Is the AI Boom a Bubble? A Developer's Read of the Numbers

Every few weeks a new headline asks whether AI is a bubble. The honest answer is the unsatisfying one: the technology is real, and parts of the financing may still be a bubble. Those two statements do not conflict, and a developer who keeps both in mind makes better decisions than one who believes only the hype or only the doom.

Here is what the evidence actually says, and where it is weak.

How much money are we talking about?#

Estimates for 2026 spending on AI infrastructure by the largest cloud companies vary widely, mostly because analysts count different companies. Published figures for the big hyperscalers' combined capital spending run from roughly $650 billion to over $800 billion for 2026, with some projections passing a trillion dollars for 2027.

Treat any single number with suspicion. What matters is the shape: spending is enormous, rising, and growing faster than the revenue it is supposed to produce.

The case that something is overheated#

Skeptics point to a few specific things rather than a general feeling:

• Cash flow strain. PIMCO has estimated that hyperscaler capital spending could consume around 94% of operating cash flow in 2026 and 2027. When a company spends nearly everything it earns on one bet, it has less room to be wrong.
• Debt and outside money. Morgan Stanley estimates that more than half of the roughly $2.9 trillion hyperscalers will spend on AI data centers between 2025 and 2028 will be financed with external capital, not from cash on hand.
• Circular demand. One research paper describes the risk when spending is justified by future demand from companies whose own revenue depends on AI spending elsewhere in the ecosystem. Money moving in a circle looks like growth until it stops.
• Depreciation. Investor Michael Burry argues that GPUs wear out or become obsolete faster than companies assume, which would overstate profits. This claim is contested, but it matters: if chips last three years and are accounted for over six, the numbers flatter.
• Weak spots are in the wrapper layer. A recent academic evaluation found the bubble evidence stronger in private AI valuations, application-layer software and speculative data-center projects than in profitable infrastructure leaders.

The case that it is real#

The bulls have facts too:

• Revenue is growing quickly. AI labs report revenue growth that few software companies have matched. Figures are inconsistent across sources and mostly self-reported, so do not quote a number, but the direction is not in dispute.
• Adoption is broad. One business-spending index reported US enterprise AI adoption passing 50% in early 2026.
• Physical constraints are real. Power, land and memory are scarce, which supports pricing for the companies that own capacity. A research paper framing the question as "boom, bubble, or buildout" argues scarce power and data-center capacity could produce durable returns.

The sobering data point#

An MIT NANDA report from August 2025 found that only about 5% of enterprise generative AI pilots reached rapid revenue growth, with most stalling. It was based on 150 interviews, a survey of 350 employees and 300 public deployments, a small sample, so the headline "95% fail" is far less solid than it sounds. But the finding that matters survives: buying tools from specialist vendors and partnering worked about twice as often as building internally. Value came from integration into real workflows, not from the model itself.

The dot-com lesson, used carefully#

The usual comparison is the dot-com era, and it is useful in one way: a bubble can be wrong about timing and prices while being right about the technology. The internet changed everything, and many of the companies that overbuilt fiber still went bankrupt. Both are true.

Applied to AI, the likely pattern is not "AI fails" but some mix of: infrastructure overbuilt in places, weak application companies wiped out, strong ones consolidating, and the capability itself staying and getting cheaper.

What this means if you write software#

• Don't tie your career to one vendor's survival. Prefer skills that transfer across models and providers: evaluation, retrieval, tool design, cost control.
• Learn the economics. A feature that costs more in inference than the customer pays is not a product. Knowing token costs, caching and model routing is a real advantage.
• Build on value, not novelty. The pilots that worked solved a specific workflow problem. That is a software engineering skill.
• Expect price swings. Compute and API prices have moved both down and up. Keep model choices behind an interface so you can switch.

The takeaway#

Nobody can say whether the AI market is in a bubble until after it ends, and anyone who sounds certain is selling something. What you can say is that capital is being committed faster than proven returns, that financing is leaning more on debt, and that the useful part, software that solves a real problem, will outlast any correction.

Sources#

• What must happen for AI's trillion-dollar gamble to pay off (MIT Technology Review)
• Boom, Bubble, or Buildout? (arXiv)
• MIT report: 95% of generative AI pilots at companies are failing (Fortune)