Is AI a Bubble? We Checked the Numbers (November 2026)

This guide was first published on saurabhbhayana.com.

Anthropic made about $33 billion in the first nine months of 2026. It has also signed up for about $895 billion of computing and data center deals, and the press expects its IPO to value it at somewhere between $1.5 and $2 trillion. Numbers like these are why "is AI a bubble" is one of the most searched questions about technology right now.

Here is the honest caveat first: nobody can tell you when a bubble will burst, and this is not investment advice. What we can do is check the numbers. A popular developer video on the AI bubble has been doing the rounds this month, and its story is mostly right. Several of its headline figures are not, so we went back to the filings and reports and corrected them below.

We build and run AI tools every day, so we also care about the practical side: if the AI bubble bursts, what happens to the APIs, prices and products you depend on? That part is at the end, with what we do ourselves.

Is AI a bubble? Our short answer

Yes and no. The money around AI looks like a bubble: companies are spending far more on chips and data centers than AI brings in, much of it with borrowed money and long contracts that cannot be cancelled. The technology does not: people and businesses use and pay for AI every day, and that is growing fast.

A bubble is about the price paid, not about whether the thing is useful. The internet was useful in 2000. It was also overpriced.

That is the same pattern as the dot-com era. The internet was real, the spending on it ran years ahead of the revenue, and the crash wiped out the companies that had bet on the timing, not the internet itself. Our view is that AI is on a similar path: the tools stay, the prices and valuations reset.

The fix: separate the two questions every time you read a headline. "Is AI useful?" and "Is AI priced right?" have different answers, and mixing them is how both the hype and the doom takes go wrong.

How we got here: from one research paper to a spending race

The story starts in June 2017 with a Google research paper called "Attention Is All You Need". It introduced the transformer, the T in ChatGPT. Google published it openly. OpenAI, founded in 2015 as a non-profit, built its GPT models on it, changed its structure in 2019 to raise money, and took $1 billion from Microsoft that year.

ChatGPT launched on November 30, 2022 and reached about 100 million users in two months, according to a UBS estimate. Microsoft added a reported $10 billion in January 2023. Google rushed out Bard, which gave a wrong answer about the James Webb telescope in its first demo, and Alphabet lost about $100 billion of market value in a day. From then on, every big tech company decided it had to train frontier models, whatever the cost.

And the cost kept rising. GPT-3's training run was estimated at about $4.6 million. Sam Altman said GPT-4 cost more than $100 million. By 2024, Anthropic's Dario Amodei was talking about models that cost around $1 billion to train. Meanwhile each new model felt like a smaller step up than GPT-3 to GPT-4 did. More money for smaller visible gains is the first sign that a race is running ahead of its returns.

The peak moment came on September 10, 2025. Oracle said it had signed hundreds of billions of dollars of new cloud contracts, including a reported $300 billion compute deal with OpenAI over about five years. Its stock closed up about 36% in a day, and co-founder Larry Ellison briefly passed Elon Musk as the richest person in the world on Bloomberg's index.

The fix: when you hear "the AI story", ask which part. The research worked. The products found users. The open question is the financing, and that is where the rest of this guide looks.

The AI bubble numbers, checked

Here are the figures you will hear most often, next to what the filings and reports actually show. Where sources disagree, we say so.

Two things stand out. First, the direction of the video is right: spending is far ahead of revenue, and lenders and rating agencies have started to say so. Second, the most dramatic numbers are the least reliable. A bubble story does not need exaggerated figures; the real ones are striking enough.

The fix: when a number about the AI bubble sounds wild, look for the filing (S-1, 10-Q, earnings release) or a named outlet. If you only find it in videos and posts quoting each other, treat it as a rumour.

Why people call it a bubble: spending runs ahead of revenue

Look at what the biggest players reported in the last few weeks. The money is going out much faster than it is coming in, and the gap is now visible in cash flow, not just in forecasts.

Revenue is growing fast. The commitments are growing faster.

