---
title: "How the AI Bubble Crash Could Happen"
description: "AI model labs can become very large companies and still disappoint public markets if valuation runs ahead of adoption."
date: 2026-05-28
tags: [ai, strategy, economics]
url: https://nem035.com/thoughts/how-the-ai-bubble-crash-could-happen
---


[Anthropic at a $965 billion private valuation](https://www.anthropic.com/news/series-h) and [Big Tech planning up to $725 billion of capex in 2026](https://www.bloomberg.com/news/articles/2026-04-30/us-big-tech-ratchets-up-ai-spending-past-700-billion-this-year) are the AI bubble problem in two numbers. The ROI is still hard to see in a lot of places. Companies are spending because they do not want to fall behind, but it is not always clear what they are getting for the money.

Nobody knows if there will be a crash or when it would happen. I am not a public markets expert, and this is mostly personal intuition from watching the setup. My default read is that a crash is more likely than no crash if the big AI labs go public while valuations are this high. The setup has very large private valuations, very large compute bills, unclear enterprise ROI, circular deals, and a public market waiting for a way to buy the companies directly.

The argument does not require AI to stop working. The more interesting version is that AI keeps improving, adoption keeps growing, and model labs still capture less of the value than public investors expect. A crash can come from the economics settling differently than the story, not only from the technology disappointing.

Private markets are strange, but they are still limited. A few investors mark a company at a high price, the company gets the money, and everyone else watches from the outside.

Public markets add a different kind of pressure because once the stock trades, anyone can buy it. Fund managers, retail traders, index buyers, options desks, and momentum investors can all buy at the same time. With AI, the pressure is stronger because many people already feel behind, whether they are companies late to adoption or investors who missed NVIDIA. An OpenAI or Anthropic IPO would give them a simple way to feel caught up.

Once the public gets involved, greed can become a major input into the price. If enough people want to get rich from the same trade, blind wishful thinking can turn a real technology into a bad entry price.

During the dot-com bubble, IPO pricing became detached from normal business reality. [Loughran and Ritter's dot-com IPO paper](https://archive.nyu.edu/jspui/bitstream/2451/27205/2/S-FI-01-12.pdf) says first-day returns averaged 88% in 1999 and 2000. That means the average IPO buyer at the offering price could almost double their money on day one. [PBS's summary of Jay Ritter's IPO data](https://www.pbs.org/wgbh/pages/frontline/shows/dotcon/thinking/stats.html) also documents large first-day IPO pops during the internet bubble.

That kind of market teaches people the wrong lesson. It rewards getting allocation before everyone else more than owning a business at a sensible price. Once that mindset spreads, access can become more important than valuation.

The SPAC boom was a newer version of the same behavior. [S&P Global's 2021 market-debut report](https://www.spglobal.com/market-intelligence/en/news-insights/research/equities-update-2021-deal-volumes-a-record-year-q4-outlook) says 2021 had 776 market debuts and SPACs made up 60% of IPOs. [Boardroom Alpha's SPAC database](https://www.boardroomalpha.com/spac-statistics/) says 2021 had 614 SPAC IPOs raising $144.9 billion. Many of those deals were sold on future expectations before the business quality was clear.

Retail participation also changed the speed of speculation. CNBC reported that [2021 options activity hit records and retail investors made up more than 25% of total options trading](https://www.cnbc.com/2021/12/22/options-trading-activity-hits-record-powered-by-retail-investors.html). Options can make speculation move faster because they give people leverage.

Public markets turn a private company into a daily price, and that price gets pushed around by IPO access, index demand, options, social proof, analyst targets, and career risk for fund managers who do not own it.

AI has the ingredients for that behavior: a simple pitch, hard cash-flow modeling, a product many people have touched, large potential upside, and the fear of missing the next NVIDIA.

## The Numbers

[Anthropic announced on May 28, 2026 that it raised $65 billion at a $965 billion valuation, with run-rate revenue above $47 billion](https://www.anthropic.com/news/series-h). [OpenAI announced in March that it raised $122 billion at an $852 billion valuation, while making $2 billion of revenue per month](https://openai.com/index/accelerating-the-next-phase-ai/).

Those are large businesses already. Anthropic's latest round implies about 21 times run-rate revenue, while OpenAI's March round implies about 36 times its disclosed monthly revenue annualized.

