AI Infrastructure Bubble: How to Test Whether Cloud Growth Is Sustainable
Microsoft, Google, and Amazon are spending huge sums on GPUs, data centers, and power. Yet a large share of their AI demand may come from only two customers: OpenAI and Anthropic. That raises a key question for investors watching the AI infrastructure bubble: are hyperscalers building for broad demand, or helping fund the same companies that drive their AI revenue?
The risks include cloud revenue concentration, circular funding, delayed data-center use, heavy losses, and slower productivity gains. The figures below come from analyst estimates, media reports, and comments from research firm CEO Ed Zitron. They point to risks, not proof that the AI sector will collapse.
Hyperscaler AI growth may be more concentrated than expected
Cloud companies often present AI demand as broad and fast-growing. But some estimates suggest that two loss-making model companies account for a large share of that growth.
Google Cloud’s exposure to OpenAI and Anthropic
UBS estimates cited in the interview suggest that OpenAI and Anthropic could account for 27% of Google Cloud revenue in the referenced year. That share could rise above 48% the following year. Zitron said the later figure would amount to more than $124 billion.
A customer share near 50% creates a serious concentration risk. If either company cuts spending, loses access to funding, or shifts its cloud strategy, Google Cloud could face a sharp revenue shock.
OpenAI is also a major Google Cloud customer, despite receiving less public attention than its Microsoft relationship. That matters because cloud revenue tied to an AI company buying compute is different from demand spread across thousands of profitable business customers.
AWS and Microsoft face related risks
Barclays estimates cited by Zitron put OpenAI and Anthropic at 13% of AWS revenue in the referenced year and 18% the following year. AWS is a much larger business than Google Cloud, so the percentage is lower. Still, dependence on two fast-growing but unprofitable customers can affect pricing, planning, and revenue quality.
Microsoft may have even greater exposure through its Intelligent Cloud segment. Zitron reported that 69% of the segment’s year-over-year growth in calendar year 2025 came from OpenAI. Without that contribution, he said growth would have been about 8%.
That figure refers to incremental growth, not total Microsoft revenue. It also may not capture indirect gains from Azure tools, software sales, or other parts of Microsoft’s AI business. Even so, it shows how one customer can make a segment’s growth rate look much stronger.
Circular financing may support the AI infrastructure bubble
The concern is not that cloud providers invest in customers. Strategic investments can create useful partnerships. The concern is whether AI demand can continue without repeated funding from those same partners and outside investors.
Cloud providers can act as investors and suppliers
Hyperscalers may fund AI companies through equity deals, cloud credits, and long-term contracts. Those AI companies then spend the money on cloud capacity, often from the same firms that invested in them.
That creates a revenue loop. A cloud provider records more sales, expands its data centers, and gains a higher valuation. The AI company receives more compute and reports faster growth, even if its own customer revenue cannot cover its bills.
This structure does not prove fraud or improper accounting. It does create a key investor question: who pays for the compute when new funding slows? Zitron compared the “smartest guys in the room” idea to past Enron concerns, but he did not claim the same conduct is happening today.
Google’s TPU deal shows how revenue can circulate
The interview described a Google TPU arrangement involving Broadcom and Anthropic. Broadcom supplies TPUs to Google, Google makes that capacity available to Anthropic, and Anthropic rents it through Google Cloud.
That setup can produce several layers of revenue across hardware sales and cloud rentals. Yet gross sales do not equal lasting profit if the end customer depends on fresh investment to keep buying capacity.
Investors need to separate hardware revenue, cloud rental income, investment gains, and revenue from paying end users. Those figures can look strong together while the underlying model remains dependent on outside capital.
OpenAI and Anthropic must grow fast to support new data centers
The scale of planned AI infrastructure creates a difficult math problem. Data centers require steady use for years, while the largest AI customers continue to post major losses.
Data-center plans require vast demand
Lightline Climate reportedly estimated that about 190 gigawatts of data-center capacity was built or under planning for the coming years. Data-center power use is often measured with power usage effectiveness, or PUE. A PUE of 1.3 means a facility uses 1.3 units of total power for every unit used by computing equipment.
