- Steve Chiavarone, chief investment officer for global equities at Federated Hermes, argues investors worried about an AI bubble may be right but several years early, since the infrastructure buildout needs another three or four years before a hangover becomes the relevant question.
- His central argument is about financing discipline. Markets are being discerning about which companies receive capital and at what rates, which he calls the opposite of bubble behaviour, where everybody gets financed at the cheapest rate. The past fortnight supports him: Firmus had an IPO refused twice, Elice fell about 30% on its Korean debut, AirTrunk postponed a listing and CoreWeave’s swaps price better-than-even default odds, while TSMC guided above expectations.
- His weaker claim is that most capital spending is funded from free cash flow with debt playing a limited role. Goldman Sachs counts roughly $578 billion of AI-related debt issued globally this year, with hyperscalers accounting for only 40%, leaving about $347 billion from borrowers without comparable balance sheets. He does acknowledge over-levered firms are coming under stress.
- He sees commoditization risk for model companies including Anthropic and OpenAI, in contrast with larger technology firms he says are using AI to build bigger moats around existing businesses while earning returns on the investment.
What Happened?
His comments follow Ken Griffin’s warning at a New York conference on Tuesday that investors are almost certainly in an AI bubble, while stressing the technology itself is real. Chiavarone compared the cycle to the US highway programme and post-war European reconstruction, expecting the same pattern of job creation during construction, a subsequent lull, then displacement of work before new industries emerge, perhaps a little more intensely. On politics he dismissed the midterms as a risk, calling them the most meaningless election in US history because no configuration of results would produce a filibuster-proof Senate majority or veto-proof Congress capable of changing national AI policy, and arguing that by 2029 most data centre infrastructure will already be built.
Why It Matters?
The financing discrimination argument is the strongest available case against the bubble framing and it is empirically testable rather than rhetorical. In a genuine bubble, capital is indiscriminate: weak and strong borrowers finance on similar terms because nobody is distinguishing between them. What has happened instead is that investors refused a $30 billion valuation twice, punished a data centre listing by 30% on day one, pushed one borrower’s credit default swaps past 60% implied default odds, and simultaneously funded a chipmaker’s record capital programme and a listing with half the book pre-committed. That is a market allocating capital by quality, which is the behaviour Chiavarone describes. His political argument is the one to push back on. He is correct that no plausible midterm result produces the majorities needed to legislate national AI policy, but federal legislation has not been the binding constraint. San Francisco approved a data centre moratorium, environmental groups defeated Oracle’s New Mexico permits to the point of a force majeure claim, roughly $68 billion of projects were blocked or delayed in a single quarter, and a Kansas developer is paying households $1,000 a year for a decade to secure consent. None of that required Congress. Describing the election as meaningless to the AI trade overlooks where the opposition has actually been effective. The historical analogy is useful for the labour question running alongside this. If the pattern follows highways and post-war reconstruction, construction employment comes first, then a lull, then displacement, then new industries, which is a longer and more disruptive sequence than either the optimists or pessimists usually describe.
What Next?
Whether capital allocation stays discriminating is the test of Chiavarone’s thesis, and the next financings are Nscale’s listing and Blue Owl’s $6 billion loan for Stack. Watch the proportion of capital expenditure funded by debt rather than cash flow, since that is where his argument is weakest. Meta, Tesla and Netflix report this week. For the political question, local permitting outcomes matter more than the November 3 result on the evidence so far.
Affected Tickers and Coins: SPY, NVDA
Source: Bloomberg
