评论文章质疑AI泡沫论是否成立

The AI bubble is a speculative concept that the AI industry is extremely over-valued. If true, the value of most companies involved in AI (several of whom are worth well over a trillion dollars) will collapse, devastating the entire economy. If you check any online discussion about the financial situation of the AI industry, it's essentially taken for granted that the bubble exists and is going to pop any day now. I don't think that this is a reasonable assumption, though. Nobody can see the future, but I think that there's a pretty realistic chance that the meteoritic growth of AI companies in the past few years is justified, and that they have even more room for growth.

To say that AI companies are spending an absurd amount of money is an understatement. The debt and spending commitments for AI companies are handled through complex vehicles that make it difficult to tell exactly how much their expenses actually are (a serious problem, and one of the most salient points in favour of AI being a bubble), but they've committed to spending at least 3 trillion dollars so far. It's also been estimated that AI companies have already spent 1.9 trillion, bringing the overall expense of the AI industry to around 5 trillion dollars. This is unprecedented levels of investment, and AI just can't be profitable if this level of spending is maintained. However, there's no reason for why this level of spending has to be maintained in the first place. The reason why AI companies are spending so much money is to rollout data centres to run the AI in the first place. These expenses are one-off purchases, not regular operating expenses. Once all the data centres are built, AI spending is going to plummet.

Outside of these expenses, AI doesn't cost that much to run. We don't have the full numbers for the entire industry here, but Anthropic's S-1 filing says that they make $2.08 for every $1 of compute. In other words, once you ignore the cost of building infrastructure, AI is very profitable. Crucially, the S-1 filing indicates that Anthropic has started to make more than their compute and infrastructure investments combined, which cuts against the idea that the cost of building data centres is inherently unsustainable. In a pinch, AI companies can also cut costs by no longer allowing people to use their products for free and temporarily suspending expensive AI research. This can be implemented instantly, whereas other kinds of expenses like leases and spending commitments are a lot harder to cut.

The only serious long-term expenses to contend with here is the massive amount of debt that's been taken out to pay for these data centres in the first place. Like I said before, it's hard to really figure out these commitments because of the complex way that they're issued, but we can do some back of the envelope math. If the AI industry has to pay off 5 trillion dollars over the course of 20 years, that's 250 billion dollars a year. For context, Alphabet (parent company of Google and Youtube) reported earning almost 162 billion dollars of revenue just in 2019. Adjusted for inflation, that's an even more impressive 213 billion dollars. Because this is very rough math, the total needed revenue might come in to well over 250 billion dollars a year, but the point stands. Even though this is an enormous expense, paying it off isn't unthinkable with the data we have so far. Some of the debt is in convertible notes as well, which can be repaid through shares instead of cash. In the first 6 months of 2026, Anthropic has recorded 72 billion dollars of loss in the form of shares that are due to lenders. The cost is only so high because the valuation has skyrocketed to an estimated 1 trillion dollars, which means that Anthropic has more than enough equity to make lenders whole.

In their S-1 filing, Anthropic records their revenue going from 41 million in 2023, 386 million in 2024, 4.585 billion in 2025, and 16.3 billion in the first 6 months of 2026. That's an insane increase in revenue, and it's safe to assume that other major AI companies like OpenAI are also seeing similar increases as well. At some point, the rate of increase has to level off, but that doesn't have to happen anytime soon. When I look at the current capacity of AI, it really feels like there's so much potential for frontier models to be used that just aren't being applied yet. Most (or maybe even all) programmers and mathematicians are worse at their respective field than an AI swarm equipped with a good harness, and a good deal more expensive as well. AI can't replace lawyers and doctors, but they can automate their paperwork and free up time for them to do other things. Outside of generative AI, self-driving cars are becoming a reality, with some evidence to suggest that they're even better than human drivers. AI adoption takes time, but there's no reason to think that AI usage isn't going to massively skyrocket, even if AI doesn't get any better than it is today (and it certainly will get better). In other words, there's massive unmet demand for AI companies to profit from.

There's a weird contradiction you see from AI skeptics where they're afraid of how AI will change the world, but also seem convinced that AI will only ever produce slop that nobody would ever pay money for. You can't really have it both ways here. It's completely valid to be afraid of AI taking your job, but money is going to change hands when that happens. I don't know exactly what the future will look like or how long it'll take for mass AI adoption to really happen, but when I imagine how AI will change the world, it's hard for me to see AI consumption not being 10x what it is now.

