The $1 Trillion Question: Is the AI Boom Built on Circular Deals or Real Demand?

What Circular Deals Actually Mean

A circular deal happens when the same money and capacity move through interconnected partners. One company finances another’s expansion, which then uses that money to buy the first company’s products.

In AI, these cycles can include equity stakes, credit lines, or “capacity backstops,” where vendors commit to buying unsold inventory from one another. A well-known example is the CoreWeave–Nvidia agreement, in which Nvidia reportedly has access to unused GPU capacity in CoreWeave’s data centers, effectively guaranteeing utilization even if demand softens (Financial Times).

Circular deals aren’t inherently bad. They can align incentives and make it easier to finance large projects. But if too much growth relies on these internal loops, real-world demand may be overstated.

Why Everyone Is Playing the Same Game

The AI boom has pushed every player to secure as much compute as possible. Supply chains for advanced chips remain constrained, so large buyers like OpenAI, Anthropic, and Meta are signing long-term contracts to lock in capacity years in advance.

OpenAI’s diversification strategy makes sense on paper. Nvidia remains the dominant supplier, but competition is intensifying. AMD’s Instinct GPUs and Broadcom’s custom chips offer alternatives that could cut costs and improve efficiency. By spreading orders, OpenAI reduces risk, gains leverage in pricing, and ensures redundancy.

For suppliers, these deals are irresistible. They guarantee revenue streams for years and attract investor confidence. AMD’s stock, for example, surged following its OpenAI announcement, and Broadcom’s market cap hit record highs after news of its custom chip partnership.

The result is a feedback loop: AI labs buy chips to train larger models, chipmakers expand factories to meet that demand, and capital markets reward everyone involved. But as analysts have warned, such loops can amplify both growth and fragility.

Where the Risks Begin

1. Overstated Demand

If AI adoption in businesses and consumers slows, the enormous capacity coming online may exceed actual usage. With data centers costing roughly $50 billion per gigawatt to build and equip (Financial Times), even a small gap between demand and supply can create massive financial stress.

2. Vendor Financing and Hidden Exposure

When suppliers help fund their own customers — or take equity in them — the line between investment and sales blurs. Reported revenue may rise, but underlying cash flow could remain weak. Analysts at Morgan Stanley and Morningstar have both flagged this issue as a sign of “circular financing” risk in the current AI market.

3. Contractual Backstops

Many infrastructure contracts include take-or-pay clauses, meaning buyers must pay for capacity whether or not they use it. These backstops keep utilization high but can strain liquidity if AI demand softens.

4. Limited Transparency

Companies rarely disclose how much of their reported backlog or revenue comes from related parties. Without that detail, investors can’t tell how dependent growth is on internal financing.

The Case for Optimism

To be fair, not all this investment is speculative. There are strong structural reasons to believe AI infrastructure will be in high demand over the next decade.

• Hardware Customization

By working with Broadcom to design its own chips, OpenAI is following a path pioneered by Google’s Tensor Processing Units. Custom silicon tailored for specific models can drastically improve performance per watt, reducing operational costs over time.

• Multi-vendor Resilience

Depending solely on Nvidia exposes any AI company to supply bottlenecks and pricing power. Partnerships with AMD and Broadcom not only expand supply but encourage faster innovation across the ecosystem.

• Long-term Payoff

Training frontier models like GPT-5 and GPT-6 requires exponentially more compute. If AI continues to move into enterprise workflows, real-time search, and robotics, today’s over-building may look like early preparation rather than excess.

Even Nvidia’s CEO Jensen Huang argued recently that the AI economy is driven by real demand, not hype, noting that every major industry is now integrating AI into production and customer interfaces (Windows Central).

• What to Watch Next

If you want to separate hype from genuine progress, track the following signals over the next 12 to 18 months:

• Utilization rates of large GPU clusters

• Revenue vs. capital expenditure growth for chipmakers and AI labs

• Disclosure of related-party transactions in quarterly filings

• Energy and power constraints in data center construction

• Real enterprise adoption metrics (AI revenue contribution by sector)

If utilization and revenue rise in step with infrastructure growth, the circular economy label may prove overly cautious. If not, we may be in for a correction similar to what the telecom sector saw in the early 2000s.

• The Bottom Line

The AI infrastructure race is both a marvel and a warning. It showcases unprecedented coordination between hardware makers, model developers, and cloud providers — but also introduces a layer of financial complexity rarely seen outside bubbles.

Circular deals have allowed the industry to move at breakneck speed, ensuring compute availability and accelerating product cycles. Yet they also mean that parts of this trillion-dollar machine depend on money flowing in circles.

The test will come when external demand, from businesses, governments, and consumer, has to justify the capital that has already been spent. If AI tools deliver transformative value across industries, the loop holds. If not, some of those circles may start to break.

And for companies building practical AI applications instead of megascale infrastructure, the lesson is clear: agility matters more than excess. Platforms like Beam AI are showing that intelligent automation doesn’t require trillion-dollar capacity, just smarter deployment of the compute we already have.