AI can transform the economy and still be a bubble for investors. That was Bret Taylor’s argument in a September 2025 interview: the technology may create substantial value even as many companies and investors lose money. His comparison with the dot-com boom is a warning to separate the promise of a technology from the prospects of any particular investment.
What Bret Taylor said about an AI bubble
Taylor, OpenAI’s board chair and the CEO of AI-agent company Sierra, discussed the AI boom in a podcast conversation listed by Sierra on September 11, 2025. TechCrunch reported his remarks on September 14, 2025. Taylor’s central point was that AI could transform the economy while the market around it remained a bubble in which many people lose money. TechCrunch’s report covers the interview, and Sierra’s episode listing dates the discussion.
“Bubble” in this argument refers to financial and commercial excess, not proof that AI systems are useless. A technology can work and attract real customers while investors pay prices that assume growth, margins, or breakthroughs that never materialize. Taylor’s claim is a view about the market, not a dated forecast for when a crash will happen.
Why a transformative technology can still be a bad investment
Three questions often get collapsed into one: Does the technology do useful things? Can a company turn that capability into a durable business? Is the price of its shares or private valuation justified by the returns it can deliver? A positive answer to the first question does not guarantee positive answers to the others.
#1 Best Overall
- Technology: AI systems can perform useful tasks, but capability alone does not establish that a product is reliable or economically valuable in a particular setting.
- Adoption: A successful demonstration does not prove that an organization can integrate a tool into daily workflows, manage errors, and sustain usage.
- Revenue and margins: A vendor may attract customers yet struggle to earn durable profits after paying for model usage, computing, support, and implementation.
- Valuation: An investment can lose value if its price assumes unusually fast growth or high margins, even if the business continues to operate.
- Capital spending: Chips, data centers, and model development require major investment. Spending ahead of proven returns can leave investors exposed if demand or utilization falls short.
These risks vary across the AI market. Model developers, chip and data-center providers, cloud platforms, enterprise software firms, agent startups, consumer apps, and implementation businesses do not have identical costs or competitive advantages. A correction need not affect them equally, and falling valuations would not by itself mean that businesses stopped using AI.
What the dot-com comparison gets right—and what it cannot tell us
Taylor’s analogy is about the difference between a correct technology thesis and a correct bet on a particular company. In the late-1990s internet boom, belief that the internet would matter proved broadly right, but many individual businesses and valuations did not last. TechCrunch and Fortune’s coverage describe Taylor’s examples: Amazon and Google as lasting beneficiaries, contrasted with failed companies such as Pets.com and Webvan.
Rank #2
That history is easier to read with hindsight than it was to live through. It does not show that today’s eventual AI winners are already obvious, nor that every company associated with a transformative technology will share in its gains.
Where the analogy is useful
- Investors can overpay for companies with limited operating histories when expectations are unusually high.
- A new technology can attract investment before business models and customer demand are settled.
- Infrastructure built during a boom may remain useful, but its builders and funders are not guaranteed to earn an adequate return.
- The technology’s long-term importance does not ensure the survival of the firms that first popularize it.
Where AI differs from the early commercial web
- AI products can have substantial usage and revenue before they become profitable; usage is not the same as healthy unit economics.
- Economics depend on the costs of computing and inference, energy, and access to specialized hardware.
- Application companies may rely on a small number of model or cloud providers, creating exposure to supplier pricing and access decisions.
- AI features can be bundled into existing software, so competition may come from established vendors as well as standalone startups.
- Models and products are improving and changing quickly, which can help customers but also make an application’s advantage easier to copy or replace.
Why Taylor says a bubble can be “OK”
In Taylor’s historical framing, a boom can speed up research, infrastructure construction, hiring, and experiments with business models. Some of those investments may support useful products after weaker companies disappear. That is a possibility, not a guarantee that infrastructure will retain its value or that losses are harmless.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Taylor explicitly acknowledged that many people could lose money. A correction can hurt investors, workers whose compensation depends on startup valuations, companies that committed to costly systems, and customers relying on vendors that shut down or change their products. Calling a bubble potentially productive in hindsight does not make its losses painless.
