跳到正文
DEV Community· Md Kawsar·· 55 分鐘前AI 評分65

AI 泡沫之爭:2026 年會破裂,還是繼續膨脹?

The AI bubble debate: will it burst in 2026 or keep inflating?

AI 導讀

文章將 AI 泡沫分為兩部分:大型科技股的市盈率泡沫已回落,基建融資週期仍在擴張。五大超大規模雲端服務商承諾於 2026 年投入逾 $700 billion 資本開支;企業 AI 軟件收入約 $100 billion,與基建投入所隱含的收入需求相差約六比一。文章並指出,生產力效益尚未普遍反映,AI 大型 IPO 的定價可作為觀察市場需求和估值的壓力測試。

正文

The hyperscalers have committed more than $700 billion, and their free cash flow has dropped to a decade low. The question Sequoia posed at $200B in 2023 has grown considerably since. Here's how to read it.

On June 23, 2026, South Korea halted trading on the KOSPI for twenty minutes. Samsung and SK Hynix had lost 12% in a single morning, and the Nasdaq fell 2.2% that session. Micron Technology, up nearly 800% over the previous year on AI demand, dropped 13%.

The following Monday, Alphabet placed an order with Intel for more than three million tensor processors, to be delivered in 2028.

Both events fell in the same week, and together they capture the AI bubble debate in miniature. The market panics, confidence returns within days, and the spending commitments never budge.

I've followed this debate for more than a year, and one observation keeps coming back to me: most people are arguing about the wrong question. They ask "is there a bubble?" as if the answer were yes or no. It isn't. This spring, Capital Economics chief markets economist John Higgins wrote that one bubble has already burst: the inflated P/E ratio of Big Tech stocks. The sector's price-earnings ratio now sits at its lowest since the pandemic. What keeps growing is something else entirely, an infrastructure and private-market bubble that never shows up in public P/E ratios.

Two bubbles, then. One has already deflated, and the other is still filling. That distinction matters a great deal for what comes next.

A question that keeps getting bigger

In July 2024, Sequoia Capital partner David Cahn published "AI's $600 Billion Question." It updated his 2023 piece, "AI's $200 Billion Question," and its method is simple and brutal.

Start with Nvidia's data center revenue run rate. Double it, because GPUs make up roughly half the total cost of running a data center (the rest goes to energy, land, buildings, and backup generators). Then double it once more to account for the 50% gross margin that end customers need in order to justify their AI compute spending. What you get is the annual revenue the AI industry must earn for the infrastructure buildout to make economic sense.

In late 2023 that figure was $200B. By mid-2024 it was $600B. In 2026, with Nvidia's revenue substantially higher, the implied requirement is larger still.

And actual enterprise AI software revenue? Roughly $100 billion a year.

That's a six-to-one gap between what the infrastructure implies and what users actually pay. It isn't a rounding error. It's a structural mismatch that has only widened as spending has accelerated.

The five largest hyperscalers committed more than $700 billion in capital expenditure in 2026, roughly double the 2025 level, with about three-quarters tied directly to AI. The slide in their combined free cash flow is striking. In 2024 the group generated around $45 billion per quarter. By Q1 2026, that had dropped to roughly $4 billion. Amazon is projected to run negative free cash flow of up to $28 billion in 2026, and it has filed with the SEC indicating it may need to raise additional equity and debt.

"The most dangerous manias often form around technologies that really do work. Investors see the future coming, then rush too far ahead of it. They pay too much, too early, for the wrong opportunity."
Forbes analysis of the 2026 AI infrastructure cycle, May 2026
Some of the most cash-generative software companies in the world are starting to look like heavy industry. That's not a normal state for them, and it deserves to be taken seriously.

Why this isn't 1999, and why that still matters

Both camps reach for the dot-com comparison. Skeptics say this feels like 1999, while optimists insist it absolutely doesn't. Each side is partly right, and the specific differences are worth knowing.

