An analyst asked Microsoft, on an earnings call, the question much of the market has been muttering: “Are we in a bubble?” On Alphabet’s call, another asked what early signs showed the spending was producing better long-term returns. A technology-spending roundup dated 1 November 2025 recorded both. These were not posts from the cheap seats. They were questions put to the companies writing the checks.
The wider mood is just as blunt. In Bank of America’s Global Fund Manager Survey, as reported in a 23 November 2025 postscript to that roundup, 53% of participating managers thought AI stocks had reached bubble proportions. That was a slight step down from a record 54% the month before. In the November poll, 45% named an AI bubble as the top tail risk. A majority that is worried is not a crash that has already happened. It does mean the overvaluation argument is mainstream, not a side hobby for people who dislike software.
The spending has not blinked. The roundup’s headline figure put Big Tech’s aggregate investment at about $400 billion for 2025, with more expected in 2026. Treat that as the compilation’s total, not a line from a single filing. Inside it, Alphabet raised full-year capital-expenditure guidance from $85 billion to $91-93 billion. Meta’s capex nearly doubled from the prior year to $72 billion, and CFO Susan Li said 2026 would be “notably larger.” The checks are concentrated in a handful of names. A slogan about “tech companies” hides that.
The fight is over payback, not whether AI is fake
Listen to the people defending the buildout and you do not hear a defense of chatbots as a novelty. Youssef Squali, lead internet analyst at Truist Securities, has put the bull case cleanly: whoever reaches artificial general intelligence first gains a large competitive advantage, and the greater risk is underspending. On that logic, a data-center bill starts to look like insurance. It is also how a maybe becomes a multi-year mortgage. That is his case for the companies in the race. It is not yet a law of corporate finance.
The companies themselves are not speaking with one voice, which is more informative than any bubble headline. Amazon CEO Andy Jassy has said, “As fast as we’re adding capacity right now, we’re monetizing it.” Anat Ashkenazi, Alphabet’s CFO, has said the company is “already generating billions of dollars from AI in the quarter, ” and that long-term investments pass through a “rigorous framework.” Neither line specifies the margin, the products inside the number, or whether the return clears the cost of capital.
Microsoft’s constraint, on the roundup’s account, has been supply. Capacity shortages were expected to limit both current businesses and AI research until at least the first half of 2026, with Azure taking most of the revenue hit. Management saying capacity is tight is evidence of demand relative to current supply. It is not proof. It says nothing about returns once the buildings are full.
Meta has been the most explicit about optionality. Mark Zuckerberg has described front-loading capacity for an optimistic AGI case and, if that case is wrong, slowing new construction “while we grow into what we build.” Investors got a capex figure and an off-ramp. They did not get a date when AI revenue covers that capex. The roundup said that gap unsettled some of them. It should have. Superintelligence, if it shows up, would redraw competitive maps. The financing gets redrawn first, on a calendar nobody has published.
None of the revenue comments cancels the skeptic’s question. The same roundup said critics point to limited paying users for current tools, and to a long period before a workforce uses the technology well. Jassy monetizing capacity, and Ashkenazi citing billions, show that someone is paying. Reported revenue is not the same thing as a proven margin, broad adoption, or a return above the cost of capital. Huge spend, unproven consistency. That is what those earnings-call analysts were circling.
Coverage and borrowing are different stories
Cash-flow coverage is not the same thing as a debt boom. One figure can say a group still has residual cash. Another can say two members of that group are already borrowing heavily. Both can be true.
Bank of America credit strategists Yuri Seliger and Sohyun Marie Lee have said AI capital spending at five megacaps (Amazon, Alphabet, Microsoft, Meta, and Oracle) is growing faster than cash flow. The November roundup reported their reading of consensus estimates: AI capex at 94% of operating cash flow minus dividends and buybacks in 2025 and 2026, up from 76% in 2024. They treated a ratio still under 100% as a sign those firms did not yet need to borrow. That reading is an aggregate for five companies. It does not mean every name is equally covered. And it does not mean nobody is borrowing.
A 23 November 2025 account of the same boom said Meta and Oracle issued $75 billion in bonds and loans in September and October 2025 alone, more than double the annual average of the past decade. That account also said Bank of America had warned the buildout was “hitting a money wall.” Keep the $75 billion with those two companies and those two months. It is a credit-market warning about the pace of issuance, not a receipt showing the five-company cash math has already failed.
Behind those financing claims, commentators have alleged circular deals, with capital moving among investors, model labs, and the firms selling the computing. A late-November account described the expansion as spending funded mostly with borrowed money or circular financing. That description is stronger than the cash-flow ratio. The public summaries do not name who funded whom and then received the purchase order back. Until that map exists, the loop is a risk to watch, not a diagram to trade on.
A cheaper stock is not a dead technology
The late-1990s comparison is popular because the shape is familiar. A real technology. A market that ran ahead of proven business models. Speculative money that did not all survive the repricing. The internet kept changing the economy after that washout. The analogy does not come with a loss table. It is a split worth keeping, not a forecast.
