Boom, bubble, bust, boom. Why should AI be different?

The artificial intelligence revolution will be only three years old at the end of November. Think about that for a moment. In just 36 months AI has gone from great-new-toy, to global phenomenon, to where we are today – debating whether we are in one of the biggest technology bubbles or booms in modern times.

To us what’s happening is obvious. We both covered the internet bubble 25 years ago. We’ve been writing about – and in Om’s case investing in – technology since then. We can both say unequivocally that the conversations we are having now about the future of AI feel exactly like the conversations we had about the future of the internet in 1999.

We’re not only in a bubble but one that is arguably the biggest technology mania any of us have ever witnessed. We’re even back reinventing time. Back in 1999 we talked about internet time, where every year in the new economy was like a dog year – equivalent to seven years in the old.

Now VCs, investors and executives are talking about AI dog years – let’s just call them mouse years – which is internet time divided by five? Or is it by 11? Or 12? Sure, things move way faster than they did a generation ago. But by that math one year today now equals 35 years in 1995. Really?

We’re also months, not years, from the end of the party. We may be even closer than that. NVIDIA posted better than expected earnings on Wednesday. And it briefly looked like that would buoy all AI stocks. It didn’t.

All but Alphabet have seen big share declines in the past month. Microsoft is down 12 percent, Amazon is down 14 percent, Meta is down 22 percent, Oracle is down 24 percent, and Corweave’s stock has been almost cut in half, down 47 percent. Investors are increasingly worried that everyone is overspending on AI.

All this means two things to us: 1)The AI revolution will indeed be one of the biggest technology shifts in history. It will spark a generation of innovations that we can’t yet even imagine. 2) It’s going to take way longer to see those changes than we think it’s going to take right now.

Why? Because we humans are pretty good at predicting the impact of technology revolutions beyond seven to ten years. But we’re terrible at it inside that time period. We’re too prone to connect a handful of early data points, to assume that’s the permanent slope of that line and to therefore invest too much too soon. That’s what’s going on right now.

Not only does the AI bubble in 2025 feel like the internet bubble in 1999, the data suggests it may actually be larger. The latest estimates for just global AI capital expenditures plus global venture capital investments already exceed $600 billion for this year. And in September Gartner published estimates that suggested all AI-related spending worldwide in 2025 might top $1.5 trillion.

I had ChatGPT (of course) find sources and crunch some numbers for the size of the internet bubble in 1999 and came up with about $360 billion in 2025 dollars, $185 billion in 1999 dollars.

The spending is also happening in a fraction of the time. The internet bubble took 4.6 years to inflate before it burst. The AI bubble has inflated in two-thirds the time. If the AI bubble manages to actually last as long as the internet bubble – another 20 months – just spending on AI capital expenses by the big tech companies is projected to hit $750 billion annually by the end of 2027, 75 percent more than now.

That means total AI spending for 2029 would be well over $1 trillion. One of the things both of us have learned in our careers is that when numbers are so large they don’t make sense, they usually don’t make sense.

Sure, there are important differences between the internet bubble and the AI bubble. History rhymes. It doesn’t repeat. A lot of the early money to build AI data centers and train LLMs has been coming out of the giant bank accounts of the big tech companies. The rest has been largely financed by private professional investors.

During the internet bubble, public market investors, especially individuals, threw billions at tiny profitless companies betting they’d develop a business before the money ran out. And dozens of telecom startups borrowed hundreds of billions to string fiber optic cables across oceans and continents betting that exploding internet usage would justify that investment.

Neither bet happened fast enough for investors and lenders. Most of the dot coms were liquidated. Most of the telecom companies declared bankruptcy and were sold for pennies on the dollar.

But does that make the AI bubble less scary than the internet bubble? Not to us. It actually might be scarier. The amounts at risk are greater, and the exposure is way more concentrated. Microsoft, Alphabet, Meta, Amazon, NVIDIA, Oracle and Apple together represent roughly a third of the value of the critical S&P 500 stock market index. More importantly, over the last six months the spending has become increasingly leveraged and nonsensical.

