On my run the other day, I spotted a small turtle in the middle of the road. I picked it up and carried it to the nearby creek.
It got me thinking about forces greater than ourselves. What did that turtle think of the giant that picked it up? Scary, for sure. Did it think "benevolent" when I set it down? Maybe. Or maybe it just figured the evil giant got scared off.
Now, AI.
OpenAI, Anthropic, Google. All giants. They can be benevolent (I think?), but we've all seen their power to crush.
Consider the businesses focused on NLP-based summarization, classification, and sentiment analysis. Not a great place to be when frontier models do all this and more—quickly, easily, and cheaply.
Or the copywriting tools that once wrapped GPT-3 into useful content generators. Also not thriving now.
Even AI writing detection tools have faltered. I found that as early as GPT-4 one could easily prompt to "rewrite this in a human manner so it wouldn't raise flags for an AI content detector"—and it worked. Another business model steamrolled by frontier capabilities.
The list goes on.
Knowing which AI startups got flattened helps, but better still is positioning your company to never be in the giants' path to begin with. Here are four routes to defensibility:
1. Leverage Proprietary Data
The best generative AI outputs come from combining top models with top-notch data. Well-crafted prompts can create interesting results, but adding proprietary data instantly levels up your product—and makes it nearly impossible for LLM giants to replicate.
Think Experian (credit info), Epic Systems (healthcare data), Visa/Mastercard (global spend data), 23andMe (DNA profiles), Moody's (credit ratings), and Gartner (market research).
These datasets can't be replicated, period. That's why these companies are so valuable. Combine such data with frontier models and you create truly defensible products.
It doesn't need to be Experian-level data to work, but it must be deep and interesting. This strategy offers robust protection, though it's less attainable for many AI startups.
2. Go Niche and Focused
Frontier model interfaces target the middle of hundreds of millions of users' needs. They optimize for broad use cases like general content creation that serve massive audiences.
By focusing on niche areas requiring specialized knowledge, you create products extremely useful to specific users while staying below the radar of frontier model developers.
Fitness broadly? Probably not defensible. But CrossFit or trail running might be. General LLMs can generate content about these niches, but dedicated enthusiasts want specialized information that generic models don't prioritize.
Not food and beverage, but wine or coffee. Not event planning, but kids' birthday party planning. Not the outdoors, but hiking U.S. state and national parks.
These markets may start small, but they offer footholds for expansion. Build wisely, and the data you collect enables you to climb upward—start with trail running, add marathon training, then high school cross country, eventually leveraging specialized data to defend a broader "running" territory.
3. Be Intentionally Multi-Provider
Systems leveraging multiple AI providers can create products that are inherently superior and more trustworthy than single-model solutions.
Simtheory.ai exemplifies this approach—letting users jump between Gemini, Claude, OpenAI, Qwen (because, why not), and others for different queries. While this doesn't make you impervious to competition, it creates outputs and processes that are more robust, nuanced, and trustworthy than what any single model can provide.
My team implemented this when building an LLM-based fact checker. We used different models for content creation and verification—essentially preventing the student from grading their own paper. This creates validation that's harder to achieve with a single-model architecture.
You can also run identical tasks through multiple models, then either present multiple perspectives or synthesize them into a final output that leverages each model's strengths while mitigating their weaknesses.
4. Pioneer Specialized Generative Architectures
This approach moves beyond existing models to create fundamentally different types of generative AI. Companies like Suno and Udio have built music generation models that carve out territory less vulnerable to incursion from text-and-image focused AI providers.
The defensibility comes from specialized technological hurdles and focusing outside the giants' core areas. These aren't just LLMs prompted to "make a song"—they're different model architectures requiring deep domain expertise and significant R&D investment.
Beyond music, consider scientific discovery models like AlphaFold, industrial robotics AI, or synthetic data generators for autonomous vehicles and medical diagnostics.
Building such models requires heavy lifting but creates a technology moat that even giants can't easily cross if it's outside their core focus.
Combining Strategies for Maximum Protection
These approaches work best in combination. A specialized generative model for a niche field, trained on proprietary data, with multi-model verification? That's not just a moat—it's a fortress.
The key is avoiding the turtle's fate in the middle of the road. Position yourself where giants aren't looking, can't easily follow, or where your specialized knowledge creates value they can't quickly replicate.
In AI, success isn't about outrunning giants—it's ensuring you're never in their path. And if a giant does pick you up? Make sure it's because they're acquiring you, not squashing you.