The AI Bubble Is Waiting to Burst

On 21st October, Meta reported a very good quarter.

Revenue was $64.6 billion, up 26%. The stock rose in after-hours trading. Then the earnings call began, and the mood changed. The next day, Meta fell 7.8%. October ended with the stock down 15%, its worst month in four years.

Why? The business isn’t weak. The problem sits in one line of the numbers.

Meta spent $34.91 billion on capital expenditure in the quarter. It earned $32.79 billion in operating cash. So it spent more than it made. Free cash flow went negative by $2.12 billion. A week later, S&P Global changed Meta’s credit outlook to negative.

Now here’s the thing. This is not only a Meta story. It is the whole AI industry in one quarter.

Let me say upfront where I stand. I am not an AI sceptic. I want this technology to win. I think it will change how we work. But I also think the way the world is paying for it is a bubble. Both things can be true.

Here is my view in one line: AI is real. The financing of AI is a bubble. And bubbles end.

The AI Hype: Are Expectations Outpacing Reality?

Let me show you the data.

1. The biggest factory in the history of mankind

Think of AI as the largest factory we have ever built.

Four companies are building most of it: Amazon, Alphabet, Microsoft and Meta. In 2025 they spent about $410 billion on capex. For 2026, their guidance now adds up to roughly $730 billion. That is my math at the midpoints of what each company has said:

And it won’t stop there. Analysts expect hyperscaler capex to cross $1 trillion next year. S&P’s Visible Alpha puts 2027 at $1.2 trillion. That means nearly $2 trillion spent in just two years. UBS projects $4.1 trillion from 2026 to 2028. That is three times what was spent in the previous six years combined.

To be fair, the money is not all coming from savings. Over $500 billion of AI-related debt has been issued by late 2026, and $466 billion of it this year alone. Meta has even created a new class of non-voting shares to help fund the buildout. When the richest companies on earth need new share classes and bond markets to pay for their factory, you should pay attention.

2. Who has actually made money? The people who sold the parts.

So who has benefited so far?

The people who supplied the factory. Chip makers. Electrical equipment makers. Power companies. Land and real estate for data centers.

Look at Nvidia. In the quarter ended 26 July, Nvidia’s revenue was $96.2 billion, up 106% from a year ago. Data center revenue alone was $89.0 billion, up 117%. Gross margin was 75%. Net income was about $59.7 billion in one quarter. Its guidance for the next quarter is $108 billion. After the results, its market value reached $5.49 trillion.

I’m not saying Nvidia did anything wrong. It built a remarkable business. Frankly, I admire it.

But this is why I call Nvidia the drug dealer. AI capex is the drug. The high it gives is hope. And the supplier wants you to buy more, bigger and faster. A supplier never wants the party to end.

Notice what is changing, too. Analysts at XTB noted after the August results that Nvidia’s ecosystem now rests not only on selling chips but increasingly on helping to finance entire data center projects. When the seller starts helping the buyer pay, ask yourself why.

3. A correction to something I used to say

I used to say the AI factory makes nothing. That is not true, and I want to be honest about it.

The factory is making products. Real ones, with real revenue:

• OpenAI told investors its annualized revenue was about $50 billion at the end of September, on its own net-of-partner basis.
• Anthropic crossed a $65 billion annualized run rate by the end of July, according to Bloomberg figures reported by Axios.
• Microsoft said its AI business passed a $37 billion annual run rate, up 123%.
• Cloud is growing fast. AWS grew 39% last quarter to $45.98 billion. Azure grew 48%.

A caution on these numbers. They overlap. Part of what labs earn shows up inside cloud revenue too. And the two labs count revenue differently. Anthropic books some sales through cloud partners gross. OpenAI counts its Microsoft sales net. You can’t just add them up.

So the real question is not whether AI makes anything. It is whether it makes enough.

And this is where it gets hard to look away.

Gartner published a study on 14 October. It says AI data center owners would need to earn a cumulative $8.5 trillion in revenue through 2030 to cover announced capex, financing costs, and a 12% return on invested capital. Gartner’s forecast for all enterprise and consumer AI revenue that can be attributed to AI data centers over the same period is no more than $3.6 trillion.

Eight and a half trillion needed. Three and a half trillion expected.

There is a second sting. Gartner expects less than half of AI revenue to go to companies that own significant infrastructure. Most AI companies are asset-light. They rent the computing from someone else. So the people who built the factory may not be the people who collect the money.