• Alphabet: revenue up about 25% to $127.6 billion in Q3 2026 and Google Cloud up from $15.2 billion to $30.7 billion. Capital expenditure that quarter was $51.4 billion, and free cash flow was minus $9.1 billion, the second negative quarter in a row.
• Meta: capital expenditure guided at $135 to $145 billion for 2026. S&P now expects Meta's free operating cash flow to stay negative through at least 2027.
• Anthropic: revenue grew from about $11.6 billion in Q2 to about $17 billion in Q3, but the company swung from a $974 million operating profit to a $567 million operating loss, which it said was "primarily driven by $5.0 billion of compute under short-term commitments".
• OpenAI: investors pulled back enough that the round target dropped by about a third.
• Oracle: its New Mexico data center project, Project Jupiter, hit a force majeure notice in September 2026, and the roughly $18 billion loan behind it has been quoted at 89 to 91 cents on the dollar.

Think of a restaurant owner who borrows to build ten kitchens because everyone says the town will only eat out from now on. Business is good and growing. But ten kitchens need rent, staff and power every month, and the bank starts asking when the customers for all ten will arrive. Nobody is saying food is bad. They are asking whether the kitchens were built too early.

The fix: when you judge an AI company, look past revenue growth to free cash flow and commitments. Fast growth with negative cash flow and years of fixed bills is a bet on timing, and timing is exactly what bubbles get wrong.

What is circular financing in AI?

Circular financing is when the same money moves around a small group of companies and each one counts it as revenue. In AI, the loop runs between chip makers, AI labs and the clouds that build data centers. The IMF and the Bank of England both warned in October 2025 that AI valuations could correct sharply, and circular deals are a big part of why.

Each step is a real sale. The question is how much new money enters the loop from paying customers.

1. A chip maker invests in an AI lab. In September 2025, Nvidia said it planned to invest up to $100 billion in OpenAI as OpenAI deployed Nvidia systems.
2. The AI lab uses that money, plus money from other investors, to rent computing power from clouds and data center companies.
3. The data center companies borrow against those rental contracts and buy chips from the chip maker.
4. The chip maker books record sales, its value rises, and it has even more money to invest in AI labs.

In plain numbers: A lends B $100, B spends it at C, and C spends it at A. All three report $100 of sales, so the group looks like $300 of business, but only $100 ever came in from outside. Similar loops show up elsewhere too, such as OpenAI's October 2025 chip deal with AMD, which came with the right to buy AMD shares.

To be fair to the other side: circular deals are not fake. The chips are real, the data centers are real, and plenty of outside customers do pay. The risk is concentration. If one link, usually a lab that cannot raise its next round, stops paying, the revenue at every other link shrinks at the same time.

The fix: the number that matters is money from end customers, meaning businesses and people paying for AI products. That is why Anthropic's 11,900 customers paying over $100,000 a year is a more useful figure than any chip sales headline.

Chips age fast: the depreciation question

Bulls compare today to the fibre optic cable laid during the dot-com boom. Most of it sat unused for years, then carried the streaming video of the 2010s. The data centers built now, they say, will be used the same way later.

We think the comparison breaks on one point: cable lasts decades, AI chips do not. Nvidia has moved from Hopper (2022) to Blackwell (2024), with Rubin lined up after it, and each generation does much more work per watt. A chip from five years ago is still a chip, but few people want to pay to run it when the new one is cheaper per answer.

• The big clouds depreciate servers over about five to six years, which spreads the cost thin and makes profits look higher.
• Amazon cut part of its server fleet from six to five years from January 2025, pointing to how fast AI hardware is changing.
• Investor Michael Burry argues the real useful life of AI chips is closer to two or three years, and that the industry is understating depreciation by tens of billions of dollars a year.

Who is right depends on whether older chips keep earning money on cheaper work like running models (inference) after the newest chips take over training. Some will. Whether enough of them will is part of the same bet.

The fix: when a company says its AI spending will pay off over "the life of the assets", check how long it assumes that life is. Six years for a chip that is two generations old in four is optimistic.

When will the AI bubble burst?

Nobody knows, and anyone who gives you a date is guessing. Bubbles can run for years after people start calling them bubbles, and they often end with a trigger that only looks obvious afterwards. What you can do is watch the signals that would show the money drying up.

The watch list we use. None of these alone pops a bubble; several going wrong together would.