In 2023, a memo from Google called "We Have No Moat, And Neither Does OpenAI" was leaked, which outlined how the open source community was keeping pace with AI labs. More broadly, the memo rejected the idea that any company can really have a durable advantage when it comes to AI development. It would certainly be profitable if they could, but there's already so much information and research out there that nobody can ever hope for their frontier models to stay on the frontier for long. That means that there will never be a point where a company wins the AI wars, just nonstop fighting as each company in the industry tries to one-up the others. You know what other company didn't have a moat and found itself in a war to sell a product that its competitors could easily copy? Popeyes during the chicken sandwich war. I know that sounds goofy, but from a business perspective, it's remarkably similar. None of the restaurants involved in the chicken sandwich war had an enduring competitive advantage over any of the others, and all of them had to deal with potential customers choosing to just cookat home instead, but it was still highly profitable for everybody involved. Simply saying that a company doesn't have a moat isn't enough to say that the company can't make money.

In theory, an AI company has to be worried about their users switching over to a competitor or using an open source model instead. In practice, this isn't a major concern. For starters, the average person isn't going to use a locally hosted model, ever. I know that it's a common hobby in tech circles, but normal users don't want to have to deal with the hassle or expense of setting up something that's going to perform at most the exact same as ChatGPT. That means that in principle, there's only going to be a few models that the vast majority of people will ever use. People can theoretically switch from model to model, but people tend to settle in with a product and develop brand loyalty. For example, I pretty much only use ChatGPT, not because I have any really reason to, but because it was the first LLM that I ever used. This partially offsets the risk of losing customers to a different product. Another way that companies can make money without having a moat is through institutional lock-in. Once a large corporation has integrated a particular AI into their day to day operation, transitioning away isn't costless. They may even have contracts agreeing to buy a certain amount of tokens, which binds them to keep using the AI instead of transitioning.

A potential way for a bubble to pop is for the companies involved to run out of capital to spend while they grow, causing them to fail. With the enormous amount of churn and debt that AI companies have, this is a serious risk to consider. Even so, I think that it's premature to worry about these companies failing to raise capital when most of them haven't even had an IPO yet. When SpaceX (which is partially an AI company than a space company, despite the name) went public, they raised more than 80 billion dollars. Besides that, many AI companies haven't issued public debt yet, which is a major way to get loans if private lenders decide that they don't want to put more of their capital into AI. The ones that have issued public debt, like Alphabet and Meta, have excellent bond ratings. If they need more capital, they won't struggle to find people to give them a loan.

Of course, investors aren't the only way to get capital. Once AI companies become profitable (as they already seem to be), they can use those profits to fund growth and pay off debts. The idea that AI companies are forced to constantly look for new investors to keep their doors open is probably a few months antiquated. From how it looks from Anthropic's S-1, the AI world may be past the horizon when it came to taking risky bets and can now focus on reaping the rewards.

When a bubble happens, it's usually preceded by the companies involved reaching obviously inflated evaluations, which makes the burst obvious in hindsight. With public AI companies, we don't see that happening. Alphabet, Meta, and NVIDIA are all AI companies that have seen their market capitalization explode over the past few years, but all of them have stable P/E ratios in their 20s. If the increase in their valuations were irrational (or even just in expectation of future growth as opposed to current earnings), you would expect the P/E ratios to go up. Instead, their valuations look exactly what we would expect from a healthy trillion dollar company. When the dot-com bubble happened, it was largely driven by companies that couldn't turn a profit or had P/E ratios in the hundreds. I'm not going to pretend that I'm an expert on financial bubbles, but I've never heard of a bubble where the companies involved had such a modest valuation based on their earnings, even if their earnings (and therefore their evaluations) are enormous.

A potential way that this could be a bubble is if AI proves to be a fad. A lot of this rise in valuations has been because of increased demand for GPUs and data centre utilization, which is downstream of the demand for AI. We might imagine that the AI labs are propped up by an enormous amount of invested capital, and this capital ends up propping up the valuations for companies that make or rent out GPUs. However, this doesn't seem to be happening. Alphabet has a full-stack AI supply chain, where they create ASICs for AI, train their own AI, integrate AI into their products, and rent out compute to customers. Their prospectus helpfully divides their revenue streams into Google Cloud for their rented compute and Google Services, which makes it easy to tell if Alphabet's increased earnings are mostly because of their data centres. The answer is no. Google Cloud revenue is growing more in terms of percentage, but Google Services is growing far more in absolute terms. Specifically, Google Services revenue grew by 37.8 billion dollars for a 12% increase, while Google Cloud revenue grew by 15.5 billion dollars for a 36% increase. Alphabet is definitely benefiting from renting out their data centres, but it's not the main reason why their earnings have skyrocketed. Meta doesn't rent out their data centres at all, but their pros