Using that PUE and the capacity estimate, Zitron calculated that more than $1.6 trillion in annual revenue could be needed to support the planned infrastructure. This is a utilization argument, not proof that all 190 gigawatts will be built.
It also does not mean OpenAI and Anthropic must fund the entire amount. The point is that cloud providers need many paying customers, high utilization, and strong prices to earn acceptable returns.
Losses increase the need for new funding
Zitron said OpenAI lost $20.9 billion in 2025, based on financial reporting he reviewed. The company can grow revenue quickly and still need outside money for chips, data centers, employees, research, and model training.
He also said more than $800 million of OpenAI revenue came from SoftBank’s “Crystal Intelligence” program. The accounting treatment and details of that program matter when judging how much revenue came from normal customer demand.
The interview also cited a possible delay in an OpenAI IPO until 2027, based on reporting from The New York Times. An IPO could provide new capital, liquidity, and public valuation support. A delay could make it harder to fund large compute commitments, though a reported delay is not the same as a confirmed listing schedule.
Construction delays could expose weak demand
AI infrastructure faces limits beyond money. Power connections, permits, land, cooling systems, chips, and skilled labor can all slow construction.
Data centers take years to complete
The interview estimated that data centers can take 12 to 36 months to build, depending on size and complexity. That timing creates a gap between an announced project and the revenue it can produce.
If demand grows more slowly than expected, a completed facility may sit below capacity. Hyperscalers would still face interest costs, depreciation, power expenses, and maintenance bills.
The reverse problem also matters. If construction cannot keep pace, cloud providers may report strong orders without having enough finished capacity to recognize the full revenue.
Productivity gains may not cover AI spending
The best case for AI spending is higher productivity: workers produce more, businesses earn more, and customers pay for the added value. The weaker case is that AI tools remain costly, unreliable, or too hard to fit into daily work.
Lower enterprise adoption would reduce demand for premium models and cloud compute. It could also pressure the shares of Microsoft, Alphabet, Amazon, chip firms, data-center operators, and power suppliers.
The AI infrastructure bubble does not need to burst completely for investors to suffer. Slower growth, lower prices, or weaker returns could be enough to reduce valuations.
Investors need a demand-quality test
AI growth headlines reveal little about the strength of the underlying business. Investors should review customer concentration, cloud commitments, strategic investments, and revenue recognition details.
A useful test asks whether new customers are paying from recurring business revenue or from fresh equity and debt funding. It also asks how long contracts last, whether customers can cancel, and whether hyperscalers provide credits that reduce the customer’s real cost.
Capital spending deserves the same review. Compare AI investment with operating cash flow, free cash flow, depreciation, and expected returns. Then ask how much revenue each model company must generate to cover compute, power, staff, leases, and financing.
The core questions are practical:
- What utilization rate must each data center reach?
- Can the AI company stay solvent without another funding round?
- What happens if model prices fall faster than computing costs?
- Are customer savings measurable, or do projected returns depend on future productivity gains?
- Does demand extend beyond OpenAI and Anthropic?
Conclusion
The AI infrastructure bubble thesis rests on a clear concern: hyperscalers may be building enormous capacity for a small group of heavily funded, loss-making companies.
The interview cited UBS estimates of 27% and more than 48% Google Cloud exposure to OpenAI and Anthropic, Barclays estimates of 13% and 18% AWS exposure, and a reported 69% contribution from OpenAI to Microsoft Intelligent Cloud growth in 2025. It also cited The Information’s claim that 89% of revenue among the largest AI companies came from OpenAI and Anthropic, along with 190 gigawatts of planned data centers and a $20.9 billion OpenAI loss.
Those numbers are risk signals, not a guaranteed crash forecast. Sustainable AI growth requires diverse paying customers, positive unit economics, recurring cash flow, and productivity gains large enough to support the infrastructure now under construction. That is the standard investors should apply before treating AI spending as proof of lasting demand.
Post a Comment