Is AI a bubble? The money is real, and so is the circle it moves in

You cannot go a week in 2026 without someone declaring that AI is a bubble about to burst, or that it is the biggest opportunity in history. Both camps are loud, both cite huge numbers, and the truth is more interesting than either. The AI build-out is real, the money flowing into it is staggering, and the way that money moves has started to worry even people who are bullish on the technology. Here is the honest version: what the "circular financing" actually is, what is genuinely worrying, and why "is it a bubble" is the wrong yes-or-no question.

What is the "circular financing" everyone is worried about?

Start with the loop, because it is the heart of the anxiety. In effect, dollars can leave Nvidia's balance sheet labelled "investment" and return to its income statement labelled "revenue", having travelled through OpenAI and a cloud provider on the way. Bloomberg flagged exactly this circular structure across Nvidia's deals: supporters call it ordinary vendor financing, while critics call it round-tripping, a way to make demand look larger than it independently is.

In plain terms: Nvidia invests in OpenAI; OpenAI commits to spend hundreds of billions on computing power; that computing runs in datacenters built by cloud firms like Oracle; and those firms buy their chips from Nvidia. The money that started as Nvidia's investment can end up back as Nvidia's sales. Do that across enough deals and you can inflate how much real, independent demand there looks to be.

The specific deals are real, but the numbers need care, because a lot of the eye-watering figures are commitments and letters of intent, not cash already spent:

• Nvidia and OpenAI signed a letter of intent in September 2025 under which Nvidia "intends to invest up to $100 billion in OpenAI progressively as each gigawatt is deployed." Reporting since suggests the full $100 billion stalled and only a fraction, around $30 billion, was actually committed.
• OpenAI's compute commitments reportedly total over $1 trillion across the decade, spread among Oracle (a reported ~$300 billion deal), Microsoft, Amazon, CoreWeave, AMD and Broadcom. Oracle's disclosed backlog of contracted future revenue ballooned past $500 billion, much of it tied to a single customer widely understood to be OpenAI.
• The chip deals are signed: AMD granted OpenAI a warrant for up to 160 million AMD shares tied to deploying 6 gigawatts of its GPUs, and Broadcom agreed to co-develop 10 gigawatts of OpenAI-designed accelerators.
• Most recently, the picture got bigger. The Wall Street Journal reported in late July 2026 that Nvidia was in talks to provide roughly $250 billion in financing guarantees for OpenAI's Ohio datacenters, and in mid-August 2026 Bloomberg reported that this sits inside a far larger roughly $500 billion AI-financing consortium Nvidia is assembling with Wall Street heavyweights including Apollo, Blackstone, BlackRock, Goldman Sachs and KKR to underwrite the build-out. It is the most aggressive version of the loop yet: the chipmaker helping to fund the very demand for its own chips, now at a scale big enough to move credit markets.

Is it actually a bubble?

This is where serious, informed people genuinely disagree, so here are both sides in their own words.

The bears focus on the financing, not the technology. CNBC's Jim Cramer, who lived through the dot-com crash, warned that "a spectre is haunting this market. The spectre of the year 2000," and put the danger bluntly: "what we learned in 2000 is that you don't lend to customers who buy your goods. They might default and your earnings get smashed." Apollo's chief economist Torsten Slok goes further, arguing "the top 10 companies in the S&P 500 today are more overvalued than they were in the 1990s." Note what Cramer is not saying: he still calls Nvidia "an exceptional company." His worry is the pattern, not the product.

The bulls say this rhymes with 1999 but does not repeat it. Nvidia CEO Jensen Huang responded to the bubble talk directly: "there's been a lot of talk about an AI bubble. From our vantage point, we see something very different." The strongest version of his case is that, unlike the fibre-optic firms of 1999 that had valuations and no revenue, the AI leaders have enormous and rapidly compounding revenue, and that the build-out is funded largely from the cash flows of profitable giants rather than junk debt. The US Federal Reserve chair played down interest rates as a driver in late 2025, noting that datacenter economics carry "very high present value" and the decision is "not about 25 basis points here or there."

But isn't the revenue real?

Yes, and this is the single most important fact separating 2026 from 2000. The money coming in is not imaginary. OpenAI's revenue reportedly reached about $13 billion in 2025, with an annualised run-rate that reached about $25 billion by early 2026 and has held roughly there since. Rival Anthropic reportedly filed confidentially for an IPO at a roughly $965 billion valuation on a revenue run-rate near $47 billion, and Microsoft has reported an AI business growing well over 100% a year. People and companies are paying real money for these tools, at scale, right now.

But hold the other half of the ledger in view too. Growing revenue is not the same as profit. OpenAI reportedly ran an operating loss of around $21 billion in 2025 on that ~$13 billion of revenue, because the compute, talent and build-out cost far more than the product brings in. And be careful with "run-rate," a figure that annualises the best recent month; it is a real signal of momentum, but it is not money in the bank for the year. The revenue is real. So are the losses funding the growth.

What could actually go wrong?

Strip away the noise and the bear case is really about a few concrete risks, not a belief that AI is fake:

• Circularity and default risk. If demand is partly propped up by vendors financing their own customers, a single big customer stumbling could ripple back through everyone who booked that spending as revenue.
• Concentration. When a huge share of one cloud provider's future revenue depends on one unprofitable AI lab, that is a lot of eggs in one basket.
• Depreciation. The chips at the centre of all this lose value fast as newer generations arrive, so the returns have to come quickly to justify the spend.
• Power. These datacenters need gigawatts of electricity the grid does not yet have, which is why the AI build-out has become an energy story as much as a computing one.

None of these means a crash is coming. All of them mean the trillion-plus dollars of commitments have to turn into durable profit, not just usage, for the maths to work.

The circular money, at a glance

Deal	Between	Reported size	Status
Compute + investment	Nvidia → OpenAI	"up to $100B"	Letter of intent (Sept 2025); ~$30B reportedly committed
Cloud compute	OpenAI → Oracle	~$300B	Reported multi-year contract
GPUs + warrant	AMD ↔ OpenAI	6GW + 160M-share warrant	Signed (Oct 2025)
Custom chips	OpenAI ↔ Broadcom	10GW of accelerators	Signed (Oct 2025)
Datacenter financing	Nvidia (+ ~$500B consortium) → OpenAI	~$250B guarantee	Reported Jul-Aug 2026; part of a larger consortium
Overall infrastructure	Stargate (OpenAI/SoftBank/Oracle)	up to $500B	Announced target

So, is it a bubble?

The honest answer refuses the headline. This is not 1999, because the revenue is real and large and growing faster than almost anything in business history. But it is also not risk-free, because the financing has become genuinely circular, the spending is concentrated in a handful of intertwined players, and the losses funding the growth are enormous.

The useful question is not "is AI real." It obviously is; hundreds of millions of people use it and pay for it. The useful question is whether the returns will ever justify the more than a trillion dollars in deals now committed to building it, and whether the circle of money holds if one big link weakens. Bulls and bears are really arguing about that, not about whether the technology works. When you next see "AI is a bubble" or "AI is unstoppable," the honest reading is that both are half-right, and which half wins is still being decided. For more, see the AI section and our guide to the best AI chatbot, the kind of product all this money is racing to build.