Brace Yourself: Understanding the AI Bubble Argument and What It Means for Tech

01 The Argument, Briefly

In a video that has drawn more than two million views, former US Labor Secretary Robert Reich makes a compact and deliberately uncomfortable argument: the money pouring into artificial intelligence behaves less like sober industrial investment and more like a classic asset bubble. His evidence, drawn from public reporting on deals among chipmakers, model labs, and cloud companies, centers on three claims: financing arrangements that loop back on themselves, private valuations far ahead of revenue, and a market whose confidence in AI stands in for measurable end demand.

Bubbles are only obvious in retrospect, which is why the strongest skeptics and the strongest bulls express equal certainty. The Wikipedia article on the AI bubble catalogues a debate that intensified through 2025 and 2026: comparisons with the dot-com era made by investors such as Ray Dalio, circular-investment concerns raised by equity analysts, and counterarguments built on real revenue growth. Nothing about the argument is settled fact, and this piece treats it accordingly: as a case under active debate.

Our goal here is to reconstruct the argument the video makes, lay out the underlying numbers from institutional sources, and give the strongest counterargument a fair hearing. Where a figure is measured, we say so; where it is an estimate or an interpretation, we say that too.

02 How Circular Investment Works

The heart of the bubble argument is what analysts call circular, or vendor-financed, investment: capital that flows from a supplier to a customer who then spends some of it back with the same supplier. Reporting by Reuters during 2025 documented a chain of exactly this shape. Nvidia announced it would invest up to $100 billion in OpenAI, while OpenAI simultaneously committed to purchase $100 billion of Nvidia systems. OpenAI signed cloud-compute contracts with Oracle reported at roughly $300 billion over five years, and Oracle is one of Nvidia's largest customers. OpenAI signed multi-year compute deals with CoreWeave, a cloud provider that builds its data centers overwhelmingly on Nvidia hardware, worth tens of billions of dollars in aggregate.

Follow the dollar and the pattern appears. A supplier invests in a customer; the customer buys from the supplier; revenue shows up on multiple ledgers. Each individual deal is legal and was announced publicly, and the headline figures are upper bounds rather than sums already spent. But revenue booked inside a loop is weaker evidence of independent demand than revenue from customers who pay with money that did not originate in the same web. That is an interpretation, not a measurement, and it is the crux of the whole debate.

The historical parallel is the 1990s telecom boom, when equipment vendors such as Lucent and Nortel financed carriers that bought their gear. When the carriers failed, the vendors absorbed billions in write-offs, turning a sector downturn into an equipment-industry collapse. The bubble argument holds that today's chip-to-cloud financing web could propagate losses through the AI supply chain in the same way. The counterargument holds that the modern web is collateralized by real revenue from outside the loop, a claim we examine below.

03 The Valuation Gap

Start with the anchors that were actually reported. In October 2024, OpenAI closed a funding round at a reported $157 billion valuation; in March 2025, a SoftBank-led round reportedly valued it at $300 billion. Against that, press reporting placed OpenAI's annualized revenue near $3.7 billion in late 2024, with an internal projection of roughly $12 billion for 2025. The revenue figures are reported estimates, not audited financials; private companies disclose little, and OpenAI's numbers cannot be verified the way a listed company's can.

Even taking the estimates generously, the market was paying something on the order of twenty-five times projected annual revenue for the company at the center of the boom. Mature public software companies trade in the middle single digits on that metric. Defenders reply that the growth rate justifies the multiple, and that private valuations are negotiated prices rather than liquid market-clearing ones. Both statements are true, which is exactly why the disagreement persists rather than resolving.

The chart makes the bubble argument visible in a single image: the valuation bars tower over revenue bars that barely register at this scale. That disproportion is either the defining feature of the boom or the ordinary arithmetic of venture investing in a category-defining company, depending on which analyst you read. The number alone cannot settle it.

04 The Capex Numbers

The bubble argument also points at capital expenditure, the spending side of the boom. Between December 2024 and February 2025, the four largest cloud companies announced 2025 spending plans that added up to more than $300 billion: roughly $100 billion from Amazon, $80 billion from Microsoft, $75 billion from Alphabet, and $60 to $65 billion from Meta, per company statements on earnings calls. Those are announced guidance figures, and they were revised upward repeatedly through 2025 and 2026.

