The AI IPO Wave Is The Dot-Com Bubble’s Evil Twin. Anthropic’s Delay Just Proved It.

By Anurag Gurtu

When a company delays its IPO, or first sale of shares to the public, to show investors better financial results, that is a strategy. But when New York City's pension fund publicly asks you to delay the sale because of AI safety concerns, your main rival's CEO calls his own IPO "ill-advised" and puts it off until at least next year, tech stocks fall because investors fear open-weight AI models, and your own prospectus warns that the product you sell could create "existential risks to humanity," something more serious is happening.

The market is sending you a message.

Anthropic's numbers are extraordinary. Its revenue in the third quarter of 2026 was reportedly $16.8 billion, up 45% from the previous quarter. At that rate, its yearly revenue would be about $67 billion.

What the S-1 actually says

Here are the numbers from Anthropic's leaked prospectus, a document that explains the company and its finances to potential investors:

• $4.6 billion: Anthropic's total revenue for 2025.
• $7.33 billion: The amount Anthropic spent on computing power and infrastructure in 2025 alone, which was more than all the money it earned that year.
• $42 billion: Anthropic's reported net loss in 2025. During its investor presentations, the company will quickly explain that $34 billion of this loss came from a noncash accounting charge. This was an adjustment to the estimated value of financial agreements that could later turn into company shares, rather than an actual cash payment.
• $518 billion: The minimum amount Anthropic expects to spend over 10 years on AI infrastructure with six partners, including Google, Microsoft and Amazon, according to its confidential IPO prospectus reviewed by Reuters. About 80% of this total must be paid even if Anthropic changes its plans or does not use the resources.
• $2 trillion: The company is still aiming for an IPO valuation of $2 trillion in November. That would value the company at 435 times its 2025 revenue.

There is one more number from the S-1 that the company may not emphasize during its investor presentations: Two unnamed customers accounted for one-quarter of Anthropic's revenue in 2025.

Three signals the market is sending

The IPO delay is the biggest headline, but it is not the only sign investors should pay attention to.

Signal one: New York City Comptroller Mark Levine. Last week, the official who oversees New York City's pension funds urged Anthropic to stop its IPO plans for now. The funds reportedly have about $7 million invested in Anthropic today. They could have much more at stake if Anthropic joins the Russell 3000 stock index after going public. Levine sent a letter to Anthropic CEO Dario Amodei asking the company to put its IPO on hold.

Levine pointed to the same dangers that Anthropic describes in its own prospectus: AI systems that can act more independently may behave in unexpected ways, cause security problems and, in the document's words, pose "existential risks to humanity."

Signal two: Sam Altman delays OpenAI's IPO. OpenAI has said it will not go public in 2026. CEO Altman called an IPO at this time "ill-advised" because of AI safety concerns. Part of this message may be a business strategy to make OpenAI look like the more responsible company. But it is still an important sign when the head of the world's second-most valuable AI company looks at the current situation and decides to wait.

Signal three: The chip selloff. Last week's $2 trillion fall in chip stocks was partly driven by worries about the cost of building AI infrastructure. A forecast from Gartner helps explain these worries. By 2030, Gartner estimates that AI data centers could bring in at most $3.6 trillion in revenue, while $8.5 trillion would be needed to cover planned investments, borrowing costs and a 12% return on investment.

The open-weight problem

Here is the basic problem that even strong revenue growth in the third quarter cannot solve: Anthropic's $2 trillion valuation depends on its closed AI models continuing to charge more than open-weight alternatives. But that price advantage is steadily shrinking.

Open-weight AI models are putting pressure on the business model behind high valuations for closed models. In 2026, business use of open-weight models rose from 23% to 67%, and companies reported saving 70% to 90% on costs. DeepSeek R1 delivers performance close to the best available models on most business tests and costs just 7 cents per million input tokens. Independent analysis estimates that Anthropic charges about 18 times more for similar input tokens and 23 times more for output tokens.

DeepSeek makes its model weights free to download. This lets companies run the model on their own systems without signing an API contract with a model provider or paying for each token. Reports also suggest that the performance gap between open models and closed models shrinks within six months of a new leading model being released.

In this situation, New Constructs, the research firm that called Anthropic's planned IPO "an unprecedented test of investor gullibility," calculated what it would take to justify a $2 trillion valuation. Anthropic would need to achieve a 10% operating profit margin after taxes right away and maintain it for 10 years, while growing revenue at an extraordinary rate. By 2035, it would need to make $358 billion a year in operating profit after taxes.

Why this is worse than 1999

The dot-com bubble had one real excuse: Nobody knew what the internet would become. The AI IPO wave of 2026 does not have that excuse.

According to Gartner's forecast, worldwide spending on AI is expected to reach $2.59 trillion in 2026, up 47% from the previous year. Yet a July 2025 report from MIT NANDA estimated that 95% of the organizations it studied were getting no return from generative AI. Meanwhile, KPMG said in its own report earlier this year that only 7% of the leaders surveyed said they had shown that AI was delivering a return on investment.

This is the situation in which Anthropic is asking investors to value the company at 435 times its 2025 revenue. The company also has $518 billion in infrastructure spending commitments it cannot easily escape, gets 25% of its revenue from just two unnamed customers, and faces open-weight models that are catching up in performance within six months of a leading model's release.

Moving the IPO from October to November does not change any of these numbers. It gives the company one more quarter of revenue data.

What actually survives this correction

Here is something no AI IPO presentation will say openly: The biggest winners in business AI will not necessarily be the companies that own the most expensive model. They will be the companies that build systems to turn any model into results that businesses can control, check and measure. These systems will use private business data, models adjusted for specific fields, groups of AI agents working together, and safeguards that make it possible to trust AI systems to carry out tasks on their own.

This value does not depend on keeping AI models rare or achieving a $2 trillion valuation. It grows as these systems are used in real business operations, where customers understand that the model itself is not the product. How the model is put to work is what matters.

The dot-com crash did not happen because the internet was bad. It happened because investors paid too much for companies. AI investors face the same problem: Real technology does not automatically justify a company's valuation. One strong quarter and a one-month delay do not fix that gap. November simply gives Anthropic more time to explain how the numbers add up.

Anurag Gurtu, co-founder and CEO of Airrived, is an entrepreneur and product leader with more than 20 years of experience in cybersecurity and AI. He has built three generative AI companies, including one that reached a valuation of $1 billion or more. Gurtu is known for turning advanced AI into platforms that can grow and work in real-world settings. He has also helped achieve successful company sales, including acquisitions by Splunk, Tripwire and Sun Microsystems, as well as the public company FireEye. He has a master's degree from the University of Southern California and holds the CISSP certification.

Illustration: Dom Guzman