What Sam Altman’s warning adds
Taylor’s remarks followed Sam Altman’s warning that someone would lose an enormous amount of money in AI while others would make enormous amounts, as reported by TechCrunch and Fortune. Both comments distinguish the possibility of broad technological change from the financial success of every company pursuing it.
Readers should also weigh Taylor’s perspective in context. He chairs OpenAI’s board and leads Sierra, an AI-agent company; both organizations have an interest in AI’s long-term importance. That does not disprove his argument, but it makes his remarks an informed industry view rather than neutral investment advice. OpenAI’s April 9, 2025 court filing identifies him as board chair.
Who could be exposed if expectations reset?
“Investors” are not one group, and bubble risk can reach people who never buy a technology stock directly.
Best Value
- Startup investors: Companies with little differentiation, weak customer demand, or high cash needs can struggle if new funding becomes harder to obtain.
- Public-market investors: A company’s AI plans do not, by themselves, establish that its stock price is justified. Valuation risk exists even for established firms.
- AI startups: A product may be vulnerable if it depends on one model provider or offers a feature that a larger provider can reproduce.
- Infrastructure builders: Expensive capacity can disappoint if utilization or customer demand falls short of what investors expected.
- Startup employees: A highly valued company can face a down-round, layoffs, or closure; equity compensation is not cash and its eventual value is uncertain.
- Business customers: A buyer can be left with migration costs if a vendor discontinues a product, changes access, or cannot sustain support.
These are risks, not claims that a particular share, company, or category is certain to fail. The available evidence does not establish a failure rate for AI startups or a date for a market correction.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How to evaluate an AI business or investment thesis
For personal-finance decisions, avoid treating enthusiasm about AI as a substitute for examining the business and the price. For a company buying software, test the operational case before making a broad commitment. The same underlying questions—whether customers stay, whether costs are controlled, and whether value can be measured—help in both settings, though they are not a formula for predicting returns.
For investors and people considering startup equity
- Check revenue quality: Distinguish recurring, recognized revenue from pilots, one-off projects, usage spikes, announced partnerships, and projections.
- Look for retention: Renewals and expansion are more informative than a list of trials or initial sign-ups.
- Understand unit economics: Ask how model usage, computing, implementation labor, and customer support affect gross margin as usage grows.
- Test defensibility: Identify a real advantage such as distribution, workflow integration, proprietary data, specialized expertise, brand, or compliance capability—not just an interface to a general model.
- Map supplier dependence: Determine whether one model provider, cloud platform, chip supplier, or sales channel can change the company’s costs or access.
- Stress-test funding needs: Consider whether the business can operate if growth slows and outside financing becomes less available.
A technically impressive demo may not survive the move to production; pilot announcements may not convert to lasting contracts; and a software-like price can hide substantial model and human-support costs. Those are reasons to ask for evidence, not proof that a company is doomed.
For businesses buying AI
- Name the task. Specify the workflow to improve, rather than buying a broad promise to “use AI.”
- Set a baseline. Record current cost, completion time, error rate, or conversion rate so that a pilot has a meaningful comparison.
- Define acceptable risk. Decide how reliable the system must be, what human review remains necessary, and what happens when it makes an error.
- Calculate the full cost. Include model usage, integration, support, and ongoing oversight—not only the vendor’s headline price.
- Review data and compliance terms. Check privacy, security, data retention, and regulatory obligations before sending sensitive information.
- Plan for supplier changes. Ask what happens if the model provider changes pricing, access, or behavior, and how difficult it would be to migrate.
- Measure outcomes before scaling. Expand only when the product produces repeatable, material results under real operating conditions.
Sierra describes its approach in terms of AI agents and business outcomes on its podcast listing. That is the company’s own positioning, not independent evidence of product performance.
What the bubble argument means for ordinary users
A downturn in AI investment could affect company valuations or the availability of particular products without making AI tools disappear. If you depend on an AI service for important work, consider how easily you could export your data, switch providers, or return to a non-AI process. Avoid confusing a vendor’s continued operation with the reliability of any specific feature, and check its terms before placing sensitive or hard-to-replace work in the service.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