The bubble risk sits around companies with real products and real revenue. ChatGPT has broad reach, with OpenAI saying it has [more than 900 million weekly active users, more than 50 million subscribers, and enterprise at more than 40% of revenue](https://www.techradar.com/pro/we-are-growing-revenue-four-times-faster-than-the-companies-who-defined-the-internet-and-mobile-eras-openai-says-its-making-usd2-billion-a-month-mostly-from-enterprise-users).

The bearish case is that the price may get too far ahead of the business model, even if AI keeps working.

## The NVIDIA Anchor

The public-market anchor is NVIDIA.

NVIDIA is the AI winner investors can already buy. As of late May 2026, its [market cap is around $5.3 trillion](https://companiesmarketcap.com/nvidia/marketcap/), and in its latest reported quarter NVIDIA reported [$81.6 billion of revenue, up 85% from a year earlier, with GAAP gross margin of 74.9%](https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Announces-Financial-Results-for-First-Quarter-Fiscal-2027/default.aspx).

Investors will have that benchmark in their heads. If they believe OpenAI and Anthropic can become the next platform layer for the economy, even briefly, the market can justify very high prices.

But NVIDIA also shows the problem. AI is priced like software, but it is built through chips, fabs, power, data centers, cooling, networking, land, debt, leases, and long-term capacity contracts.

Normal software does not usually require this much physical capacity before the next dollar of revenue can show up.

## The Cost Problem

AI model labs need constant access to expensive compute, and the companies that supply that compute are also spending a lot of money.

Bloomberg reported that [Microsoft, Alphabet, Amazon, and Meta could spend up to $725 billion on capex in 2026](https://www.bloomberg.com/news/articles/2026-04-30/us-big-tech-ratchets-up-ai-spending-past-700-billion-this-year), mostly for AI data center equipment. Axios framed the same earnings cycle as investors asking companies to [show the return on AI spending](https://www.axios.com/2026/04/30/ai-meta-alphabet-microsoft-amazon), not just talk about AI.

Revenue and spending are both going up fast. Large spending is not automatically a problem, but spending at this scale needs large justification. The valuation can work if revenue compounds faster than compute cost, capacity stays well used, and margins improve as the systems get more efficient. If compute cost stays too visible, the market starts valuing the labs less like software and more like infrastructure.

The physical buildout is already big enough to show up outside tech earnings calls. Coverage of US Census data says [data center construction spending surpassed office construction spending in late 2025](https://www.bdcnetwork.com/building-sector-reports/data-centers/news/55365223/data-center-construction-spending-outpaces-office-market-for-the-first-time). That does not include the racks and servers inside the buildings.

Calling AI a software business leaves out the factories, power contracts, data centers, financing, and hardware supply chain required to serve the next wave of usage.

## The Circular Deal Problem

The circular deals make this harder to understand. NVIDIA [intends to invest up to $100 billion in OpenAI](https://nvidianews.nvidia.com/news/openai-and-nvidia-announce-strategic-partnership-to-deploy-10-gigawatts-of-nvidia-systems) as OpenAI deploys NVIDIA systems. Cloud providers invest in model labs that buy cloud capacity, including Anthropic's expanded compute deals with [Amazon](https://www.anthropic.com/news/anthropic-amazon-compute) and [Google Cloud](https://www.anthropic.com/news/expanding-our-use-of-google-cloud-tpus-and-services). Model labs sign infrastructure deals that help suppliers report demand, and that supplier demand helps justify the model lab valuation.

Some of this is normal because strategic capital can be the right move when supply is tight. If compute is scarce, you want partners, guaranteed capacity, and alignment between the chip company, cloud company, and model company.

But public markets eventually ask a simple question: which dollars are normal customer demand, and which dollars were pulled forward by the buildout itself?

Microsoft's OpenAI partnership is one example of how tangled this can get. Microsoft says [OpenAI remains its primary cloud partner and revenue-share payments continue through 2030](https://blogs.microsoft.com/blog/2026/04/27/the-next-phase-of-the-microsoft-openai-partnership/).

The revenue can be real while the revenue quality is hard to understand, and revenue quality becomes important when a company is valued at 20, 30, or 60 times revenue.

## The Customer Budget Problem

There is also a customer version of the same issue. Teams adopt AI tools, usage grows faster than expected, and then someone realizes the budget was built for experiments while the usage looks like production.

