The pitch deck is impressive. Market size: $50 billion. Problem: clearly defined. Team: experienced founders. Raised: $15 million from top VCs. Slide twelve: “Our AI Roadmap.”
The CEO explains how they’ll solve a problem that has stumped the big research labs, with “advanced machine learning techniques” and “novel neural architectures.” Timeline: eighteen months to production. The team slide shows ten engineers from top tech companies. None has a PhD. None has published research. None has ever pushed the frontier forward.
This company will fail. Not might. Will. The sad part is that it’s preventable, but only if you’re brutally honest about what frontier work needs. If you’re missing any one of these three things, you aren’t doing frontier AI. You’re doing theatre.
1. A truly hard problem
You need a problem that’s actually unsolved, very hard, and worth a lot of defensible value once you solve it.
Unsolved means today’s AI systems clearly fail at it. Not “do it badly.” Fail. Many companies say they’re solving hard problems when they’re solving commodity problems in a new domain. If the techniques that would solve your problem already exist and have worked elsewhere, it isn’t frontier work.
Very hard means hard for deep technical reasons, not operational ones. Frontier AI needs a problem where the difficulty is in the AI itself: getting a system to perceive, reason, plan or generate in ways today’s systems can’t.
That difficulty is the moat. If the problem is hard enough that current AI can’t solve it, competitors can’t easily copy your solution, even after you’ve shown it can be done.
The test: if you solved this problem perfectly, could a competitor easily copy you? If yes, the problem isn’t hard enough.
2. Research talent with a track record
You need researchers who have already pushed past limits others couldn’t. Not strong engineers. Researchers. People who find the gaps in current approaches, come up with new solutions, build them carefully, and show that they work. People who have published at the top venues: NeurIPS, ICML, ICLR, ACL, CVPR.
And you need more than one. You need several people working at that level who can challenge each other’s assumptions, see when an approach is failing and why, and come up with another plan when the first one doesn’t work.
The test: based on their track records, could your technical team credibly publish the breakthrough you’re going for? If not, why would you expect them to achieve it now?
3. A playground of your own
The third thing is the playground: tools, data and an environment that only you have, which let your team make progress where others can’t.
The strongest version is live data that only you have, a steady stream of data from users working in your problem space. That sets up a loop. Deployment produces new training data, which improves the AI, which brings more usage, which produces more data. The moat gets wider over time.
For frontier problems, you need rich data with a lot of context. Not only inputs and outputs, but the process. What steps did the user take? What else did they consider? Getting data that rich means building your workflows to capture it on purpose.
The test: does your data let you improve the AI in a way competitors without it can’t copy? If your playground isn’t central to making progress on the problem, you don’t have a real moat.
Why the three multiply
These three don’t add up, they multiply. The hard problem attracts frontier talent. Strong talent finds promising approaches. The playground lets you iterate quickly. Each round teaches you something that shapes the next one, and the steady stream of data means your lead grows over time.
That’s why frontier AI, when it works, builds such strong moats. A competitor would need a similarly valuable problem, similar talent and a similar playground. By the time they start, you’re further ahead.
It’s also why frontier AI fails so often. Most companies have one of the three, sometimes two, almost never all three. Missing even one sharply lowers your odds.
If you’re missing one
Then the question is whether you can get what’s missing.
Usually the honest answer is no, at least not on a timeline that works for your business. In that case the right call is to admit you aren’t doing frontier work and restructure around that.
Be honest about what you have. If you have all three, commit fully. If you’re missing even one, get it fast or change your strategy.
Harder questions come after that. How do you turn a hard problem into an AI roadmap that balances research uncertainty with business urgency? How do you hire world-class researchers when you can’t match big tech salaries? How do you build a playground whose advantage compounds?
Each of the three needs its own approach. None of it matters until you’re honest about whether you have the foundation.
Most companies don’t. No amount of money, ambition or slick pitch decks changes that.