Pets.com raised $82 million and earned $600,000 in revenue before collapsing, a 137-to-1 ratio of investment to revenue. OpenAI, by contrast, has more than $20 billion in annual recurring revenue, and Anthropic expects a profitable quarter in Q2 2026 with quarterly revenue above $10 billion. Today's AI companies have real products and real customers, and the Magnificent Seven together carry net margins above 25%, against an S&P 500 average of around 13%. Nvidia's net margin is roughly 53%.

Amundi, Europe's largest asset manager, ran a formal bubble detection test on current AI valuations. Its April 2026 finding: AI stocks do not show the explosive, non-stationary P/E dynamics that defined the dot-com era from 1997 to 2000. Back then, tech P/E ratios kept expanding even as the businesses underneath deteriorated. Today, AI stock P/E ratios have been falling since 2023, because earnings growth has outpaced price gains.

Credit markets are calmer as well. From 1998 to 2000, credit spreads widened ahead of the crash and gave investors an early warning. That stress signal hasn't appeared this cycle.

So why does the 1999 comparison keep coming back? Because a different historical parallel fits better.

Consider the British Railway Mania of the 1840s. Andrew Odlyzko, the financial historian whose research on that episode is the standard academic reference, found that even a calculation made at the time showed the projected railway mileage implied economically impossible returns. The railways were overbuilt by a margin that bankrupted most investors, yet the infrastructure they left behind powered British industrial growth for generations.

Jeff Bezos recently described the AI situation as an "industrial bubble": money will be wasted and companies will fail, but society inherits powerful infrastructure either way. That framing is honest and probably correct. What it doesn't tell you is whether your particular investment survives the correction.

The prisoner's dilemma no one can exit

Here's the part of this story I find genuinely uncomfortable.

The hyperscalers know the revenue gap exists. Their CFOs can run the same Sequoia calculation Cahn published in 2024. Yet they're spending more, not less. Why?

Writing on Substack in June 2026, Nick Wade named the mechanism precisely: it's a prisoner's dilemma. No single hyperscaler can safely cut AI spending without risking being overtaken by the others. The outcome is collective over-investment that every participant can see and none can stop on its own. Each player's rational choice adds up to an irrational result for all of them.

Amazon knows it may run $28 billion in negative free cash flow, and it's still spending. Google cut more than a third of its small-team managers in 2026. Its cloud revenue grew 63% to exceed $20 billion a quarter, yet the cuts continued. These companies are posting their best-ever cloud numbers while dismantling the teams that built those businesses, because the new race calls for different people and different capital.

Then there's xAI. SpaceX filed its S-1 in May 2026, putting xAI's financials on paper for the first time. In 2025, xAI posted $3.2 billion in revenue and a $6.35 billion operating loss, a margin of negative 200%. In Q1 2026 alone, it reported $818 million in revenue and a $2.47 billion operating loss, an annualized margin of negative 302%. Scaling is making the economics worse, not better. Each quarter at a larger scale produces a larger loss percentage.

"If you compare xAI to a traditional SaaS company, the financials look reckless. If that payoff doesn't happen, we are looking at the largest venture-funded correction in history."
Harrison Rolfes, senior research analyst at PitchBook
OpenAI has $20 billion in annual recurring revenue against $1.4 trillion in data center commitments, funded partly by debt. On the current trajectory, debt service alone is an enormous constraint on its ability to ever reach conventional profitability.

These aren't startups living on hope. They're companies with real revenue, real products, and financial structures that any other industry would be told are unsustainable.

What the productivity data actually says

The structural case for this investment is that productivity gains will eventually justify the spend. So let's look at the productivity numbers.

A National Bureau of Economic Research study published in February 2026 surveyed firms on AI's effect on workplace productivity. Ninety percent reported no measurable impact. Yet executives at those same firms projected that AI would raise productivity by 1.4% and output by 0.8%. That's a gap between stated belief and observed reality. A separate study discussed in Tom's Hardware found that 95% of enterprise generative AI deployments had no measurable effect on profit and loss.

Here, optimists and skeptics should be able to agree on at least one point: the gains are real in specific, narrow use cases and absent or invisible in most broad corporate deployments. The open questions are whether the narrow cases will grow large enough to cover the cost of the broad ones, and how long that will take.

