Is AI a Bubble? Four Likely Paths From Late 2026 Into 2027

Published September 13, 2026

Few questions matter more to investors, workers, and household budgets right now than whether the artificial intelligence boom is a durable technology shift, a financial bubble, or some combination of both. This post looks at the numbers behind that debate and lays out the most likely directions the industry could take between now and the end of 2027.

This post is for educational purposes only. It is not investment, tax, or financial advice.

The bull case: spending at a scale rarely seen outside wartime

The five largest Western hyperscalers — Amazon, Alphabet, Meta, Microsoft, and Oracle — have announced hundreds of billions of dollars in infrastructure spending. Estimates vary with the date and the spending included: Futurum put their 2026 plans at $660–690 billion in February, while tooldirectory.ai’s May compilation estimated $725 billion, compared with roughly $410 billion in 2025.¹² Fortune reports JPMorgan’s 2027 forecasts of over $1.1 trillion in capex and $900 billion in operating cash flow.³ Gartner’s May forecast puts broader global AI spending at approximately $2.6 trillion in 2026 and $3.5 trillion in 2027. That includes categories beyond hyperscaler capital spending and should not be added to it.⁴

Supporters of this spending point to revenue growth in Microsoft’s Azure and Google Cloud, as reported in company earnings discussions.⁵⁶ Research reviewed by Stanford SIEPR also finds productivity gains in some tasks, although results vary across workers and settings.⁷ The optimistic interpretation is that infrastructure built today could eventually be absorbed by demand that takes years to develop. That is an investment thesis, not a result established by spending announcements.

The bear case: a widening revenue gap and circular financing

The skeptical case rests on three questions: whether revenue can justify the investment, how financing connects the companies, and how much debt supports the buildout.

1. The revenue gap. In June 2024, Sequoia Capital partner David Cahn estimated that the AI ecosystem would need roughly $600 billion in annual revenue to support the infrastructure economics in his model. His calculation started with Nvidia’s projected revenue and made assumptions about other data-center costs and customers’ gross margins. It was a revenue requirement under those assumptions, not a measured $600 billion loss or a current 2026 shortfall.⁸ The underlying question still matters: will customers generate enough additional revenue to pay for the infrastructure and earn a return?

2. Circular financing. Bloomberg maps investment and purchasing relationships among Microsoft, OpenAI, Nvidia, and other AI companies.⁹ One directly documented example is the November 2025 Microsoft–Nvidia–Anthropic partnership: Nvidia and Microsoft committed to invest up to $10 billion and $5 billion respectively in Anthropic, while Anthropic committed to purchase $30 billion of Azure computing capacity.¹⁰ These are announced commitments, not proof that all the money has already changed hands. Our concern is that overlapping investor, supplier, and customer relationships can make firms more exposed to one another’s financial difficulties.

3. Debt is entering the picture. Federal Reserve Governor Lisa Cook said in May 2026 that hyperscalers were issuing bonds to fund AI investment and smaller data-center developers were borrowing through private debt funds and asset-backed credit markets. She warned that a sustained debt boom could eventually raise financial-stability concerns, while noting that even ambitious projections did not imply a return to leverage levels before the Global Financial Crisis.¹¹ Our interpretation is that debt can spread the consequences of disappointing infrastructure returns beyond shareholders to lenders and other credit investors.

What the return-on-investment data actually shows

Away from the infrastructure layer, enterprise adoption data is more sobering than the capex numbers suggest. Project NANDA’s preliminary 2025 report described 95% of organizations as seeing no measurable return from generative AI. Its research combined a review of more than 300 publicly disclosed initiatives, interviews at 52 organizations, and responses from 153 senior leaders. These were different evidence sources, not a representative survey of 300 deployments, and the finding should not be treated as a universal failure rate.¹² Separately, RAND’s 2024 report cited estimates that more than 80% of AI projects fail, twice the rate for other IT projects. RAND interviewed 65 practitioners to investigate causes; it did not establish that percentage through a new census of projects.¹³

IBM’s discussion of AI returns emphasizes business objectives, governance, workflow changes, and organizational readiness.¹⁴ Drawing on that guidance and the implementation research, our suggested checklist is to tie AI to a specific revenue or cost line, define success before scaling, give the teams using it a role in deployment, and measure the full cost of adoption. Those questions are useful for a nonprofit or small business as well as a large enterprise.