Gartner’s conclusion is blunt. More than $3.5 trillion of AI infrastructure spending could be exposed to write-downs by 2030. If current investment continues, more than $4 trillion could be exposed by 2032.

And even the winners are not yet comfortable. OpenAI is on course, by one analysis, to pile up hundreds of billions in losses before turning cash positive around 2029 or 2030. Anthropic told investors it expected its first operating profit in the second quarter of this year. That is real progress. But it is one expected profitable quarter against a trillion-dollar buildout.

4. The ghost city problem

Now think of China’s ghost cities.

Developers built millions of apartments. Estimates of how many sit empty vary a lot depending on the method, from about 20 million to well over 60 million. Whatever the exact number, the lesson is simple. If you build far more than people will pay for, someone loses.

Let me be fair about where the comparison works and where it doesn’t.

Where it doesn’t work: AI chips are not sitting idle today. Nvidia is still supply constrained. Microsoft’s CFO said chip delays will push some capacity into the first half of 2027. The racks are being used.

Where it does work: Two places.

First, price. It isn’t enough that the chips are busy. They have to earn more than they cost, including the cost of borrowed money. That is the Gartner gap.

Second, lifespan. An apartment can sit empty for twenty years and still be an apartment. A GPU cannot. Forbes notes that state-of-the-art AI hardware may become obsolete in three to five years. The building lasts decades. The chips inside it don’t.

5. The depreciation trap

This is the part almost nobody talks about, so let me explain it simply.

When a company buys servers, it doesn’t count the full cost as an expense on day one. It spreads the cost over the servers’ useful life. Hyperscalers currently write off servers over roughly five to six years.

Now take a simple example. Say a company buys $400 billion of servers.

• Spread over six years, the yearly charge is about $67 billion.
• Spread over three years, the yearly charge is about $133 billion.

Same servers. Same cash spent. Reported profit is dramatically different.

Investor Michael Burry argues that the big companies are using lives that are too long. By his estimate, the industry could understate depreciation by about $176 billion between 2026 and 2028. That is his estimate, not an audited fact. But it is the right question to ask. If chips age faster than the accounting assumes, today’s profits are flattered, and tomorrow brings write-downs.

6. The credit market is saying what the stock market won’t

Stocks tell you what people hope. Credit markets tell you what lenders fear. And lenders are getting nervous.

Here are the facts from October.

CoreWeave, a company that rents out AI computing, carries more than $51 billion of debt and obligations. On 15 October its five-year credit default swap spread hit about 1,195 basis points, a record. At that peak, the market was pricing a default probability above 60% over five years. Its 2031 bond yielded close to 16%.

Oracle has about $132 billion of net debt and a BBB- rating, the lowest rung of investment grade. Its five-year default insurance cost was about 308 basis points on 21 October. On 2 January it was 141. That is more than double in under ten months.

Meta’s own default insurance and the Hyperion data center bond tell the same story. Meta’s credit default swap hit a record 108 basis points in October. The $27.3 billion Hyperion project bond, backed by Blue Owl and Meta, slid to about 91 cents on the dollar.

And in July, Nikkei Asia’s analysis of SEC filings found that Alphabet, Amazon, Meta, Microsoft and Oracle together carry about $1.35 trillion of balance sheet debt. They also carry another $1.65 trillion of commitments that are not on the balance sheet, mainly purchase commitments and leases not yet started.

Who holds this debt? Pension funds, insurers and asset managers. And ordinary savers. AI-related debt is now about 10% of the largest investment-grade corporate bond ETFs, up from 7% a few years ago. If you own a plain corporate bond fund, you may own part of this factory.

7. Which stage are we in?

Here is how I read it.

In July, Alphabet raised its capex forecast and the market sold off Alphabet, Amazon, Meta and Microsoft. That was the first signal. In October, with Meta, Alphabet and Amazon, the signal got louder. All three were punished for spending more than they earned.

The market is no longer rewarding the promise. It wants proof. It wants money coming in, not just money going out.

I believe that is the stage before the final reckoning. The last dance has begun. When a stock falls while the CFO is still talking about capex, investors have stopped paying for hope.

And remember what Meta’s CEO said on that call. Muse, Meta’s new AI assistant, is expensive to run. Meta has not even launched it worldwide because of the compute cost. He said Meta would try to make money from it sooner than usual, and that the math on all of this has to work out. That is an honest comment. It is also the whole problem in one sentence. The math has to work out.

