Can AI startups outrun dot-com bubble comparisons? Investors aren’t so sure.

ChatGPT fever hit something of a peak last year, investors said this week at Collision, a tech startup and investor conference in Toronto. The aftermath, and a sense of déjà vu, are giving them a clearer sense of what’s investible.

“Everybody thought dot-com was useful. ‘Oh I can go buy things online,’ and look what happened,” Wesley Chan, a co-founder and managing partner at FPV Ventures, said in Toronto. “I’m the anti-AI guy,” he added, describing his investment strategy.

Of the 1,623 startups that exhibited at Collision this year—the highest number of any Collision event—20% are building AI products, according to organizers. A spokesperson said that doesn’t include a large proportion of startups that now have “AI components” in their business.

Only a small number of them will survive and break through the AI hype, investors said this week.

Some said they’re increasingly looking for startups with business models that have long-term viability, products that solve corporate business problems and those with access to stores of private or unique data to train AI models.

The dot-com boom of the late 1990s got “messy” because every venture-capital firm needed its bet in the space—leading to inflation in the costs of expenses such as hiring and office space, according to Mike Schroepfer, the former chief technology officer of Facebook parent Meta Platforms and a partner at Gigascale Capital.

A similar dynamic is playing out now with the AI boom, investors say.

Nowadays, scrutiny should be applied to AI startups, where a rush to fund them has created a lot of noise, Schroepfer said in Toronto. Plus, training large language models at the scale of OpenAI requires millions of dollars for computing and AI chips, so that’s not an area in which new startups can be competitive.

“What you’re left with is, ‘Is there a market where I can get a very unique data moat, where I have the data to train the model?’” he said. “And when my customers use it, they give me new data, so I’m building my data flywheel.”

Since the start of the AI boom, investors have poured record levels of funding into AI companies such as Mistral AI, which raised $650 million earlier this month at a roughly $6 billion valuation. Tech company Amazon invested $2.75 billion in Anthropic in March, bringing its total investment in the AI company to $4 billion. CoreWeave, an AI computing startup, raised $7.5 billion in private-debt financing in May.

Those are just a few of the biggest checks—investors put $21.8 billion into generative AI deals last year, up fivefold from the prior year, according to research firm CB Insights. The average round size for those deals was $51 million, compared with the industry average of $8 million, CB Insights found.

Judging by the type of startups on display at Collision, some investors say the market has reached an inflection point. With a handful of companies such as OpenAI and Anthropic dominating the building of large language models, startups like Databricks and Scale AI providing data capabilities, and others in image generation and customer service, everyone else is fighting to stand out.

Wall Street Journal owner News Corp has a content-licensing partnership with OpenAI.

Alex Mans, founder and chief executive of AI-powered travel and transportation platform Flyr, said he encountered a large number of AI startups at Collision working on products that do the same thing as existing AI models.

“There’s a lot of companies that look like vertical software-as-a-service, but they’re just a pretty interface on top of a large language model,” he said. Some offer products that help companies analyze their invoices, but there’s no reason to use them when an AI chatbot works just as well, Mans said.

That’s the same reason Matt Wood, vice president of AI products at Amazon’s cloud unit, said some AI startups have been outmaneuvered by the technology’s pace. Taking an input and giving it to an AI model like OpenAI’s GPT-4 doesn’t offer much differentiation, he said in Toronto.

One solution, Wood said, is for startups to build “autonomous agents,” or virtual AI workers that can perform specific tasks on behalf of humans. The technology is developing quickly, he said, and names of AI agents may soon become as valuable as web domain names.

The cloud giant is putting $230 million into generative AI startups this year, Wood added. “What we’re seeing is the doubling down on what’s working,” he said. “Now, there’s a few breakout areas where we have high competence, and they’re attracting more of the investment as an overall percentage.”

Joseph Dormani, a partner at Thomson Reuters Ventures, said his team met with about 50 startups at Collision this week. He said some of the most interesting weren’t those with general AI capabilities, but those offering the ability to search across databases using AI, and that helped companies use multiple AI models.

It’s likely that the Thomson Reuters venture arm will invest in some of them, he said. Last year, it invested in a startup building AI summaries for medical insurance claims.

“When people say, ‘We’re investing in AI,’ at this point it’s almost like saying you invest in software. It doesn’t really mean too much,” Dormani said. “The product is what matters.”

Write to Belle Lin at belle.lin@wsj.com