Who’s paying for the AI boom? The AI bubble, explained

01

What just happened?

Three numbers ruined a lot of portfolios this month:

40%

China’s STAR 50 index of AI and chip stocks, down from its July high in under four months. About $480B gone.

12.6%

Alphabet’s fall on Oct 29, its worst day ever. Roughly $500B gone by the closing bell.

36%

CoreWeave’s share price fall in October, as of Oct 21. You’ll meet this cheerful company shortly.

A bubble is when prices climb far ahead of the money actually coming in. Everyone buys because everyone else is buying, right up until someone asks to see the receipts.

☕ At the canteen

A chai stall sells 100 cups a day, but people pay for shares in it as if it sells 10,000. The chai may be excellent. The price still has to meet the customers one day.

So is this the pop? Not yet. The US is in a sharp selloff, not a collapse: the Nasdaq closed Oct 30 about 8% below its record. When the dot-com bubble burst, it fell almost 80%. An 8% fall feels terrible in a trading app, but history has seen much worse.

Not a pop, at least not yet. But somebody has started reading the bill.

02

How does an AI chip get paid for?

AI runs on GPUs (the chips that train and run AI models), and they’re extremely expensive. Giants can pay from their own profits. Everyone else has to get creative.

Meet the neocloud: a company that mainly rents GPUs to others, such as CoreWeave. Think of a chip landlord who has to buy the whole building first and find the tenants later. Here’s how that building gets paid for.

Try it: follow the money

Neocloud

Nvidia

Lenders

At the canteenA new stall wants costly chai machines. Its savings: optimism.

In AIA neocloud wants Nvidia GPUs it can’t afford with the money it has.

1 / 6

Nvidia invests in a company. The company buys Nvidia chips. Nvidia’s sales go up. Nvidia invests in more companies. If you’ve ever split a bill with friends and somehow ended up owing yourself money, congratulations: you understand circular financing.

Is this a scam?

No. It’s legal and disclosed, and Nvidia’s CEO, Jensen Huang, calls the circular financing claims “ridiculous”. The real concern is narrower: when a supplier funds its own customers, demand can look stronger than it really is.

How big does this get?

Nvidia has guaranteed up to $105B of lease obligations on an OpenAI data center campus in Ohio. That’s a cap, not a cheque. It only applies if the tenant defaults, it builds up over nine construction phases (the first expected in Nvidia’s fiscal 2029) and reports say OpenAI must reimburse anything Nvidia pays. Still, when the chip seller co-signs the customer’s rent, it’s more than a chip seller.

Not a scam. Just a loop that works beautifully while everyone keeps paying.

03

Why did the loop start wobbling?

The loop needs three things: chips that earn more than they lose in value, customers who keep renting and lenders who keep lending. In October, the lenders got nervous, and they’re the ones holding the cash.

CoreWeave’s 9.625% bond traded around 80 cents on the dollar in October, pushing its yield (the yearly return a buyer gets at today’s price) to about 15%.
CoreWeave’s debt was $35.6B at the end of June, about $52B with leases. S&P rates it B+, which is officially called junk. Finance is many things, but subtle isn’t one of them.
Oracle was cut to BBB- by S&P on July 9, one notch above junk. S&P estimates that about half of Oracle’s $638B in contracted future revenue is tied to OpenAI. One customer, half the order book, no pressure.
Elon Musk said debt markets have “almost entirely shut off” for companies dependent on AI data centers, and that 200+ neoclouds could face “guppy-level death and consolidation”. That’s his opinion, posted while also saying SpaceX had “already closed $40 billion in debt financing”. Timing is everything.

☕ At the canteen

You lend a stall ₹100 for ₹9.6 a year. You get nervous and sell the loan for ₹80. The buyer earns 12% from that ₹9.6 on their ₹80. They also get the full ₹100 back when the loan ends in 2032, which adds about 3% a year. Together, that’s about 15%. Your panic became their yield, as long as the stall survives until 2032.

A falling bond price is lenders saying “we’re less sure”, in numbers.

04

Why are chips sitting unplugged?

Money is only half of the problem. A data center also needs electricity, and a grid doesn’t care about your funding round.

Microsoft: Satya Nadella said in late 2025 that Microsoft may have “a bunch of chips sitting in inventory that I can’t plug in”. Trillion-dollar company, shortage of sockets.
Virginia: Dominion Energy has roughly 70 GW of projects waiting to connect, and data centers wait seven years on average. For scale, a large nuclear reactor makes about 1 GW. Seven years is also longer than your degree, your internship and your first two job changes.
The PJM grid: PJM runs the largest US grid. Its independent market monitor says data centers drove $29.4B of capacity charges over four auctions, 46% of the total.