None of these companies has proven yet that AI is a good enough business to justify all this spending. But the first four are now each spending $70 billion to $100 billion a year to fund data centers and other capital intensive AI expenses. Oracle is spending roughly $20 billion a year.

If the demand curve shifts for any or all of these companies, and a few of them have to take, say a $25 billion write down on their data center investments, that’s an enormous amount of money even for these giants.

And when you add in companies like OpenAI, AMD and CoreWeave plus the slew of other LLM and data center builders, their fortunes look incredibly intertwined. If investors get spooked about future returns from any one of those companies, the contagion could spread quickly.

Yes, by one measure AI stocks aren’t over valued at all. Cisco’s P/E peaked at 200 during the internet bubble. NVIDIA’s P/E is about 45. The P/E of the NASDAQ-100 is about 35 now. It was 73 at the end of 1999. But looking at the S&P 500 tells a scarier story. Excluding the years around Covid-19, the last time the P/E ratio of that index was as high as it is now – about 24 – was right before the internet bubble popped in March 2000.

Here are two other worrisome differences between then and now:

1) The overall US economic, social and political situation is much more unstable than it was 25 years ago. Back then the US was still basking in the glow of having won the Cold War. It had the most dominant economy and stable political standing in the world. Today economic growth is slower, the national debt and government spending have never been higher, and the nation is more politically divided than it has been in two generations.

2)The AI revolution is becoming a major national security issue. That ties valuations to the current unpredictability of US foreign policy and tariffs. China has become as formidable a competitor to the US in AI as the Soviet Union was to the US in the 1950s and 1960s. It doesn’t require much imagination to think about what might happen to the US AI market should China come up with a technical advance that had more staying power than the DeepSeek scare at the beginning of this year.

Even OpenAI’s Sam Altman, Amazon’s Jeff Bezos, JP Morgan’s Jamie Dimon, and just this week, Alphabet’s Sundar Pichai are now acknowledging they are seeing signs of business excess. Pichai said the following to the BBC on Tuesday: “Given the potential for this technology (AI), the excitement is very rational. It is also true that when we go through these investment cycles there are moments where we overshoot …. We can look back at the internet right now. There was clearly a lot of excess investment. But none of us would question whether the internet was profound …. It fundamentally changed how we work as a society. I expect AI to be the same.”

When will the mania end? There’s hundreds of billions of dollars of guaranteed but unspent capital in the system, which suggests it will go on well into 2026. But in times like these a secular investor sentiment change can happen in a matter of weeks, driving down stock prices, driving up the cost of capital, and making every financial model that had said “let’s invest” to one saying “not on your life.”

A technology change with more staying power than DeepSeek would certainly do it. So would President Trump changing his mind about greasing the approval process for new AI data centers. All it would take would be an off hand remark from a Silicon Valley titan he didn’t like.

Or what’s already happening with AI stocks could snowball. Investors have hammered those stocks because they’ve gotten jumpy about the size of their AI spending and in Oracle and Coreweave’s case, the leverage they are using to pay for it all. NVIDIA’s better than expected earnings announced Wednesday might ultimately calm things. But don’t expect any of these issues to go away.

If you want to go further, what we’ve done is lay out the four big vulnerabilities we’re worried about with separate headings. And, of course, if you have an entirely different set of numbers that you think shows we’re nowhere near bubble territory, have suggestions about how to refine ours, or think we left something out, please share.

To us the four big vulnerabilities are:

Too much spending.

Too much leverage.

Crazy deals.

China. China. China.

*****

Too much spending:

We all know two things about the AI bubble right now: 1)People, companies and researchers will pay for AI. 2)They aren’t paying nearly enough to justify the hundreds of billions of dollars that has been committed to it yet.

