Trillions are pouring into AI infrastructure, from data centers to GPUs, but questions around overbuilding, capital efficiency, and deal circularity are becoming harder to ignore.

Artificial intelligence is no longer just a software story. It’s an infrastructure arms race.

Over the next several years, trillions of dollars are expected to pour into GPUs, data centers, and the energy systems needed to run them. The scale is staggering – on par with the early internet buildout, and in some ways closer to post-war industrial mobilization. Amazon, Microsoft, and Google are committing record capital expenditures, making clear that AI is not a cyclical bet. It’s a structural one.

But beneath the momentum, two questions are starting to matter more: Are we overbuilding? And how real is the demand underpinning this boom?

We’ve been here before. Infrastructure cycles, from railroads to telecom, tend to overshoot. Capital floods in, supply races ahead of demand, and the unwind can be brutal.

The late-1990s telecom bubble is one obvious parallel. Equipment vendors effectively financed their own growth, extending credit to carriers while booking revenue upfront. The result was a mirage of demand. When usage failed to catch up, overcapacity crushed pricing, balance sheets broke and the sector collapsed.

AI, at least today, looks different.

Data center vacancy rates remain near historic lows, around 6–7% globally. New capacity is being absorbed as quickly as it comes online, often pre-leased before completion. Pricing remains firm. Most importantly, demand is visible: enterprises are deploying AI into workflows, developers are building on top of models and usage continues to scale.

This is not speculative capacity waiting for a use case. It is capacity chasing real consumption.

That doesn’t mean the risk is gone. It just means it’s delayed. Infrastructure cycles rarely stay in balance for long, and even a modest slowdown in demand could tip the system into overcapacity. The key signals to watch – GPU utilization, vacancy rates, and enterprise spend – will determine how long this alignment holds.

The New “Circularity” Isn’t What It Seems

A second concern is more subtle: the growing web of financial and strategic relationships tying the AI ecosystem together.

On the surface, deals between cloud providers, chipmakers, and AI labs can look circular, with capital flowing in loops and companies investing in each other while transacting heavily across the same relationships. That’s a red flag in most cycles.

But here, the structure matters.

Take Microsoft and OpenAI, or Amazon and Anthropic. These partnerships combine equity investment with long-term compute commitments and usage credits. While money moves both ways, revenue is ultimately tied to real usage. Credits function as incentives, not artificial demand since they only convert into revenue when compute is actually consumed.

The strategy is straightforward. Cloud providers are willing to trade near-term margins for long-term dominance, locking in demand and positioning themselves as the default infrastructure layer. In return, AI labs gain access to capacity, predictable pricing, and speed.

This isn’t financial engineering. It’s platform economics.

The same logic applies elsewhere. AMD is using partnerships with AI labs to accelerate software optimization and validation. NVIDIA is structuring deals where equity is swapped for future hardware demand, effectively capturing upside while reinforcing its position at the center of the ecosystem.

Still, not all deals are equally clean. Some arrangements, like NVIDIA’s financing relationship with CoreWeave, come closer to traditional circularity, where lending supports hardware purchases that in turn drive revenue back to the lender.

These aren’t systemic risks – yet. But they do complicate the picture. In a world where capital and revenue are increasingly intertwined, headline growth can overstate underlying demand. Free cash flow – not revenue – will be the more reliable signal.

Governments Are Raising the Floor

There’s one more factor that makes this cycle different: governments.

AI is no longer just a commercial opportunity. It’s a strategic priority. The U.S. CHIPS and Science Act, China’s national AI initiatives, and similar efforts across Europe, India, and the Middle East all point to the same conclusion: control over AI infrastructure is now tied to economic and national security.

That creates a floor under demand. Even if private investment slows, public spending is unlikely to. Countries are not going to underinvest in the technologies that will define competitiveness over the next decade.

The AI buildout is not a bubble in the traditional sense. Demand is real. Utilization is strong. The strategic rationale is clear.

But that doesn’t make it immune to the laws of capital cycles.

Overbuilding is still the most likely endgame – it’s just a question of timing. And while most deal structures are economically sound, the increasing interdependence of major players adds a layer of complexity that investors can’t ignore.

The next phase of AI won’t be defined by whether the technology delivers. It will be defined by whether the industry can scale it without repeating the excesses of the past.

History suggests that’s a high bar.

Contributing author: Kanishk Malhotra