I’ve spent about a decade watching capital cycles, and I don’t think “bubble or not?” is the right question about AI infrastructure spending. It fits in a headline, which is why everyone asks it, but the honest answer is more complicated, more interesting, and more useful to your portfolio than a yes or a no. So here’s the version I’d want handed to me before sizing a trade.
Start with the scale, because most people haven’t absorbed it
The four biggest spenders — Amazon, Alphabet, Microsoft and Meta — are on track to spend roughly $725 billion on capital expenditure in 2026. That’s up about 77% from last year’s record $410 billion. By company: Amazon near $200 billion, Microsoft around $190 billion, Alphabet $175–185 billion, Meta $125–145 billion after it raised guidance at Q1. Analysts already have combined spending north of $1 trillion penciled in for 2027.
That’s more than the annual GDP of most mid-sized countries, spent in one year, by four companies, largely on GPUs and the buildings and power to run them. Microsoft’s CFO has said about $25 billion of its own capex increase came purely from higher memory-chip prices. One encouraging data point in the pile: Alphabet’s cloud backlog has roughly doubled, to around $460 billion.
The bull case, stated fairly
Before the scary part, the bull case deserves a real hearing, because a lot of bubble commentary skips straight to the fear without engaging with it. Jefferies’ Brent Thill put it bluntly: the economy is healthy and the bear thesis is “garbage.” His words, not mine, though I did laugh when I read it. And his reasoning isn’t crazy. In a single quarter, Google Cloud revenue rose 63% year over year to about $20 billion. AWS added roughly $8.3 billion of year-over-year growth to reach $37.6 billion. Azure added about $7.9 billion to reach $34.7 billion. That’s revenue climbing alongside the capex, which is exactly what you’d want to see from a real infrastructure buildout rather than a speculative mania.
Same story one layer down. Nvidia’s data-center revenue hit $75.2 billion in one quarter, up 92% year over year, on actual GPU shipments. Bernstein has a line I keep coming back to: Big Tech is “the new Big Oil,” and its combined capex passed the old Big Oil’s 2013 level of about $166 billion years ago and never looked back. Huge capital cycles aren’t new. They built the railroads, the telecom networks, the electrical grid. The companies that came out the other side had real, durable moats.
The bear case, and this is the uncomfortable part
The bear case is about circular financing, and the mechanism is almost embarrassingly simple once you strip the jargon off it. A chip maker or cloud provider invests billions in an AI startup. The startup spends that same money buying chips or cloud capacity from the investor. The investor books it as revenue. Everyone applauds. No actual outside cash has entered the loop.
The clearest example is Nvidia and OpenAI. It started in September 2025 with a letter of intent for Nvidia to invest up to $100 billion in OpenAI’s infrastructure. By early 2026 the deal had stalled; when OpenAI’s $122 billion round closed at an $852 billion valuation that March, Nvidia’s actual check was $30 billion, and Jensen Huang said a $100 billion outcome was “probably not in the cards.” Separate reporting has described an additional, debt-backed capacity commitment on top of that: Nvidia, in effect, helping underwrite its own customer’s ability to keep buying Nvidia chips. UBS has estimated the OpenAI relationship could account for as much as 13% of Nvidia’s 2026 revenue.
Michael Burry, who built his reputation shorting the 2008 housing bubble, responded to the arrangement in one line: history repeating. Bernstein’s Stacy Rasgon had flagged the pattern back in September 2025, noting Nvidia had made more than 50 AI-startup venture investments in a year, some of which came back as hardware orders. The dot-com comparison gets repeated so often it’s easy to tune out, but the sharper analogy is the late-1990s telecom vendor-financing episode, when companies like Nortel funded their customers’ purchases, booked the sales as organic growth, and then collapsed when the demand didn’t show up.
Where reasonable people actually disagree
Taking only the frightening data and ignoring the rest is lazy, so here’s the serious counterargument. The research team at Acadian argues the real warning signs would be equity going into these deals, hidden leverage on the balance sheet, or distorted fundamentals, and in most recent deals, including AMD and OpenAI, those aren’t present. Their framework lists four bubble markers: heavy net equity issuance, a wave of AI IPOs, acquisition frenzies fed by inflated stock, and megacaps ending buybacks. All four were flashing in 1999–2000. On their read, none are flashing hard right now.
That’s a good argument, and it’s roughly where the thoughtful disagreement lives. The counter to it, which comes from economist Noah Smith, is that Nvidia’s trade receivables have been growing faster than its sales. Vendor financing that shows up as receivables doesn’t announce itself with IPO mania. It sits quietly on a balance sheet until it doesn’t.
The part that gets almost no coverage: depreciation
Put $150 billion or more into assets that last five to six years and you’re booking something like $17–20 billion a year in depreciation, per company, before that particular batch of spending has earned anything back. That’s not a forecast; it’s accounting arithmetic. Free cash flow is already feeling it. Meta’s FCF margin is narrowing as operating cash generation falls relative to capex. Microsoft’s capex-to-operating-cash-flow ratio has pushed past 50%.
The market has tolerated this because cloud revenue keeps accelerating and because the alternative, under-investing and losing the infrastructure race, looks worse to every CEO in the room. But tolerance is temporary. If cloud growth slows while capex stays at this pace, the multiple compression could hit the whole sector at once, because markets don’t reprice hundred-billion-dollar assumptions one company at a time.
My actual read
I don’t think this is cleanly one thing or the other. I think it’s a genuine productive buildout and a 1999-style excess happening in the same market at the same time, which is an uncomfortable place to invest from. The hyperscaler-to-hyperscaler spending looks structurally sound to me, because the cloud revenue growth underneath it is real. The Nvidia–OpenAI financing loop, where a company posting large losses collects hundred-billion-dollar backstops from its main chip supplier, is the part of the cycle that could turn ugly.
So here’s what I’m doing with my own money. I’m not worried about Nvidia or the hyperscalers themselves; their revenue isn’t mostly circular, and their balance sheets don’t yet show the equity-issuance red flags Acadian’s framework watches for. I am being more conservative about valuations that lean heavily on a single financing deal, because one stalled headline between two megacaps can panic all three.
Big capital cycles have built durable industries: railroads, telecom, electrification. They have also produced spectacular wreckage when demand didn’t arrive on the spending schedule. The honest answer to “is this a bubble” isn’t yes or no. It’s: watch the receivables, watch for the start of heavy equity issuance, watch whether cloud revenue growth slows or reaccelerates. Those tells matter more than the number of zeros in any one headline. I’ve been burned before holding a clean story into a bad market, and this is about as consequential a capital cycle as I’ve seen in ten years. Be skeptical of anyone, bull or bear, selling you the one-sentence version.