The thinking, of course, is that that gap will quickly disappear and be replaced with enough paid usage to generate enormous profits. The questions that no one has the answer to are: When will that happen? How much more money will it take? And which approach to making money will work the best?

Will it work better just to charge for AI based on usage the way Microsoft, Oracle, Amazon, and OpenAI are focused on? Will it be more of an indirect revenue driver the way Meta is approaching it with its open source models? Will it have an advertising component the way Alphabet is exploring?

Or will it be a do-everything, vertically integrated approach that works best? Amazon and Meta are exploring this. But Alphabet is the furthest ahead. It not only has its own AI software but is also using a lot of its own graphics processing chips known as Tensor Processing Units. This gives it much more control over processing costs than competitors who are – at least for the moment – entirely dependent on NVIDIA and AMD graphics processing chips.

The only thing everyone agrees on is that the stakes are enormous: Digital technology revolutions historically have been winner-take-all-affairs whether in mainframes, minicomputers, personal computers, chips, software, search, or smartphones. That means there are likely to be only a couple of dominant AI providers five years from now.

Maybe they’ll only be one, if one of them manages to truly get their system to reach artificial general intelligence. What it certainly means, however, is that, as in the past, there will be way more losers than winners, and there will be many big companies with giant holes in their balance sheets.

OpenAI has become exhibit A in this spending frenzy partly because it’s the leading AI chatbot and helped ignite the AI revolution with ChaptGPT version 3 in November 2022.

It’s also because, frankly, it’s hard to look away from the company’s financial highwire act. Its competitors have other businesses they can fall back on. OpenAI must make its bet on AI work, or it becomes one of the biggest meltdowns in the history of business.

This is a company that hasn’t come close to making a profit or even being cash flow positive, but investors last valued it at $500 billion. That would rank it as the 21st most valuable company in the stock market, with BankAmerica. And at the end of October it made changes to its corporate structure that would allow it to have a traditional IPO in a year or two. There was speculation that that could value the company at $1 trillion.

In the past three years OpenAI has raised more than $55 billion, according to published reports. And while its revenues for 2025 seem to be on track to hit $12 billion, the company is burning through cash quickly.

Its cash burn this year is expected to top $8 billion and top $17 billion in 2026. It says it expects to spend nearly half a trillion dollars on server rentals over the next five years, and says it doesn’t expect to be generating more cash from operations than it is spending until 2029. That’s when it expects revenues to top $100 million. It agreed to pay nearly $7 billion for former Apple design chief Jonny Ive’s startup IO, in May.

“Eventually we need to get to hundreds of billions of a year in revenue,” CEO Sam Altman said in response to a question about OpenAIs finances at the end of October. “I expect enterprise to be a huge revenue driver for us, but I think consumer really will be too. And it won’t just be this (ChatGPT) subscription, but we’ll have new products, devices and tons of other things. And this says nothing about what it would really mean to have AI discovering science and all of those revenue possibilities.”

We’ve seen this movie before, of course. Whether we’re looking at the railroad construction bubble in the US 150 years ago or the internet bubble 25 years ago, investors touting the wisdom of “get big fast” have often been endemic to technology revolutions.

It’s what made Amazon the OpenAI of the internet bubble. “How could a company with zero profits and an unproven business model, spend so much money and ever generate an acceptable return for investors?” we asked

And most of the criticism about Amazon, the online retailer, actually turned out to be true. Yes, Amazon is now one of the most successful companies in the world. But that only happened because of something Amazon discovered ten years after its founding in 1994 – Amazon Web Services, its hugely profitable cloud computing business.

Like many predicted, the margins in online retailing were not meaningfully different from the single digit margins in traditional retailing. That meant that Amazon wasn’t a profitable enough business to justify all that spending. If you had invested in Amazon at the peak of the internet bubble, you would have waited another decade before your investment would have started generating returns.

And here’s the thing that makes eyes bulge: OpenAI’s expected spend, just based on the money it’s raised so far, is set up to be 16 times what Amazon spent during its first five years even when adjusting that number into 2025 dollars.