Write-downs are coming. I don’t know the date. Nobody does. But the arithmetic says they are coming.

8. What could prove me wrong

I am passionate about this. But passion is not proof. So let me give you the other side.

• Demand is real and growing. Azure up 48%. AWS up 39%. Nvidia guiding to $108 billion. These are not the numbers of a dead market.
• Valuations are lower than in 2000. In late 2025, Bloomberg data showed the Nasdaq 100 trading at about 26 times expected earnings. At the dot-com peak it was over 80.
• The giants have cash. Microsoft, Alphabet, Amazon and Meta can absorb a lot of pain. The weakest links are the borrowers, not the giants.
• Capex growth may be flattening. Microsoft trimmed its forecast, partly because chip suppliers were late. Alphabet held its guidance. S&P expects Meta’s capex growth to flatten from 2028, though it said it is less sure Meta will cut.
• A bubble does not mean a fake technology. The internet was real. Fibre laid in the late 1990s was used years later. But the investors who paid for it were not the ones who benefited. That is the point.

So my claim is narrow. I’m not saying AI will fail. I’m saying the amount being spent has run ahead of the money coming in, and a lot of it is financed with debt. When that happens, someone takes the loss. History says it is usually the late investor.

9. What this means for India

India is not outside this story.

CBRE says India’s data center sector has attracted about $173 billion in investment commitments since 2021, including about $38 billion in the first half of 2026 alone. More than 15% of those commitments are tied to AI. KPMG puts installed capacity at about 1.9 GW in FY26. CBRE counts operational IT load at about 1.3 to 1.5 GW. The pipeline is two to three times existing supply.

Some of the announcements are huge. AirTrunk has committed $30 billion for 5 GW by 2030. Google is putting about $15 billion into an AI hub in Visakhapatnam. Microsoft has announced $21.5 billion in two phases.

But announced capacity and working capacity are very different things. Power, water and imported GPUs are the real limits. The question we must ask is the same one Gartner is asking the world: who will pay for all of this, and how much?

For us, there is an opportunity too. If the infrastructure gets built and the price of computing falls, the winners may be the people who build useful products on top of it. That is where Indian founders can win. We have the talent, the customers and the cost discipline.

10. What I’d do

I’m not a financial adviser, and this is not investment advice. But I care about the founders and investors who read this, so here is what I would think about.

If you are a founder or run a tech business:

1. Build products that earn money from the first customer. Price on outcomes, not on hype.
2. Don’t tie your product to one model provider. Keep the ability to switch.
3. Watch your cost per task closely. Today’s prices may be supported by investor money.
4. Fix one real workflow and measure it. An MIT study from 2025, based on more than 300 AI initiatives, found that most pilots showed no measurable profit impact. The ones that worked were focused and tied to real operations. The study’s method is debated, but the lesson holds.
5. If you depend on venture funding for AI, plan for a tighter funding market and keep more cash than you think you need.

If you are an investor:

1. Separate the shovel sellers from the shovel buyers. They are in very different positions when the cycle turns.
2. Check capex against operating cash flow every quarter. Meta’s was about 106% in the latest quarter.
3. Watch the credit market. CoreWeave and Oracle spreads often move before stocks do.
4. Ask what useful life the company assumes for its chips.
5. Don’t borrow to buy AI stocks. Size every position so that a 50% fall does not change your life.

My scorecard

I don’t want to write a post and walk away. So here is a test you can hold me to. I will revisit this post on 1st May 2027. These are the signals I’m watching:

• Any of the big four cutting capex guidance for two quarters in a row for reasons other than accounting or supply delays
• Nvidia missing or cutting its own guidance
• CoreWeave or Oracle credit spreads moving higher again, or a refinancing failing
• The first large, named write-down of AI data center assets
• A major AI lab raising money at a lower valuation than before

If none of these happen, I was too early or too harsh, and I will say so publicly. If several happen, you will have seen it coming.

One last thing

I wrote this with a lot of feeling because I know many of you are building, investing and hiring around this wave. I don’t want you hurt by it. I want you ready for it.

AI will matter for decades. But the companies that matter will be the ones that earn more than they spend. Be one of those.

Look at the numbers. Ask hard questions. And build something that pays for itself.

This post reflects my personal opinion and public data available on 1st November 2026. It is not investment advice. Please do your own research or speak to a licensed adviser before making financial decisions.