Investing and economic exposure in an AI economy

Two separate exposures, not one worry

“Is AI a bubble, and am I exposed to it” is really two questions in one. The first is portfolio exposure: whether retirement accounts and index funds have grown unusually concentrated in a handful of AI-linked companies, and what happens if the spending underneath that concentration disappoints. The second is labor-market exposure — whether a household’s income, not just its savings, depends on work AI could reshape — which the household income planning and labor displacement pages cover in depth. This page addresses the first question: what credentialed voices on both sides of the “AI bubble” debate are saying, what the spending underneath the argument looks like in disclosed dollars, and how major asset managers currently frame concentration risk.

The case that this is a bubble

The skeptical camp includes people paid to be right. Investor Michael Burry disclosed large put positions against Nvidia and Palantir in his final 13F filing before deregistering Scion as an SEC adviser, and has since warned of a possible “1987-style” drop in AI stocks — though his largest recent figures come from his own newsletter, not further SEC filings. Apollo Global Management’s chief economist, Torsten Slok, calculates that AI-related data-center capital spending is rising from 1.4% of GDP in 2025 toward roughly 3% by 2027 — nearly twice the pace of the 2000s housing boom — and argues a cycle that builds this fast can unwind at a similar pace. GMO co-founder Jeremy Grantham calls the AI-driven market the most expensive in American history by market-cap-to-GDP and, applying his own “two-sigma” bubble framework, sees a plausible peak-to-trough decline of around 70%, with timing genuinely uncertain.

Institutional voices echo the concern with less theatrics. MIT economist Daron Acemoglu, a 2024 Nobel laureate, argues neither theory nor data support the most exuberant growth forecasts, modeling AI’s effect at roughly 1.1%-1.6% added to GDP over ten years, far below other estimates. The Bank for International Settlements, the umbrella body for the world’s central banks, found AI investment surging as a share of GDP and concluded financial-stability risk looks moderate only if AI firms hit substantial profit targets, while flagging a widening equity-debt valuation gap and growing reliance on debt to finance the buildout. Federal Reserve Governor Lisa Cook has warned that a sustained debt-financed boom could become a stability concern in its own right, while also crediting AI-driven growth with supporting stability on net. IMF Managing Director Kristalina Georgieva has told audiences to brace for turbulence, comparing today’s valuations to dot-com-era exuberance and warning a sharp correction could tighten financial conditions well beyond tech.

The case that it isn’t — or isn’t yet

The bullish and mixed positions come from some of the same research desks raising the alarms above, which is itself informative. Goldman Sachs Research has estimated that roughly $19 trillion of market value is already pricing in AI’s economic payoff ahead of where measured productivity gains have landed — a genuinely bearish-sounding number — while the same firm has separately argued AI capital spending remains a smaller share of U.S. GDP than prior technology cycles, and that most AI-spending companies, unlike their dot-com-era predecessors, still generate large free cash flow, buy back stock, and pay dividends. Even OpenAI’s own CEO holds a mixed position: Sam Altman has said investors are “overexcited” about AI and compared the moment to the dot-com bubble, while maintaining AI is among the most important developments in a very long time, and later predicted some investors will lose enormous sums even as others make fortunes. None of this reads as an industry closing ranks around one story — the disagreement runs through the same buildings, not just between them.

What the spending actually looks like

Whatever one concludes about valuations, the underlying spending is real, disclosed, and large. Microsoft reported fiscal Q4 2026 capex of $41 billion, roughly two-thirds of it short-lived CPUs and GPUs, with fiscal 2027 spending guided to keep growing. Meta’s Q2 2026 earnings call disclosed $31.1 billion in quarterly capex against free cash flow of just $784 million, with full-year guidance narrowed to $130-145 billion and no 2027 figure offered yet. Alphabet and Amazon each raised 2026 capex guidance more than once this year, per their Q2 2026 results and Amazon’s July call, and Oracle’s cloud backlog reached $638 billion alongside its own sharply higher capex.

Several headline mega-deal figures deserve caution: they are letters of intent and reported backlog, not settled contracts. Nvidia’s own announcement of “up to $100 billion” toward OpenAI was explicitly non-binding; by year-end, Nvidia’s own CFO said no definitive agreement existed, and later reporting found the deal that actually materialized came in well below that number. The similarly reported $300 billion, five-year OpenAI-Oracle Stargate agreement likewise traces to reported deal terms, not a fully itemized public contract. Economist Noah Smith has separately raised the question of circular financing — chip-industry investment into AI labs that returns as compute-purchase revenue — and while one estimate for such arrangements runs “north of $800 billion,” that figure traces to market-aggregator sites, not any named bank, economist, or regulator, and should be read as illustrative, not confirmed.

Concentration: why “the market” increasingly means a handful of firms

Portfolio exposure to AI is rarely a deliberate choice; it usually arrives through ordinary index exposure. The Magnificent Seven reached about 34.3% of S&P 500 market capitalization by December 2025, up from roughly 12.3% in 2015, and contributed about 42% of the index’s total return in 2025 alone — meaning a “diversified” S&P 500 index fund is, in practice, an increasingly concentrated AI bet. Why so few firms sit at the center of that bet — control of chips, data-center capacity, and frontier models — is covered separately in power concentration: why a few labs matter.

Major asset managers have started saying so explicitly, in similar language despite competing with each other. Vanguard distinguishes AI’s genuine economic upside — its own modeling gives real odds to a 3% GDP-growth scenario — from what it considers excessive tech-stock valuations, projecting muted 4-5% average returns for U.S. tech stocks over five to ten years even in a bullish scenario. BlackRock warns elevated concentration erodes the diversification benefit of index exposure, tethering portfolios to a narrow set of outcomes, and recommends blending AI-leader exposure with broader diversification. J.P. Morgan Asset Management puts it most directly: the structural case for AI remains intact, but index-level exposure concentrates risk, and recommends diversifying into AI-adjacent sectors, geographies, and private markets. None of the three calls for retreat from equities; all three call concentration itself a risk, independent of what AI delivers.

The paycheck side of the ledger

The other exposure — whether household income depends on AI-exposed work — rests on a thinner research base than the market-concentration debate above, and this manual would rather say so than borrow false precision from either side. The most-cited direct evidence is a Danish study by economists Anders Humlum and Emilie Vestergaard, which linked survey responses from roughly 25,000 workers in ChatGPT-exposed occupations to Denmark’s administrative earnings records and found precisely estimated null effects on earnings and hours worked, ruling out effects larger than about 2% in the two years after ChatGPT’s release. That is one dataset, one country, one window — a data point, not a verdict on how this settles elsewhere or over a longer horizon. For sharper, more recent U.S. hiring-pipeline evidence and concrete household planning steps, see household income planning for an uncertain labor market and labor displacement: what the data actually shows.

A framework, not a forecast

Line up the names above and a more useful pattern emerges than either side’s headline claim: this argument runs through the same institutions, not cleanly between them. Goldman Sachs published both the bubble case and its rebuttal; Federal Reserve governors credit AI with supporting growth in one sentence and flag it as a debt-financed stability risk in the next; OpenAI’s own CEO holds both positions about his own industry at once. That is not evidence the debate is unresolvable noise — it is evidence that no one, including those with the best access to the underlying data, currently has a validated model for how this resolves. A reasonable household response is not to bet on either side settling it, but to notice where concentration has crept into retirement accounts by default, and raise that with a licensed adviser — on its own terms, not a headline’s.