Every market commentator wants a simple answer. Is AI a bubble or is it a boom? They want a yes or no verdict to print in headlines. But framing the entire AI transition as a single binary question is a fundamental mistake.
The academic literature and the market data point to a segmented reality. The risk profiles of semiconductors, cloud infrastructure, data centers, foundation models, and application software are completely different. You cannot apply one broad verdict to the entire stack.
The right way to analyze the AI story is to split it by layer. Infrastructure is fundamentals-driven and largely self-funded. The equity valuation layer and the labs themselves carry real, unresolved risk.
You can be right that this is not a systemic credit crisis like 2008. At the exact same time, you can be right that specific parts of the tech stack are heavily overheated. Both positions are true.
The Bull Case: Real Numbers, Not Vibes
The bull case for AI infrastructure does not rely on empty promises. It relies on massive, cash-generative balance sheets. Unlike the dot-com era, the companies building the physical foundations of AI are among the most profitable corporations in human history.
Most AI capital expenditure is funded directly by free cash flow, not by speculative debt. The Magnificent Seven hold fortress balance sheets. Their return on equity sits near 46 percent. They trade at roughly half the earnings multiples that Cisco and Intel hit during the peak of the dot-com bubble.
Contrast the profitability of today's leaders with the darlings of the late 1990s. Nvidia's projected FY2026 revenue is $215.94 billion, representing a 65 percent increase year over year. This is not Pets.com selling dog food at a loss. This is an industrial monopoly selling high-margin infrastructure to cash-rich buyers.
- Mag 7 capital expenditure is funded by free cash flow, not leverage.
- Nvidia's $215.94 billion revenue target is backed by real purchase orders.
- Hyperscalers are reporting genuine capacity constraints, not oversupply.
Furthermore, the capacity constraints are real. Microsoft and Amazon Web Services report actual shortages of GPUs and compute capacity. Compare this to the 1990s fiber boom, where telecom companies laid millions of miles of fiber optic cables that sat at less than 0.002 percent utilization for years. Today, every GPU that is shipped is plugged in and run at capacity immediately.
The Bear Case: Steelmanning the Risk
To understand the risk, we must look below the hyperscalers. The bear case is strong when you focus on valuation multiples and rising leverage.
Several AI leaders trade at price to sales ratios above 30. Historically, paying 30 times sales is a reliable indicator of low long term returns and impending corrections.
Debt is also creeping into the ecosystem right below the top tier. Oracle raised $18 billion in debt in late 2025. CoreWeave is highly leveraged, raising billions in debt backed by its GPU inventory while simultaneously taking equity investments from Nvidia. The argument that AI is entirely self-funded is true at the top of the stack, but it breaks down as you move down the food chain.
"We are building massive infrastructure based on a promise of utility. If the utility doesn't materialize at the app layer, the infrastructure will eventually write down." — Yann LeCun, Chief AI Scientist at Meta
There is also a circular loop in AI financing. Nvidia participated in OpenAI's $110 billion capital raise by taking a stake worth up to $30 billion. At the same time, OpenAI is Nvidia's primary chip customer. OpenAI itself is reportedly not projected to be cash-flow positive until 2030.
This is the ultimate caveat. The hyperscalers are solid, but a massive portion of the demand curve traces back to foundation model labs that are not profitable today.
Labs vs. Dot-Com Companies
There is a structural difference between 2000 and today. In the dot-com era, capital went to companies built on top of the internet. Web portals, online grocery stores, and early e-commerce startups had short shelf lives because they did not own the underlying technology. Once the infrastructure existed, those early companies were disposable.
Today, capital is going directly into the labs building the core technology itself. OpenAI, Anthropic, and Google DeepMind are not simple application builders. They are creating the equivalent of the internet protocol. That IP does not have a disposability problem.
However, this does not make the labs immune. AI researchers like Margaret Mitchell have noted that the cost of training frontier models is growing exponentially, while the marginal utility of data is hitting diminishing returns.
Recession, Not Crash
If a correction occurs, it is highly likely to look like a standard market contraction rather than a systemic financial collapse.
The 2008 crisis was built on household leverage, subprime mortgages, and interconnected bank balance sheets. When the housing market cracked, the entire consumer economy collapsed. Today, the leverage is concentrated on corporate balance sheets and equity markets.
A correction in AI-heavy indices is not a credit crisis. A large market capitalization wipeout would still have real macroeconomic effects, including reduced capital expenditure, wealth effects, and tech sector layoffs. But the scale of a correction can still be massive. During the dot-com crash, the NASDAQ fell 80 percent, the S&P 500 fell 50 percent, and roughly $6 trillion in wealth was erased, all without triggering a systemic banking collapse.
The Falsifiable Marker
Instead of watching stock prices or press releases, watch one metric. Watch enterprise application-layer revenue growth.
Real annual recurring revenue expansion at the application layer, driven by actual corporate budget allocations, is the only metric that cannot be distorted by circular financing.
If enterprise software buyers show that LLM integration increases productivity enough to justify the seat licenses, the infrastructure layer will remain solvent. If application-layer revenue stalls, the demand loop will break. That is the metric that will tell you which way this breaks.
Sources
- Goldman Sachs: Generative AI Capex and Return on Investment Report
- Sequoia Capital: The $600 Billion AI Question
- Yahoo Finance: Nvidia and Oracle FY2026 financial reporting
- Yann LeCun: Public statements on AI utility and infrastructure scaling
- Bloomberg: CoreWeave debt structure and OpenAI circular funding analysis