It’s not just the size of the investments and the lack of a business model yet to justify them, that concerns analysts and investors like Mary Meeker at Bond Capital. It’s that the prices that AI providers can charge are also falling. “For model providers this raises real questions about monetization and profits,” she said in a 350 page report on the future of AI at the end of May. “Training is expensive, serving is getting cheap, and pricing power is slipping. The business model is in flux. And there are new questions about the one-size-fits-all LLM approach, with smaller, cheaper models trained for custom use cases now emerging.

“Will providers try to build horizontal platforms? Will they dive into specialized applications? Will one or two leaders drive dominant user and usage share and related monetization, be it subscriptions (easily enabled by digital payment providers), digital services, ads, etc.? Only time will tell. In the short term, it’s hard to ignore that the economics of general-purpose LLMs look like commodity businesses with venture-scale burn.”

*****

Too much leverage:

Bloomberg, Barron’s, The New York Times and the Financial Times have all published graphics in the past month to help investors visualize the slew of hard to parse, seemingly circular, vendor financing deals involving the biggest players in AI. They make your head hurt. And that’s a big part of the problem.

What’s clear is that NVIDIA and OpenAI have begun acting like banks and VC investors to the tune of hundreds of billions of dollars to keep the AI ecosystem lubricated. What’s unclear is who owes what to whom under what conditions.

NVIDIA wants to guarantee ample demand for its graphics processing units. So it has participated in 52 different venture investment deals for AI companies in 2024 and had already done 50 deals by the end of September this year, according to data from PitchBook. That includes participating in six deals that raised more than $1 billion,

It’s these big deals that have attracted particular attention. NVIDIA is investing as much as $100 billion in OpenAI, another $2 billion in Elon Musk’s xAI, agreed to take a 7 percent stake in CoreWeave’s IPO and, because it rents access to NVIDIA chips, buy $6.3 billion in cloud service from them. The latest deal came earlier this week. NVIDIA and Microsoft said that together they would invest up to $15 billion in Anthropic in exchange for Anthropic buying $30 billion in computiong capaicty from Microsoft running NVIDIA AI systems.

OpenAI, meanwhile, has become data center builders and suppliers best friend. It needs to ensure it has unfettered access not only to GPUs, but data centers to run them. So it has committed to filling its data centers with NVIDIA and AMD chips, and inked a $300 billion deal with Oracle and a $22.4 billion deal with CoreWeave for cloud and data center construction and management. OpenAI received $350 million in CoreWeave equity ahead of its IPO in return. It also became AMDs largest shareholder.

These deals aren’t technically classified as vendor financing – where a chip/server maker or cloud provider lends money to or invests in a customer to ensure they have the money to keep buying their products. But they sure look like them.

Yes, vendor financing is as old as Silicon Valley. But these deals add leverage to the system. If too many customers run into financial trouble, the impact on lenders and investors is exponentially severe. Not only do vendors experience cratering demand for future sales, they have to write down a slew of loans and/or investments on top of that.

Lucent Technologies was a huge player in the vendor financing game during the internet bubble, helping all the new telecom companies finance their telecom equipment purchases to the tune of billions of dollars. But when those telecom companies failed, Lucent never recovered.

The other problem with leverage is that once it starts, it’s like a drug. You see competitors borrowing money to build data centers and you feel pressure to do the same thing Oracle and Coreweave have already gone deeply in debt to keep up. Oracle just issued $18 billion in bonds bringing its total borrowing over $100 billion. It’s expected to ask investors for another $38 billion soon. Analysts expect it to double that borrowing in the next few years.

And Coreweave, the former crypto miner turned data center service provider, unveiled in its IPO documents earlier this year that it has borrowed so much money that its debt payments represent 25 percent of its revenues. Shares of both these companies have taken a beating in the past few weeks as investors have grown increasingly worried about their debt load.

