Oracle's September 10, 2026 earnings report captures the tension behind the AI bubble debate. Quarterly revenue rose 30% to $19.3 billion, yet the company reported negative free cash flow of roughly $5.4 billion. Sales are growing. Building the capacity to serve them is consuming even more cash.
That is a more useful starting point than declaring that AI has already collapsed, or assuming that every dollar spent on it will pay off. The question is whether the companies financing the infrastructure can turn demand into durable returns, and how much of the value ultimately stays with their customers.
The evidence now demands a more careful argument than the one that followed DeepSeek's arrival in early 2025. Chip suppliers are reporting substantial profits, cloud businesses are expanding, and the spending commitments remain enormous. Those facts can coexist with overinvestment. They do not establish that every company is in a bubble, or that a crash is inevitable.
Oracle shows why revenue and cash are different tests
In its first-quarter fiscal 2027 results, Oracle reported cloud infrastructure revenue of $7.4 billion, up 121% from a year earlier. Its cash flow statement recorded $23.1 billion of operating cash flow and $28.5 billion of capital expenditures. Subtracting those expenditures leaves the $5.4 billion free cash flow deficit.
Oracle also reported $664 billion in remaining performance obligations, a measure of contracted business still to be recognized as revenue. That is evidence of future business, but it is not $664 billion of cash already earned. Delivery schedules, customer payments and the cost of supplying the contracted services still matter.
A negative quarter does not prove that the investment is wasteful. Construction and equipment purchases can precede years of customer payments. The risk is a mismatch: spending is committed today, while the returns depend on capacity being delivered, used and paid for on acceptable terms.
The revenue behind the boom is real
Nvidia's August 26 results make it difficult to describe the entire industry as an expensive experiment without a business model. For the quarter ended July 26, 2026, it reported $96.2 billion in revenue, including $89 billion in Data Center revenue, and $63.7 billion in GAAP operating income.
That establishes that a supplier can earn very substantial returns from the buildout. It does not establish the eventual return for every company buying its systems. The same hardware purchase is revenue for the vendor and a capital commitment for the customer. Adding those figures together as though they measure separate end-user demand would exaggerate the market.
Microsoft offers a different perspective. Its fiscal 2026 fourth-quarter earnings call reported 43% growth in Azure and other cloud services. Capital expenditures were $41 billion, including finance leases, while cash paid for property and equipment was $35.8 billion. Operating cash flow of $55.4 billion left $19.6 billion in free cash flow.
These are company and cloud-business figures, not an audited profit statement for generative AI alone. Microsoft has established businesses supporting its investment. Still, positive cash generation alongside heavy spending is important counterevidence to the idea that all major participants are financing growth in the same way.
DeepSeek's $5.6 million figure needs its missing context
DeepSeek remains relevant because it challenged assumptions about the resources needed to build competitive models. But its widely repeated price tag has a precise scope.
The DeepSeek-V3 technical report estimated $5.576 million for the model's official training, using an assumed rental price of $2 per H800 GPU-hour. The total included pre-training, context extension and post-training. The authors explicitly excluded earlier research and experiments on architectures, algorithms and data.
It was not a statement that the company, its infrastructure and its entire research program cost less than $6 million. Nor was it a full development-cost disclosure for the separate DeepSeek-R1 reasoning model. The report describes training V3 on a cluster of 2,048 Nvidia H800 GPUs. Efficiency did not eliminate the need for substantial hardware.
The economic implication cuts both ways. Cheaper computation can pressure the prices model providers charge. It can also make new uses affordable and increase the volume of work people run. Whether efficiency reduces total infrastructure spending depends on how far usage expands as costs fall.
The trillion-dollar numbers do not all measure the same thing
Much of the bubble debate becomes less reliable when unlike figures are placed side by side. A forecast of global spending, a company's capital budget, a supplier's revenue and a startup's valuation each answer a different question.
For example, Gartner's March 2025 forecast of $644 billion in generative AI spending included services, software, devices and servers. Almost $398.3 billion of the forecast was devices. Treating the total as money spent solely to build models or train chatbots changes what the estimate means.
McKinsey's $6.7 trillion data center estimate is a projection of cumulative worldwide investment through 2030, covering AI and traditional computing demand. It is neither one year's spending nor money already spent. The scale is significant, but comparing it directly with a single year's narrowly defined AI revenue cannot establish a rate of return.
The same care applies to customer counts. Weekly active users and paid business seats are different populations. Dividing one by the other does not produce a valid conversion rate for an entire product, especially when consumer subscriptions and other sources of revenue are outside the comparison.
Who captures the value if intelligence gets cheaper?
A useful technology can create more value for its users than its suppliers are able to charge for. If several models can complete the same task adequately, a customer may gain negotiating power or switch to a cheaper provider. A business using AI to reduce repetitive work could benefit even while a model developer struggles to protect its margins.
There are reasons that outcome would not be uniform. Customers may pay for reliability, security, integration and support as well as raw model capability. Existing software relationships can make a service easier to sell and harder to replace. A benchmark score alone does not capture those advantages.
Infrastructure owners face another trade-off. Long-lived sites and power connections can support successive generations of equipment, while chips may need replacing much sooner. Microsoft said roughly two-thirds of its latest quarterly capital expenditures went to short-lived assets, primarily CPUs and GPUs. A busy facility still needs to earn enough to cover operating costs and future equipment replacement.
That is why a growing market does not automatically make every participant a good business. Timing, purchase price, financing and the ability to retain customers can matter as much as the popularity of the technology.
What would make the bubble argument stronger?
The most useful warning signs would connect demand to cash. Revenue growth slowing while committed spending remains high would be more revealing than another large investment announcement. So would customers reducing usage after discounts expire, or providers needing repeated new financing to support services that do not generate enough cash.
The opposite evidence matters too. Sustained customer spending, improving margins and cash generation after equipment purchases would strengthen the case that a buildout is earning its keep. Contract backlogs help describe demand, but their conversion into delivered services and collected payments is the test.
The current evidence supports neither a blanket declaration that AI is a sham nor confidence that every investment will work. Some companies are already earning large profits; others are spending ahead of the cash their infrastructure produces. Those are different positions with different risks.
AI could become cheaper, more widely used and more economically useful while parts of the investment boom disappoint. The technology does not have to fail for a business to overpay for its future. The question worth following is who can keep earning after the construction bills, computing costs and competitive price cuts arrive.