Cognizant's Chief AI Officer Babak Hodjat told Nikkei Asia this year that companies need another six months to a year before they see real productivity gains. That's not a devastating claim. But the infrastructure is being built on the assumption that those gains arrive at a scale big enough to cover $700 billion in annual CapEx, and six more months of measurement doesn't settle that.

The IPO season as a stress test

One concrete event will do more to clarify the bubble question than any analyst report: the coming wave of AI mega-IPOs.

The SpaceX S-1 is already out, with xAI's financials attached, and IPOs from Anthropic and OpenAI are expected. They'll be the first time public market investors, rather than private backers, set a price on the leading AI companies. Private valuations come from funds that must deploy capital and have long time horizons. Public markets are less patient and far more specific about what they'll pay for current losses against current revenue.

When those filings land, here's what to watch:

The price-to-revenue multiple at IPO, and how it compares with what similar SaaS companies trade at
Whether the companies project profitability on a timeline of years or decades
How much of the business model depends on continued cheap access to compute (which isn't getting cheaper as CapEx rises)
Investor demand at pricing versus the initial indication. Strong oversubscription and a barely covered book tell very different stories.
Whether Anthropic's profitable quarter in Q2 2026 holds, or turns out to be a one-time condition
Fidelity noted in February 2026 that IPO activity was on pace for roughly 100 deals in 2026, close to the 25-year average and well below the nearly 400 of 1999. The IPO market isn't in 1999-style mania, yet the individual deal sizes in AI are unprecedented. Normal volume combined with extraordinary scale per deal is itself unusual.

How to think about this if you're not a hedge fund

Most people reading this are builders, product managers, engineers, or early-career professionals deciding whether to bet their next three years on AI skills. The bubble debate isn't very useful at that scale of decision. Here's what the data suggests for anyone making practical moves right now.

Separate the infrastructure story from the application story. Infrastructure investment may correct. Applications already earn revenue and will keep existing regardless. Cursor, Perplexity, and Github Copilot won't disappear because Nvidia's stock drops 30%, and the tools you learn now stay useful through any valuation cycle.
The productivity gap is real and will close, but unevenly. If 90% of firms see no impact, 10% are seeing real gains. Find those specific cases in your own domain and study them closely. The ROI is concentrated, not spread evenly.
Watch Amazon's free cash flow over the next two quarters. Negative free cash flow of $28B plus a filing to raise debt is an indicator worth tracking. If it improves alongside genuine AI-driven revenue growth, the investment thesis gains support. If it worsens while revenue stays flat, the prisoner's dilemma begins to break.
IPO pricings matter more than IPO headlines. When Anthropic files, ignore the valuation figure in the headline. Look at the P/S multiple, the trajectory of operating leverage, and whether institutional demand requires a meaningful discount from private-round pricing.

Where this leaves the debate

Part of the AI bubble has already burst. The public stock valuation bubble, measured by P/E ratios, peaked around October 2025 and has been deflating since. That correction happened quietly enough that most people never noticed it had a name. Capital Economics gave it one, and the financial press mostly moved on.

What hasn't corrected is the infrastructure financing cycle: the $700 billion in annual CapEx, Amazon's negative free cash flow, xAI's losses growing faster than its revenue, OpenAI's $1.4 trillion in data center commitments. That's a different animal from an inflated stock P/E ratio, and it'll resolve over a longer timeline and in a messier way.

The railway builders of the 1840s were wrong about the business models and right about the technology. Most of them lost their money, and the railways worked. I suspect the same is true here. The technology will work. Some of the companies building it won't survive the financing cycle. When the correction comes, the people who lose money will call it a bubble bursting, and the people who didn't will call it an inevitable market rationalization.

Neither label changes what you should do between now and then.

Enjoyed this read? Be sure to follow me for more stories that make people feel things. You can also connect with me on LinkedIn or Telegram @kawsarlog for more updates on my work.

I do freelance scraping projects, so if you need some data collected feel free to contact me on LinkedIn or email at hi@kawsarlog.com

來源:DEV Community · dev.to