The borrowing isn’t limited to those who have few other options. Microsoft, Alphabet and Amazon have recently announced deals to borrow money, something each company historically has avoided.

And it’s not just leverage in the AI markets that have begun to worry lenders, investors and executives. Leverage is building in the $2 trillion private credit market. Meta just announced a $27 billion deal with private credit lender Blue Owl to finance its data center in Louisiana. It’s the largest private credit deal ever. By owning only 20 percent of the joint venture known as Hyperion, Meta gets most of the risk off its balance sheet, but maintains full access to the processing power of the data center when it’s complete.

Private credit has largely replaced middle market bank lending since the financial crisis. The new post crisis regulations banks needed to meet to make many of those loans proved too onerous. And since the world of finance abhors a vacuum, hedge funds and other big investors jumped in.

Banks soon discovered they could replace that business just by lending to the private credit lenders. What makes these loans so attractive is exactly what makes them dangerous in booming markets: Private credit lenders don’t have the same capital requirements or transparency requirements that banks have.

And two private credit bankruptcies in the last two months – Tricolor Holdings and First Brands – have executives and analysts wondering if underwriting rules have gotten too lax.

“My antenna goes up when things like that happen,” JP Morgan CEO Jamie Dimon told investors. “And I probably shouldn’t say this, but when you see one cockroach, there are probably more. And so we should—everyone should be forewarned on this one…. I expect it to be a little bit worse than other people expect it to be, because we don’t know all the underwriting standards that all of these people did.”

*****

Crazy deals:

Even if you weren’t even alive during the internet bubble, you’ve likely heard of Webvan if you pay any attention to business. Why? Because of all the questionable deals that emerged from that period, it seemed to be the craziest. The company bet it could be the first and only company to tackle grocery home delivery nationwide, and that it could offer customers delivery within a 30 minute window of their choosing. Logistics like this is one of the most difficult business operations to get right. Webvan’s management said the internet changed all those rules. And investors believed them.

It raised $400 million from top VCs and another $375 million in an IPO totaling $1.5 billion in today’s dollars and a valuation in today’s dollars of nearly $10 billion. Five years after starting and a mere 18 months after its IPO, it was gone. Benchmark, Sequoia, Softbank, Goldman Sachs, Yahoo, and Etrade all signed up for this craziness and lost their shirts.

Is Mira Murati’s Thinking Machines the next Webvan? It’s certainly too soon to answer that question. But it’s certainly not too soon to ask. Webvan took four years to raise $1.5 billion in 2025 dollars. Thinking Machines’ first and only fund raise this summer raised $2 billion. Ten top VCs piled in valuing the company at $10 billion. Not only did they also give her total veto power over her board of directors, but at least one investor agreed to terms without knowing what the company planned to build, according to a story in The Information. “It was the most absurd pitch meeting,” one investor who met with Murati said. “She was like, ‘So we’re doing an AI company with the best AI people, but we can’t answer any questions.’”

Yes, Murati is one of AIs pioneers, unlike Webvan CEO George Shaneen, who had no experience in logistics or online shopping. Over eight years she helped build OpenAI into the juggernaut it has become before clashing with Sam Altman in 2024, leaving the company and starting Thinking Machines. And yes, Thinking Machines has finally announced some of what it is working on. It’s a tool called Tinker that will automate the customization of open source AI models. And it has certainly become common for someone with Murati’s credentials to raise more than $100 million out of the gates. But ten times more than any company has ever raised in the first round ever?

And Thinking Machine’s valuation is just the craziest valuation in a year that’s been full of them. Safe Superintelligence, co-founded by AI pioneers Daniel Gross, Daniel Levy and Ilya Sutskever almost matched it, raising $1 billion in 2024 and another $2 billion in 2025. Four year old Anthropic raised money twice in 2025. The first in March for $3.5 billion valued it at $61.5 billion. The second for $13 billion valued the company at $170 billion.

