the biggest AI names have become difficult entry points rather than open opportunities. As capital concentrates around a small group of model companies, investors and hyperscalers are looking beyond the crowded model layer for opportunities that capture AI growth without paying the same valuation premiums.

This shift is creating a new investment landscape. The AI infrastructure stack is producing new investment opportunities layer by layer, each addressing a critical bottleneck in the build-out. Understanding where hyperscalers are directing their capital reveals which infrastructure categories will likely produce the next generation of billion-dollar companies.

What Are the Critical Infrastructure Layers Attracting Investment?

Power has become the gating constraint of the entire AI infrastructure build-out. In November 2025, Microsoft CEO Satya Nadella said the company has AI graphics processing units (GPUs) sitting idle because it doesn't have enough power to install them. That shortage has accelerated investment in nuclear energy at an unprecedented scale. Thirteen announced projects have committed more than 9.8 gigawatts of nuclear capacity to AI data centers, and every major U.S. hyperscaler has signed at least one nuclear power deal. Microsoft backed a $16 billion restart of Three Mile Island, while Meta secured 6.6 gigawatts of capacity across TerraPower, Oklo, Vistra, and Constellation.

Beyond power, several other infrastructure layers are attracting significant capital:

• Thermal Management: Liquid cooling has roughly doubled in market size in 2025 to nearly $3 billion, heading toward $7 billion by 2029, as chip power density pushes past what air cooling can handle. Cooling startup Frore Systems closed a $143 million Series D at a $1.64 billion valuation, signaling investor confidence in this layer.
• Networking Infrastructure: Networking companies are addressing the growing need to connect large-scale AI clusters, with Upscale AI reaching a $2 billion valuation as investors bet that cluster networking will become a critical infrastructure layer.
• Inference Platforms: These platforms solve the challenge of putting AI models into production. Baseten raised a $1.5 billion Series F as enterprises move from AI experimentation to real-world deployment.
• Security for AI Agents: Security is becoming another essential layer, with the top 10 agentic AI security startups raising a combined $3.6 billion to protect software agents that barely existed two years ago.

How Are Hyperscalers Securing Long-Term Infrastructure Commitments?

The strongest evidence that infrastructure demand is real is the depth of commitments already in place. North American data center vacancy has remained at a record-low 1% for a second consecutive year, while 92% of capacity under construction is precommitted through binding leases or owner-occupied development. CBRE expects preleasing to remain in the mid-70% range in 2026, compared with a historical norm of 40% to 50%, indicating that most planned capacity already has an identified user before construction finishes.

Commitments also extend deep into power procurement. Microsoft's 20-year agreement with Constellation supports the 835-megawatt restart of Three Mile Island Unit 1. Meta has assembled a broader nuclear portfolio with 20-year agreements with Constellation and Vistra, alongside development commitments with TerraPower and Oklo, together supporting approximately 6.6 gigawatts of existing and planned nuclear capacity. For investors, this means much of the current build-out is supported by visible commitments rather than demand projections alone.

What Risks Could Derail the Infrastructure Build-Out?

The same capital fueling AI infrastructure growth could also create the market's next major risk: building capacity faster than demand can absorb it. Allianz Research estimates the gap between AI capital spending and revenue growth has reached roughly 46%, already wider than the 32% gap seen during the 2001 telecom cycle. The concern is not that AI revenue is absent. OpenAI is generating $2 billion in revenue per month in 2026, and Anthropic surpassed a $47 billion revenue run rate at the end of May 2026. The question is whether revenue can grow quickly enough to justify the scale of infrastructure being deployed.

How the build-out is financed presents another risk. One company's debt has grown to more than $21 billion, up from under $8 billion in 2024, and that debt is secured by GPUs whose useful life at this scale has not been fully tested. A different company, which sells those same GPUs, holds a $6.3 billion contractual backstop to buy the other company's unsold capacity. Timing adds a separate financing risk: even with contracted demand, permitting, construction, and grid-connection delays can defer revenue while interest and other carrying costs continue to accrue.

How to Evaluate Which Infrastructure Investments Will Survive a Downturn

• Demand That Can't Be Cut: Focus on categories where the demand itself is inelastic and survives downturns. A running GPU generates heat that has to be removed, a data center draws power that has to be supplied, and an AI agent handles data that regulation requires you to secure. That demand holds at any model valuation, while demand funded by experimentation budgets and unproven AI use cases disappears first in a downturn.
• Verifiable Metrics Before and After Launch: Judge each company on numbers that can be verified. Before delivery, examine preleasing rates, binding contracts, customer concentration, and counterparty credit quality. After facilities open, utilization and renewals show whether projected demand materialized. A sustained decline in preleasing, followed by weak utilization, would indicate that construction is beginning to outrun customer demand.
• Debt Structures Aligned With Revenue Timelines: Read debt structures alongside construction schedules to understand when debt service begins and whether contracted revenue arrives on the same timeline. Strong customers reduce some risk, but the financing terms determine who absorbs the cost of construction or grid-connection delays if a project opens late or a customer walks away.

The next bubble and the next generation of durable companies are forming in the same place. The investor's job this cycle is telling them apart. As the AI economy matures, the companies that survive won't necessarily be the ones building the flashiest models. They'll be the ones solving the unglamorous but essential problems of power, cooling, and connectivity that make those models possible.