Uber is a useful example. The Information reported that [Uber's Claude Code use maxed out its full-year AI budget only a few months into 2026](https://www.theinformation.com/articles/uber-cto-shows-claude-code-can-blow-ai-budgets), and Forbes later summarized that Uber had [burned its 2026 AI budget in four months](https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-burns-its-2026-ai-budget-in-four-months-on-claude-code/). Tom's Hardware also covered Uber leadership saying [the company could not clearly connect higher AI usage to better shipped products](https://www.tomshardware.com/tech-industry/artificial-intelligence/uber-chief-warns-no-link-yet-between-ai-tokenmaxxing-and-shipping-successful-products-company-pumps-the-brakes-on-all-out-ai-spending).

Budget pressure can cut both ways. A blown budget can mean the tool is wasteful, but it can also mean the product found real demand faster than finance expected. Cloud, mobile data, AWS, and online advertising all had versions of this pattern: usage ran ahead of budgets, finance added controls, and spending kept growing anyway.

The risk is not that budget reviews stop AI adoption. The risk is that they change the shape of adoption. Customers may keep spending, but with procurement, usage caps, internal chargebacks, and harder ROI reviews.

More disciplined buying is good for mature customers, but it can be bad for a stock priced as if usage will grow without friction.

## The Pricing Problem

Tokens are not mobile data, but a similar pattern can show up when new usage overwhelms fixed infrastructure, exposes marginal cost, breaks simple pricing, and forces the market to learn the real unit economics.

Today, demand is ahead of supply, so model labs can price against customer ROI. If a coding agent saves expensive engineering time, the vendor can charge a lot, but that does not mean pricing power lasts forever.

[Epoch AI tracks LLM inference prices falling rapidly](https://epoch.ai/data-insights/llm-inference-price-trends), and [Stanford's 2026 AI Index](https://hai.stanford.edu/ai-index/2026-ai-index-report) says open models have narrowed the gap with the top closed models. As capacity grows, inference gets cheaper, customers learn which tasks are worth paying for, competitors cut prices, and apps hide the model layer underneath a workflow. Buyers start asking why one model should cost so much more than another if both are good enough for the job.

My base assumption is that LLMs commoditize. Model labs can survive commoditization, but the base model layer becomes a harder place to keep extreme margins. The durable value may move up the stack: agents, workflow software, vertical tools, enterprise systems, proprietary data, distribution, and implementation.

If that happens, model labs need to own more than the model: the customer relationship, the workflow, distribution, and a reason customers keep paying them instead of swapping in the next good-enough model.

The strongest counterargument is that the leading labs may become that stack themselves. If a lab owns the user identity, memory, agent runtime, enterprise admin layer, integrations, app marketplace, and default interface, then the base model can commoditize without the business commoditizing. The model becomes one part of the product. The durable business is the operating layer around it.

The crash case depends on that layer forming slower than the valuation requires, or forming somewhere else. If the workflow layer ends up owned by Microsoft, Salesforce, ServiceNow, Apple, Google, vertical SaaS companies, or internal enterprise systems, then the model lab captures less of the value than public investors may expect.

## Usage Is Not The Same As Depth

Usage alone can mislead. ChatGPT has broad reach, and that reach is valuable. But broad awareness can coexist with shallow habits, occasional use, and weak willingness to pay. OpenAI has said [about 95% of ChatGPT users do not pay](https://apnews.com/article/3c2674f5cdf67ac6d88eedb207de117c). That may change, but it is a real gap between usage and monetization.

The same gap exists at work. Lots of people have tried AI, fewer people use it every day, and even fewer have rebuilt a workflow around it. The jump from "I used this last week" to "my department runs on this" takes product, integration, training, procurement, compliance, change management, support, security review, data access, and process redesign.

AI software may be a bigger opportunity than raw models for this reason. Turning a model into useful software takes use-case selection, interface design, workflow integration, proprietary data, support, sales, and implementation work. Automation takes a lot of manual labor before it feels automatic.

If startups and incumbents build the workflow layer, the model lab may become the supplier underneath.

## The Task And The Job

The task and the job are not always the same. AI can do visible tasks like writing code, drafting copy, summarizing calls, generating spreadsheets, and classifying tickets. The job often includes the parts around the task: deciding what should exist, knowing which details matter, applying taste, handling exceptions, taking responsibility, and changing the surrounding process.

ROI gets fuzzy when companies buy automation for the task but still need humans to do the job.

This connects to [the illusion of software automation](/thoughts/the-illusion-of-software-automation). AI can produce work quickly, but companies still need review, integration, accountability, and process changes before they can rely on it for important systems.