Two things are true at once. The spending is enormous, and it buys real things: buildings, transformers, turbines, and accelerator chips. That is different in kind from a purely financial bubble in which money never becomes productive capacity. But the 1990s fiber buildout, the parallel most often cited by bubble proponents, had the same character, and WorldCom and Global Crossing still went bankrupt when demand did not arrive fast enough. The fiber itself was eventually used; the shareholders were still wiped out. Overbuilding real assets produces glut and falling prices, not necessarily permanent loss, and that distinction matters for how a correction would actually feel.

So the capex question reduces to utilization. If demand for inference continues to compound, the spend earns a return; if it stalls, the industry owns an extraordinarily expensive fleet of underused hardware. That is the honest shape of the disagreement, and no guidance number resolves it.

05 Echoes of the Dot-Com Era

The strongest structural parallel to 1999 is concentration. By early 2025, the ten largest companies made up roughly 39 percent of the S&P 500's weight, against roughly 25 percent at the dot-com peak, per S&P Dow Jones Indices data cited in financial press coverage; both figures are rounded. A narrower market means a technology selloff transmits directly into index funds and retirement accounts, which is a large part of what made the dot-com bust so broadly felt even for people who never bought a single dot-com stock.

Ray Dalio, Bridgewater's founder, has drawn the dot-com comparison explicitly, as summarized in the Wikipedia article on the AI bubble. The echoes he and Reich point to are pattern-matching, and the pattern is recognizable: a new technology narrative that reorders the market, retail enthusiasm arriving after institutional money, vendor financing propping the supply chain, and valuations that assume the future arrives on schedule rather than merely eventually.

The differences deserve equal billing. Many of the 1999 leaders of the boom were companies with negligible revenue funded by fresh equity issuance. The current AI buildout is led by incumbents with vast operating cash flow, and the semiconductor at its center, Nvidia, was reporting $130 billion in annual revenue at its fiscal 2025 close, a measured figure from an SEC filing rather than a projection. A bubble led by cash-rich incumbents can still deflate, but it deflates differently: more slowly, with write-offs and write-downs rather than overnight evaporation.

06 The Counterargument

The bull case does not deny the financing loops; it argues that they sit on top of measured, compounding demand. Nvidia's fiscal 2025 revenue of $130.5 billion, of which $115.2 billion came from the data center segment, is a figure from an SEC Form 10-K, not a press release. Cloud AI services at Microsoft, Amazon, and Alphabet were growing at double-digit rates in quarterly reports through 2025, and those segments were producing margin, not just bookings. Whatever the accounting web looks like from above, customers outside the web, enterprises buying inference, are paying real invoices.

There is also a physical argument. Power constraints, land, and grid interconnect queues cap how fast capacity can actually be built, which limits the wildest overbuild scenarios in a way that 1990s fiber was not limited. And data centers carry residual value: the buildings, the cooling plant, and the substations outlive any single generation of accelerator chips, leaving recoverable assets behind even in a downturn.

Finally, the skeptics' own historical example cuts against the strongest form of the bear case. The internet was not a fraud in 2000; the pricing was wrong, not the technology. Amazon, built on bubble-era financing, survived and eventually dominated. The bubble argument is a claim about prices, and prices can correct without the underlying technology failing. Even Reich's video, read carefully, is a claim about financial fragility, not about whether AI works.

07 What a Correction Would and Would Not Change

For consumers, a correction would likely be mild at first. Models already deployed keep running regardless of their makers' share prices, and an overcapacity of compute tends to push inference prices down, which is what happened to bandwidth prices after 2001. Products at the margin, features attached to thinly funded startups, would disappear first, and ecosystems that depend on free-tier subsidies would see those subsidies shrink.

For frontier development, the effects would be sharper. Training runs at the frontier cost billions per attempt, and that capital is the most bubble-sensitive component of the entire stack. A funding retrenchment would mean fewer independent labs, slower capability gains than the current roadmap projects, and consolidation around whoever possesses durable cash flow. That is an interpretation, but it is consistent with every prior capital cycle in the history of the technology industry.

For investors, the honest summary is that the bubble label is an argument under active debate, not a fact. The measured record, the financing web, the valuation multiples, the capex totals, and the revenue growth, is compatible with more than one future. The variable that would settle the argument is the variable worth watching: who pays for AI, in cash, at prices that hold up outside the loop.