Moody's expects six major US companies to spend USD 785 billion on capital projects in 2026, driven by AI demand.
Delays, lease commitments, short chip life, and credit strain are the main pressure points, and weaker borrowers feel them first.
Backlog and demand look strong, but neither guarantees profit. A correction would likely slow projects, not stop AI use.
The AI boom is creating an unusual situation for Big Tech: business is growing, but investor confidence is becoming harder to secure. Amazon’s cloud unit grew 39% last quarter, its fastest pace in nearly five years. Amazon’s shares still fell more than 6% after earnings. Over 12 months, it spent USD 42.4 billion more cash than its operations brought in.
Meta saw the same pattern. Its shares fell 7% even though results beat estimates. Demand looks real. Investors now want clear proof of returns, not just bigger AI budgets.
AI tools run inside data centers, which are huge buildings packed with computers. AI data center investment is now running at record levels. Moody's expects six major US companies to spend USD 785 billion on capital projects in 2026, driven by AI demand. Its March forecast was USD 700 billion. Close to USD 1 trillion is expected in 2027.
A data center takes 12 to 24 months to move from first spending to first revenue, according to Moody's. Spending took about 60% of operating cash flow in 2025, J.P. Morgan estimates. The share is set to reach 90% to 100% in 2026. Operating cash flow is the cash a business earns from its normal work.
J.P. Morgan counts USD 17 billion of bond sales by these companies in 2024 and USD 109 billion in 2025. The first half of 2026 alone brought USD 194 billion.
| Pressure Point | What the Data Shows | Why it Matters |
|---|---|---|
| Construction delays | 57% of projects slipped at least three months in 2025. Grid connection can take four years in major markets. | Finished buildings can sit idle while costs run. |
| Lease commitments | Five companies held USD 969 billion in future data center leases at the end of 2025. About USD 662 billion had not started. | Payment obligations build before facilities earn revenue. |
| Short chip life | Moody's puts the useful life of AI chips at four to six years. | Equipment may age before it repays its cost. |
| Debt risk | S&P cut Oracle to BBB-, one notch above junk status. | Weaker borrowers have less room for error. |
The unstarted leases alone equal about 113% of those companies' adjusted debt. A Moody's analyst stressed that this is not an accounting trick. The companies have simply not yet received the services that trigger the liability. Lease costs reach the balance sheet once each lease starts.
Short chip life adds a twist. Companies now sign shorter leases with renewal options. Each renewal becomes a much bigger investment decision, made when AI plans may have changed. Most of these firms are strong. J.P. Morgan does not expect a credit crisis. The risk sits with weaker borrowers.
Demand is the strongest argument for the builders. Moody's says AI demand is well above supply. Amazon reports an AWS backlog of USD 578 billion, which is contracted work not yet delivered. It calls that backlog well diversified. Amazon also projects AWS growth above 40% in both 2027 and 2028.
Backlog needs a careful reading. It counts contracted business, not revenue earned. Amazon gave no margins, backlog timeline, or payback figures. A Jefferies analyst says cloud firms cannot turn orders into revenue as fast as investors want, pointing to building and supply delays.
Customer mix also matters. S&P estimates that roughly half of Oracle's USD 638 billion backlog is linked to OpenAI. Some demand also circles back. Cloud companies invest in AI labs that spend heavily on computing power from the same companies. The purchases are real. The cycle depends on funding that keeps flowing.
In a December outlook, Bank of America's strategist said this is not 2000. She added that AI will not keep leading the market without a pause. Think of turbulence on a flight. The plane stays up. Passengers feel the drop.
A correction would show up as slower construction, postponed projects, lower valuations, and tighter financing. AI use would continue, but heavily indebted firms would feel the strain first, while stronger firms could afford to wait.
The dot-com era built fiber networks on overly optimistic growth forecasts, and although the internet did grow, many builders failed after overspending.
Also Read: How Much Water Does AI Use? Why Data Centers Tell the Bigger Story
The evidence supports concern about returns, financing, and capacity. It does not prove that an AI bubble is about to burst. AI may not have a technology problem. It may have a timing problem. Bills arrive years before the revenue does. How well each company manages that gap will decide who profits from AI.
Also Read: BlackRock CEO Warns AI Data Center Halt Could Trigger Recession
Some companies already build in stages. Moody's notes that firms are pacing data center builds and chip orders to limit overbuild risk. That approach may become particularly important as companies finalize their 2027 budgets. Amazon has said its own spending will rise meaningfully next year. The market will watch whether caution or ambition earns the better reward.
1. How much are big tech companies spending on AI data centers?
Moody's expects six major US companies to spend USD 785 billion on capital projects in 2026. Its March forecast was USD 700 billion. Spending is expected to approach USD 1 trillion in 2027.
2. Why are investors worried when AI demand looks strong?
Spending is taking most of the cash these companies earn. J.P. Morgan estimates it will take 90% to 100% of operating cash flow in 2026, up from about 60% in 2025. Amazon's and Meta's share drops after earnings show that investors want proof of returns.
3. What are data center lease commitments?
These are rent agreements signed for facilities that are often still being built. Moody's found USD 969 billion in future leases across five companies at the end of 2025, with about USD 662 billion not yet started. They are real obligations, not an accounting trick.
4. Is this an AI bubble?
The evidence supports concern about returns, financing and capacity. It does not prove a crash is near. The dot-com era is the closest comparison. The internet did grow, but many builders overspent on overly optimistic forecasts.
5. Which companies face the most risk?
Heavily indebted firms have the least room for error. S&P cut Oracle to BBB-, one notch above junk status. Stronger firms can wait longer, and J.P. Morgan does not expect a credit crisis.