The labor market: a narrower effect than the headlines suggest, concentrated at the bottom

The evidence does not show widespread, economy-wide job displacement, but it does suggest a harder entry point into some careers. In its August 12, 2026 revision, a Stanford Digital Economy Lab working paper found that employment among workers aged 22–25 in AI-exposed occupations was 19% below the level implied by the employment trend of less-exposed peers, using payroll data through June. The gap mainly reflected reduced hiring. The authors describe these as early, descriptive indicators rather than causal estimates: the patterns weaken with some controls and are stronger in their payroll sample than in national survey benchmarks.¹⁵ Stanford SIEPR’s broader review likewise cautions against attributing the entire difficult graduate labor market to AI.⁷

In practical terms, our reading is that career entry deserves attention even without evidence of mass unemployment. Learning to use relevant AI tools may be useful, but these studies do not establish that AI fluency guarantees a job or a wage premium.

Four likely paths from here into 2027

Putting the infrastructure, adoption, and labor data together, four broad scenarios seem most plausible. These are our scenarios, not forecasts issued by the cited researchers, and we have not assigned statistical probabilities. They are not mutually exclusive — elements of more than one could unfold in different parts of the industry at once.

1. Slow-motion correction, not a crash (currently the modal case). Spending growth decelerates in 2027 as a handful of high-profile projects get delayed, renegotiated, or scaled back, and one or two heavily leveraged infrastructure players face real financial distress. Public markets reprice AI-exposed stocks meaningfully lower without a systemic crisis, because the largest hyperscalers have diversified, profitable core businesses that can absorb write-downs. This looks less like the 2000 dot-com crash and more like a sharp, contained correction.

2. Financing-led shock. If bond yields stay elevated and one large circular-financing counterparty misses payments or needs a bailout-style restructuring, stress could spread faster than expected through the private credit and asset-backed securities tied to data centers. This scenario extends the debt-financing risk Cook described; it is our interpretation, not a specific Federal Reserve forecast.¹¹ Depending on who holds the debt, stress could reach broader credit markets and institutions such as insurers and pension funds.

3. Bifurcation — infrastructure keeps growing, but “AI-native” startups thin out. The picture over the next 18 months could split cleanly: hyperscaler capex and chip demand continue growing (because the underlying compute is useful for far more than one company’s chatbot), while a wave of thinly capitalized AI application startups run out of runway once venture funding gets more selective. In this scenario the “bubble” pops mainly at the applications layer, not the infrastructure layer.

4. Demand catches up faster than expected. A genuine breakthrough in enterprise deployment — driven by better tooling, agentic workflows that actually complete multi-step tasks reliably, or a drop in inference costs — closes some of the revenue gap faster than bears expect. One reason to consider it is the large decline in inference costs at a fixed capability level documented in Stanford’s 2025 AI Index.¹⁶ Whether that translates into profitable enterprise use remains an open question.

Our base case

Our judgment, based on the evidence reviewed here, favors a combination of scenarios 1 and 3: continued heavy infrastructure investment by well-capitalized hyperscalers, a thinning-out of undercapitalized AI startups and some financially stressed infrastructure vendors, and a market correction in AI-exposed equities that is significant but contained — unless the financing-shock scenario (2) is triggered by a specific credit event, in which case the effects would spread further and faster than most investors currently expect.

For individual investors, the practical takeaway is to examine concentration risk rather than try to time a particular scenario. The supplier, customer, and financing connections described above suggest that owning several AI-linked companies may still leave a portfolio exposed to common risks. That is our inference from the documented relationships, not a measurement of any particular portfolio’s risk.¹⁰¹¹

Sources