Beyond the Hype: Which AI Startups Are Actually Building Something That Lasts

The graveyard of AI startups is filling faster than the valley of winners. For every company that has found genuine product-market fit, there are dozens burning through compute budgets on demos that don't convert.

What separates the durable from the disposable? After speaking with founders, investors, and researchers across the ecosystem, we found the answer is less about the model and more about the moat.

The Model Is Not the Moat

This sounds obvious when stated plainly, but the behavior of many founders and investors suggests it is not internalized. The companies that will matter in five years are not the ones with access to the best model — because model access is commoditizing at a rate that makes differentiation at that layer nearly impossible for most startups. They are the companies with proprietary data, entrenched workflows, or distribution advantages that make their products genuinely difficult to replicate, even by a competitor with a slightly better underlying model.

We spoke with fourteen founders who are building companies in this space. The ones with the clearest long-term outlook shared a common framework: they could articulate precisely what would make their company defensible in a world where the frontier model available today is commoditized and generally available in 18 months. Most of them were not counting on the model at all. They were counting on the data the model ran on, the integrations that made switching painful, and the domain expertise embedded in how the product was designed.

The ones with the weakest long-term outlook were doing the opposite: betting that their choice of model, their fine-tuning approach, or their prompt engineering was itself the differentiator. In a market where the underlying models are improving at the current pace, that bet has a short shelf life.

What the Durable Companies Have in Common

Across our conversations, four characteristics appeared consistently in the companies that investors — serious, pattern-matching, been-through-a-cycle investors — were most confident about.

The first is domain depth. The best AI startups are built by people who knew the domain before the AI existed. They are not generalist engineers who identified an opportunity and learned enough to pitch it. They are people who spent years in legal, in healthcare, in supply chain, in financial regulation — and who are now applying AI to problems they understand deeply enough to build genuine solutions rather than impressive demonstrations. The demos from domain experts look different from the demos from generalists. The edge cases are handled. The vocabulary is right. The integrations make sense. The buyers notice.

The second is proprietary data. The most defensible AI companies are the ones sitting on datasets that cannot be replicated by a competitor writing a larger check. This can come from exclusivity agreements, from being the system of record in a workflow, from a community that generates data as a byproduct of using the product, or from partnerships that provide access to data that is genuinely difficult to obtain. The companies that understand this are building data acquisition into their go-to-market from day one — not as a future initiative, but as the core logic of why the product works and why the product improves over time.

The third is workflow integration. The AI companies with the highest retention are the ones that have made themselves structurally difficult to remove. Not through lock-in in the malicious sense, but through integration so deep that the product has become part of how users do their jobs. When an AI tool is embedded in the daily workflow — when outputs from it feed into other systems, when colleagues expect data from it, when removing it would require rebuilding processes around it — the switching cost is real regardless of whether a better product enters the market. The companies that understand this are not just building great AI features. They are building distribution infrastructure.

The fourth is a clear theory of value capture. This sounds like a truism, but the number of AI companies with compelling demonstrations and no clear path to monetization that scales is striking. The best companies can describe, precisely, who pays, how much, and why the willingness to pay is durable as the market matures. They are not relying on a first-mover premium that evaporates as competition increases. They are building toward a margin structure that makes sense even in a world where the AI they are using is no longer novel.

The Six Investors Worth Listening To

The investors we spoke with who have the clearest framework for evaluating AI startups are not the ones who moved fastest in 2021 and 2022. They are the ones who watched the previous generation of AI hype — the machine learning wave of 2016 to 2019, the NLP excitement that preceded the transformer era — and formed specific views about what translated into durable businesses and what did not.

The consistent thread in their framework is skepticism about any company whose core pitch depends on the continued superiority of their model choice. "I ask every AI founder the same question," one partner at a growth equity firm told us. "What does your company look like in three years when OpenAI releases this capability for free? If the answer is 'we're in trouble,' I know what I need to know." The companies that pass this test are the ones where the model is an input into a larger system — not the system itself.

What they are looking for instead is evidence of compounding: that the product gets better as it processes more data, that the customer relationships generate insights that feed back into product improvement, that the network effects (where they exist) are genuine rather than rhetorical. Compounding is the only reliable source of durable advantage in a market where the underlying technology is improving as fast as AI is. The companies that are building it are the ones worth watching.

The Honest Assessment

Most AI startups funded in the past two years will not exist in five years. This is not a prediction about AI — it is a prediction about startups, which has always been true. The difference is that the failure modes are faster and more visible in this cycle than in most previous ones. The companies that will succeed will look, in retrospect, like they always had a plan. They probably did. The ones that will fail will look like they were surfing a wave that stopped. They probably were.

The actionable insight, if there is one, is this: the question worth asking about any AI company — as an investor, as a potential employee, as a competitor — is not whether the technology is impressive. It usually is. The question is whether there is a business underneath the technology that would survive a world where the technology became free. The companies with a compelling answer to that question are the ones building something that lasts.