1. Nvidia's results on November 17, 2026: worrying if data center growth slows sharply or a handful of buyers make up most of the sales.
2. Anthropic's IPO pricing: worrying if it prices well below the reported range or falls hard in its first weeks of trading. Early investors selling to the public at the top is a classic late-bubble pattern.
3. OpenAI's round: worrying if it closes well under the cut $20 billion target or at a lower valuation again.
4. Big Tech's Q4 results in late January 2027: worrying if free cash flow stays negative and 2027 spending plans still go up.
5. Debt and ratings: worrying if data center loans keep trading below face value or more rating outlooks turn negative.

Will the AI bubble burst in 2026?

There are only a few weeks of 2026 left, and the signals above are mixed rather than all red. A sharp fall in AI stocks is possible at any time, but we would not call it a burst until money stops reaching the AI labs and builders. That shows up in funding rounds and credit, not in a single bad day for one stock.

The fix: pick three of these signals and check them once a month instead of reacting to every headline. If three or more turn bad together, take it seriously.

What happens if the AI bubble bursts?

If the AI bubble bursts, stock prices and valuations fall, weaker AI startups shut down, and the cheap AI you use today gets more expensive. AI itself does not go away. That is the lesson of the dot-com crash.

The Nasdaq fell about 78% from its March 2000 peak to October 2002, and thousands of internet companies closed. The internet did not. Gmail launched in 2004, YouTube in 2005 and Amazon Web Services in 2006, built by people who used the crash to hire, buy cheap servers and focus on things customers paid for.

The biggest change for most people would be price. Today a lot of AI is sold below cost, paid for by investors and by Big Tech's other businesses. When that money gets more careful, the subsidy shrinks. Cheap open-weight models from China, such as DeepSeek and Qwen, push prices the other way, which is one reason we do not expect AI to become unaffordable, just less of a free party.

The fix: if your business depends on an AI API, assume your cost per task could double and check that you would still be fine. If you would not be, that is the part to fix now.

What to do if you build with AI (what we do)

We run AI inside our own products, including wptaskify, our WordPress tool, every day. These are the rules we follow, and they are the same whether the bubble bursts next month or never.

1. Know your cost per task. We log tokens and spend per request. A blog cover image at medium quality costs us about 2 US cents; a long article rewrite costs more. If you do not know your numbers, a price change will surprise you.
2. Default to the smallest model that does the job. Our products run on a small, cheap model by default and only step up when a task needs it. Most of the bill comes from using a big model where a small one would do.
3. Give each project its own API key and a hard monthly budget. In October 2026 one shared key ran dry in the middle of a batch and only 5 of 40 images were made. Separate keys stop one experiment from draining the live product.
4. Keep a fallback. Write your code so you can switch providers, and know which open-weight model you would move to if prices jumped.
5. Do not build a thin wrapper. If your product is one prompt on top of someone else's API, a price rise or a new feature from that provider ends you. Own the workflow, the data and the customer.
6. Learn the fundamentals. When AI gets expensive, the developers who can cache results, cut tokens, optimise database queries and design systems are worth more, not less.

Tip: Try this today: export last month's AI usage from your provider's dashboard, divide by the number of tasks you ran, then multiply by two. If the business still works at that price, you are in good shape. If not, start with a cheaper model and caching before you touch anything else.

If you want to understand where AI costs come from, start with what LLM inference is and why every answer costs money. For the fallback plan, see whether you can run an LLM on your own laptop. And if you are a student wondering what to learn, our guide on how to become an AI engineer leans on exactly these fundamentals.

The fix: do these six things before you need them. Cost discipline is cheap when prices are low and painful to learn when they rise.

Is AI overhyped? Where we disagree with the doom takes

Some of it is. Not all of it. The spending and the valuations are ahead of the business, and a lot of AI products are demos with a price tag. But the usage is real, and a crash would not change the fact that AI already saves us hours every week.

• Real demand: Anthropic says 11,900 customers pay it more than $100,000 a year, up from 950 a year earlier. Companies do not sign that kind of bill for a toy.
• Real revenue growth: Google Cloud roughly doubled in a year, and AI is a big part of why.
• Overstated failure: the famous "95% of AI projects fail" figure comes from one 2025 study of company pilots. It says most pilots have not shown measurable profit yet, not that AI does not work.
• Overstated doom: a falling stock price is not the same as a failing technology. Cisco lost most of its value after 2000; routers kept selling.