☕ At the canteen

The stall bought ten chai machines. The canteen has two plug points, and the extension board request is with the hostel warden.

You can borrow money quickly. You can’t borrow a power station quickly.

05

So who’s paying?

AI companies describe this as an investment. Here’s who is actually funding it.

Shareholders

Meta launched about $35B of share offerings on Oct 26 and plans about $60B more in non-voting shares, though that plan isn’t signed. Owning a piece of Meta now means owning a slightly smaller piece of Meta.

Cash reserves

Amazon expects to spend about $220B this year. Alphabet’s free cash flow was negative $9.07B last quarter, meaning more cash went out than came in. Even very deep pockets have a bottom.

Employees

About 60,000 Microsoft staff in the US, UK, Canada and India were offered voluntary buyouts, roughly a quarter of the workforce. Microsoft says AI isn’t the driver. The theory that job cuts fund AI spending is just that: a theory.

Power customers

Capacity charges feed into household bills across the PJM grid. You didn’t ask a chatbot for anything, but your electricity bill may still have opinions.

Savers

Meta’s Hyperion data center sits in a $27B venture, 80% owned by Blue Owl funds, with its debt off Meta’s balance sheet. The bonds went through PIMCO, the sort of name that turns up in other people’s savings.

Microsoft’s actual layoffs this year are much smaller than its buyout offer: about 4,800 roles in July, plus dozens in Israel in October. A buyout is voluntary, so please don’t let a thumbnail tell you 60,000 people were fired.

The bill didn’t disappear. It got split, and some people didn’t get a vote.

06

Are the profits real?

A chip is paid for today but used for years, so companies spread its cost over its expected life. That’s depreciation, and it’s normal. It also means someone has to estimate how long a chip stays useful, and that estimate changes the profit you see.

Try it: stretch a $30,000 chip’s life

Cost each year: $10,000
Total cost: $30,000
Yearly profit boost: $0

The cautious version. Profit looks modest, and the accountant sleeps well.

The total never changes. Only the timing does, but profit looks higher each year. In November 2025, Michael Burry, known for betting against US housing before 2008, estimated that big AI companies’ longer chip lifespans will understate depreciation by $176B across 2026 to 2028. That’s his estimate, and it’s disputed. Useful life is a management estimate that auditors review.

Same chip, same cash. Different story, depending on who holds the calculator.

07

So, has it popped?

Signs of cracking	Signs it’s holding
China’s STAR 50 down 40%	Google Cloud up 103% in Q3
CoreWeave yield about 15%	Azure up 48%
Oracle one notch above junk	Microsoft’s free cash flow at +$18.6B
Alphabet’s record one-day fall	Nasdaq about 8% below its record, not 80%

Our read: not popped as of Nov 1. People really are paying for AI: Google Cloud doubled and Azure grew 48%. The weak spot is companies that rely on debt to buy chips.

Past bubbles at least left useful railways and fibre behind. Chips age much faster, so what this one leaves behind may hold less value. A 2026 GPU won’t be a heritage monument.

The bubble hasn’t popped. The bill has arrived.

08

What does this mean for you?

1
Bubbles can pop while the technology survives. The web outlived the dot-com crash. Plenty of companies with “.com” in their name did not.
2
The bottleneck has moved. Power, cooling and infrastructure now matter as well as chips, which opens AI work to more branches of engineering. Electrical engineers, your moment has arrived.
3
Learn to read a headline number. Ask three things: is it cash, is it an estimate, and who said it?
4
Check viral statistics. “70% of the world’s internet traffic flows through Virginia” has been repeated for years. TeleGeography’s analysis says it doesn’t hold up. Confidence isn’t a citation.

2-minute recap

GPU: The chip that trains and runs AI. Currently more popular than you.
Neocloud: A company that mainly rents GPUs to others, such as CoreWeave.
Circular financing: A supplier funding customers who then buy from that supplier.
Bond yield: The yearly return from a loan at today’s price. Price falls, yield rises.
Junk rating: A credit rating saying lending to you is higher risk. Oracle is one notch above it.
Depreciation: Spreading a chip’s cost across the years you expect to use it.

09

Sources

Every number above comes from these. Primary filings first, then reporting.

The canteen analogies are simplified on purpose, and the $30,000 chip is an illustration. Market figures change daily. This page explains the news and isn’t financial advice.