Cheap tasks can unlock new work. When a task becomes cheap enough, people stop asking how to do the old task faster and start asking what else they can now afford to do. A company might review every customer interaction, classify failure modes, generate follow-ups, test new scripts, and give product teams a map of where customers get confused.

Cheaper software creation may create more software, more internal tools, more agents, more audits, more integrations, and more operational surface area. I still think [SaaS survives AI](/thoughts/saas-in-the-age-of-ai) because building software and owning the system around it are different jobs.

The problem for an IPO buyer is timing. New categories of work can take years to become clean model-lab cash flow.

## No Crash

AI revenue keeps compounding quickly, enterprise adoption moves beyond pilots, companies rebuild real workflows around AI, and usage grows fast enough to absorb price cuts, competition, and procurement friction. At the same time, margins improve as inference gets cheaper, hardware supply opens up, model efficiency improves, and the labs keep enough pricing power to turn high usage into attractive economics.

In that world, a high valuation can be rational. [Anthropic is already claiming $47 billion of run-rate revenue](https://www.anthropic.com/news/series-h), and [OpenAI disclosed a $24 billion annualized pace in March](https://openai.com/index/accelerating-the-next-phase-ai/). If those numbers keep compounding and investors believe the margin structure will eventually look like elite software, trillion-dollar valuations can make sense.

This outcome is possible. We have already seen how quickly AI products can reach real usage, and the [AI honeymoon phase](/thoughts/ai-honeymoon-phase) exists because first contact with these tools can be surprising. The no-crash case requires the surprising part to become normal corporate habit, where the product moves from "everyone tried it" to "budgets depend on it."

## Crash

The crash version can start with real adoption. The labs go public in late 2026 or 2027, the stocks open strong, and analysts build models around very large outcomes. Investor presentations talk about knowledge work, agents, and the percentage of GDP that AI can touch. Then the market starts asking for proof on a public-company schedule.

Revenue may still grow quickly, but not quickly enough for the price. Enterprise adoption may be real but slower than the story implied, margins may improve while compute costs and capex stay visible, competition may force price cuts, customers may push back on bills, and open-source models may make the base layer feel less special.

Investors may also start haircutting revenue quality. If growth is tied to infrastructure partners, GPU commitments, cloud credits, reseller agreements, or customers whose own AI budgets depend on the same capital cycle, the reported top line can still be real while the market assigns it a lower multiple.

Investors often skip this during a boom. A technology can be transformative over ten years and still disappoint a stock that needs the next eight quarters to look perfect.

## The Valuation Math

A $3 trillion valuation at [Anthropic's disclosed $47 billion revenue run rate](https://www.anthropic.com/news/series-h) is about 64 times sales. The exact future multiple is unknowable, so 12 times sales is not a prediction. It is a pressure test. If the market eventually values the company like a very strong but more mature software business, the company needs $250 billion of revenue to justify the same market cap. If the market keeps valuing it like a scarce platform layer, the required revenue is lower. If the market starts valuing it like infrastructure, the required revenue is higher.


That is why the multiple matters so much. The crash case does not require the business to fail. It requires the market to move from "this company owns the next platform" toward "this is a large, expensive, competitive business with real margins to prove."

OpenAI and Anthropic can probably become very large companies. The harder question is whether they can grow into a public-market price before the market loses patience. At a $5 trillion valuation, a 12 times revenue pressure test would imply more than $400 billion of annual revenue.

If IPOs happen in late 2026 or 2027, my default crash window would be 2028 to 2029. That gives the stocks enough time to run up, enough time for the narrative to spread, and enough public quarters for investors to compare the story with reported numbers.

A faster crash could happen 6 to 18 months after IPO if the stocks run hard first. In that version, the first few reports show margin pressure, capex pressure, slower enterprise conversion, growth deceleration, or lower-quality revenue than investors expected.

No crash probably requires extreme revenue growth through at least 2029, plus clear progress on compute economics.

## The Part I Keep Coming Back To

AI can change a lot and still produce a bad stock outcome. People mix those up during booms. A lot depends on where value settles: some technologies replace an old market, some expand it, some create a new layer, and some make the old layer cheaper.

If frontier model progress slows, that would be a cleaner crash catalyst. The article does not need that assumption. The more subtle risk is that AI succeeds while the public market overpays for the base model layer.

I think LLMs commoditize, and I think the model labs become very large. I also think the public market may price them as if they capture too much of the value, too quickly, with too little margin pressure. The crash can happen even if AI keeps working because everyone sees the upside, everyone wants in, and the price gets set before the business model finishes settling.