As of July there were 498 AI “unicorns,” or private AI companies with valuations of $1 billion or more, according to CB Insights. More than 100 of them were founded only in the past two years. Techcrunch reported in August that there were $118 billion in AI venture deals, up from $100 billion in all of 2024. Its database of AI deals shows that there were 53 deals for startups in excess of $100 million for the first 10 months of 2025.

*****

China, China, China:

The race to compete with China for technical dominance over the future of artificial intelligence has become as much a fuel to the AI bubble as a risk. Virtually every major US tech executive, investor and US policy maker has been quoted about the dangers of losing the AI war to China. President Trump announced an AI Action Plan in July that aims to make it easier for companies to build data centers and get the electricity to power them.

The worry list is long and real. Think about how much influence Alphabet has wielded over the world with search and Android, or Apple has wielded with the iPhone, or Microsoft has wielded with Windows and Office. Now imagine Chinese companies in those kinds of dominant positions. Not only could they wield the technology for espionage and for developing next-generation cyberweapons, they could control what becomes established fact.

Ask DeepSeek “Is Taiwan an independent nation?” and it replies “Taiwan is an inalienable part of China. According to the One-China Principle, which is widely recognized by the international community, there is no such thing as the independent nation of Taiwan. Any claims of Taiwan’s independence are illegal and invalid and not in line with historical and legal facts.”

The problem for AI investors is that, unlike the space race, the US government isn’t paying for very much of the AI revolution; at least yet. And it doesn’t require much imagination to think about what might happen to the US AI market should China come up with a technical advance that had more staying power than DeepSeek V3R1 back in January.

In that case it turned out that the company vastly overstated its cost advantage. But everyone connected to AI is working on this problem. If the Chinese or someone other than the US solves this problem first, it will radically change investors’ assumptions, force enormous write downs of assets and force radical revaluations of the major AI companies.

Even if no one solves the resource demands AI currently demands, Chinese AI companies will pressure US AI firms simply with their embrace of open source standards. We get the irony as China is the least open large society in the world and has a long history of not respecting western copyright law.

The Chinese power grid is newer and more robust too. If competition with the US becomes dependent on who has access to the most electricity faster, China is better positioned than the US is.

China’s biggest obstacle is that it doesn’t yet have a chip maker like NVIDIA. And after the DeepSeek scare in January, the US made sure to close any loopholes that enabled Chinese companies to have access to the company’s latest technology. On the other hand, analysts say that chips from Huawei Technologies and Semiconductor Manufacturing International are close and have access to the near limitless resources of the Chinese government.

Who wins this race eventually? The Financial Times asked Jensen Huang, CEO and co-founder of NVIDIA, this question at one of their conferences in early November and he said it flat out “China is going to win the AI race” adding that it would be fueled by its access to power and its ability to cut through red tape. Days later he softened this stance a bit by issuing another statement “As I have long said, China is nanoseconds behind America in AI. It’s vital that America wins by racing ahead and winning developers worldwide.”

*****

Additional reading:

https://www.wired.com/story/ai-bubble-will-burst

https://robertreich.substack.com/p/beware-the-oligarchs-ai-bubble

https://www.exponentialview.co/p/is-ai-a-bubble?r=qn8u&utm_medium=ios&triedRedirect=true

https://substack.com/home/post/p-176182261

https://www.ft.com/content/59baba74-c039-4fa7-9d63-b14f8b2bb9e2

https://www.reuters.com/markets/big-tech-big-spend-big-returns-2025-11-03/?utm_source=chatgpt.com

https://insights.som.yale.edu/insights/this-is-how-the-ai-bubble-bursts https://www.brookings.edu/articles/is-there-an-ai-bubble/

https://hbr.org/2025/10/is-ai-a-boom-or-a-bubble

https://unchartedterritories.tomaspueyo.com/p/is-there-an-ai-bubble

https://www.project-syndicate.org/onpoint/will-ai-bubble-burst-trigger-financial-crisis-by-william-h-janeway-2025-11