The fix: keep using AI where it saves you real time or money, and stay sceptical of anything that only makes sense if prices stay this low forever. Our guide to which Claude model to use for which job is a good place to start trimming costs.

Takeaways

• Our verdict: the money around AI shows clear bubble signs, but the technology is not a bubble. People and businesses pay for AI every day; the open question is whether they will pay enough, soon enough, to cover what has been built.
• Anthropic's amended S-1 shows about $33 billion of revenue from January to September 2026 against about $895 billion of purchase commitments and leases, almost all for compute.
• Alphabet spent $51.4 billion on capital expenditure in Q3 2026 and had negative free cash flow of $9.1 billion, its second negative quarter in a row, even though Google Cloud revenue roughly doubled to $30.7 billion.
• S&P Global moved Meta's credit outlook to negative on October 30, 2026, citing AI spending that it expects to keep free cash flow negative through 2027. It kept the AA- rating.
• OpenAI cut its funding target from about $30 billion to about $20 billion and its valuation from $1.4 trillion to $1.25 trillion, according to reporting at the end of October 2026.
• Circular financing means money loops between chip makers, AI labs and data center builders, so each step books revenue while much less new money comes in from outside customers.
• Nobody can date a bubble. The signals worth watching are Nvidia's results on November 17, 2026, Anthropic's IPO pricing, OpenAI's round, Big Tech free cash flow in early 2027, and the price of data center debt.
• When the dot-com bubble burst, the Nasdaq fell about 78% from March 2000 to October 2002, yet Gmail, YouTube and AWS all launched in the years after. If you build with AI, plan for higher prices, not for AI disappearing.

FAQs

Is AI a bubble?

The money around AI shows bubble signs: spending on chips and data centers is far ahead of AI revenue, much of it funded with debt and long contracts. The technology itself is not a bubble, because people and businesses already pay for it and usage keeps growing. Our view is that prices and valuations will reset, not that AI will go away.

When will the AI bubble burst?

Nobody knows, and dates you see online are guesses. Watch signals instead: Nvidia's results on November 17, 2026, Anthropic's IPO pricing, whether OpenAI's round closes at its cut target, Big Tech free cash flow in late January 2027, and the price of data center debt. Several of these going wrong together would be the warning.

Will the AI bubble burst in 2026?

A sharp drop in AI stocks can happen any day, but a real burst means money stops reaching AI labs and data center builders, and the signals for that are mixed as of November 2026. We would watch Anthropic's IPO and OpenAI's funding round before calling it.

What happens if the AI bubble bursts?

Valuations and AI stocks fall, weaker startups close, data center projects get delayed, and AI that is cheap today becomes more expensive. AI tools do not disappear. After the dot-com crash the internet kept growing, and companies like YouTube and AWS launched in the years after.

What is circular financing in AI?

It is when money loops between a few companies and each books it as revenue. A chip maker invests in an AI lab, the lab rents computing power from a data center company, and the data center company buys chips from the chip maker, often with borrowed money. The sales are real, but less new money comes from outside customers than the totals suggest.

Is AI overhyped?

Partly. Valuations and spending plans assume growth that has not arrived yet, and many AI products are thin demos. But demand is real: Anthropic reports 11,900 customers paying over $100,000 a year, and Google Cloud revenue roughly doubled in a year.

Should I stop learning AI if it is a bubble?

No. A financial bubble bursting does not end a technology; it removes the hype money. Learn AI together with the fundamentals that matter when it gets expensive: caching, cost control, databases and system design. Those skills get more valuable, not less, in a downturn.

Is it a good time to buy AI stocks?

We do not give investment advice. What we can say is that the numbers in this guide, such as negative free cash flow at Alphabet and Meta and Anthropic's $895 billion of commitments, are the kind of facts to read in the filings yourself before deciding, rather than relying on videos or headlines.

Originally published at saurabhbhayana.com, where I write about AI automation, AI search and Local SEO for small businesses.