An AI valuation reset would not, by itself, mean the end of AI. It would mean some prices, some capex plans, and some private marks were wrong about timing. It would still be messy for customers. Vendors with thin cash and expensive infrastructure can vanish or get bought for parts. Access, contract terms, and model pricing can move. If a 2026 plan assumes today’s vendor list and today’s rates, a credit squeeze is a procurement risk, not a philosophy seminar.
It also does not automatically bless some other trade. U.Today reported a post on X by CryptoQuant CEO Ki Young Ju arguing that an AI-bubble burst would do little harm to Bitcoin, on the view that Bitcoin had largely missed the run-up in AI stocks. He wrote:
“I don’t think an AI bubble burst would hurt BTC much.”
No benchmark, window, or return was attached. No flow evidence supported the further guess that investors had already sold Bitcoin to fund AI-stock bets. The reported post also said an AI crash is not necessarily bullish or bearish for Bitcoin, because the two seem unrelated, and that it may be time to rotate back. Those lines pull apart.
For anyone running a company, that is useful only as a warning against stapled narratives. Three claims keep getting glued together. AI equities may be expensive. The technology would likely outlast a valuation crash. Some unrelated asset would be spared, or even rewarded, if the first claim comes true. The survey and the capex figures support the first as a live concern. The second is a fair historical analogy. The third is an opinion, and it should not move an operating budget.
Buy the workflow, not the race
The hyperscalers are pricing a contest to reach AGI. Most businesses are pricing a workflow. Those are different bets. Copying the first one on a smaller balance sheet is how a sensible firm gets into trouble.
What a company outside that race should fund is narrower than the headlines. One sales queue, one baseline, one owner: an agent drafts the follow-up the team already sends, cycle time is checked against last quarter, and a named person takes the account back when the draft is wrong. That is a usable test of AI for business. It does not require a belief that superintelligence arrives on a schedule.
The price of that workflow is the risk worth watching, because your pilot does not set it. The financing of the race does. If AI capex stays near the whole of residual operating cash flow for the biggest spenders, and if more of the buildout shifts from cash toward bonds and loans, the price of access can move even while the models improve. Microsoft’s capacity limits can lift the near-term cost of getting in. A later glut can do the opposite. Zuckerberg has already sketched the off-ramp: slow new construction and grow into what was built. If more than one company takes that path, scarcity pricing does not last. Procurement should assume neither price is permanent. Shorter commitments, a second supplier, and a clause that lets you leave are worth more than a slogan about being early.
Squali’s underspending argument is a bull case for firms whose product is the model, or whose rival may ship a clearly better one. That is where it is persuasive. It is a weak guide for a regional bank, a manufacturer, or a sales organization. Missing a disciplined pilot is cheaper than funding capacity you cannot fill. Skipping the pilot because a bubble headline made the tools sound fictional is a quieter, separate expense.
Questions worth sitting with
-
Has the AI bubble already burst?
No. A 23 November 2025 account of Bank of America’s Global Fund Manager Survey put the share of managers who thought AI stocks had reached bubble proportions at 53%, down from 54% the month before, with 45% naming an AI bubble as the top tail risk in the November poll. That is a widely held fear, not a completed crash.
-
How big is the spending, in actual numbers?
A November 2025 spending roundup put Big Tech’s aggregate at about $400 billion in 2025, with more expected in 2026. That is the compilation’s headline total, not a single company filing. Within the same account, Alphabet lifted capex guidance from $85 billion to $91-93 billion, and Meta’s capex nearly doubled from the prior year to $72 billion. Susan Li, Meta’s CFO, said 2026 would be notably larger.
-
Are the biggest spenders already forced to borrow?
Not as a group, and not name by name. Seliger and Lee, as reported in that roundup, said estimated AI capex at Amazon, Alphabet, Microsoft, Meta, and Oracle was 94% of operating cash flow minus dividends and buybacks for 2025 and 2026, up from 76% in 2024. That means the five-company group can still cover the estimated buildout from residual cash flow in aggregate. It does not mean every name is equally covered, and it does not mean no one is borrowing. A 23 November 2025 account said Meta and Oracle issued $75 billion in bonds and loans in September and October 2025 alone.
-
Would a stock-market crack end the technology, including the agents companies are starting to deploy?
Not by itself. A repricing can destroy speculative valuations and strain vendors without erasing the technology, which is the lesson people draw from the internet after the late-1990s bust. It can still change prices, capacity, and which suppliers exist next year.
-
Should a reported Bitcoin “dodge” change a corporate plan?
No. Ju’s reported view, that an AI bust would not hurt Bitcoin much, is one analyst’s opinion, offered without a benchmark or evidence of capital flows. It is not evidence that an automation budget is safe, or that it should be cut.
-
What should a company that is not building the models fund while that race is being financed?
A pilot with a cost, a baseline, and an owner. Pick a workflow you already run, measure it against the current process, and name the person who takes over when the agent is wrong. Do not fund a smaller copy of a hyperscaler’s data-center bet.
Ask what the spend is supposed to return, who is borrowing to make it happen, and whether your own project still works if model prices jump or a vendor disappears. The bubble argument is a market argument. The work is still sitting on the desk.