A startup built on someone else's large language model stays defensible by owning what a competitor cannot copy with more money or a better engineering team: proprietary data, protected intellectual property and in-house models layered on top of the LLM. That combination, together with a team that has domain expertise and a product that has proved its market fit, is what venture capitalists look for when the core technology is not yours.
Key takeaways
- Relying on a third-party LLM leaves a startup open to copying by rivals and to larger companies moving into its market.
- Innovation and IP protection differentiate a startup beyond the mere application of an LLM.
- Proprietary data is the proven defence: it deters competitors who have more funding or stronger technology teams.
- A practical path is to amplify proprietary data and technology with generative models, then consider an in-house LLM once product-market fit is proven.
- VCs want a team and a vision that go beyond a mere API call.
Why is defensibility the first question VCs ask?
As an entrepreneur and academic, I am often approached by venture capitalists (VCs) and startups with one pressing question: "How can a startup stay competitive and defensible when building its core on large language models (LLMs) such as ChatGPT?"
My answer is a careful one. Low-IP companies may raise concerns, but one factor demands our closest attention. This article explains why that element matters so much.
Why did LLM-backed startups attract so much VC interest?
LLMs have changed the way we interact with technology, enabling major advances in natural language processing and understanding. The technology has found applications across many sectors, from customer service and content generation to data analysis and decision-making. As a result, the startup landscape has seen a surge in ventures built on these powerful language models.
After OpenAI released ChatGPT, the interest it generated among VCs around the world was remarkable. Although some saw it as a lesser disruption, the enthusiasm surrounding LLMs caught the attention of renowned investors.
At the time of writing (2023), thousands of generative AI startups were attracting billions of investment dollars, and PitchBook expected the global generative AI market to reach $42.6 billion in 2023.
What makes nonproprietary technology risky for startups?
Transformative language models have sparked immense interest and potential among startups, but it is crucial to recognise the challenges of building on nonproprietary technology. Startups, unlike industry giants such as Google, often lack the resources and infrastructure to fine-tune, retrain or fully exploit the capabilities of LLMs.
This reliance on third-party technology leaves them open to replication by rival startups and to being overshadowed by larger corporations seeking to capitalise on growing markets. With well-established technologies and a constant hunger for new applications, these giants can deploy their own resources at scale and erode the competitive edge of smaller ventures.
A simple test makes the risk concrete. If a competitor could rebuild your product in a few weeks with the same model access and a well-written prompt, your advantage lies in speed and distribution, not in technology. Speed and distribution matter, but they rarely survive a larger company deciding to enter your market.
How do innovation and IP protect an LLM-backed startup?
To reduce the risk of technological replication, startups must actively find ways to differentiate themselves beyond the mere application of LLMs. The key lies in a culture of innovation and IP protection. A team that understands the nuances of the technology and also brings domain expertise and creative problem-solving can extend its offering and build a competitive advantage that goes beyond the underlying model.
Coupling this with the strategic acquisition and protection of IP rights can strengthen a startup's market position and shield it from imitators.
Why is proprietary data the key element?
In venture capital, owning unique data is a proven way to protect a technology's defensibility. VCs acknowledge that proprietary datasets can deter competition, preventing rivals from simply copying ideas with superior funding or technology teams.
The value of unique data goes beyond slowing down replication. It gives startups valuable insight, helping them identify hidden patterns, understand customer behaviour and stay ahead of market trends. By putting proprietary data to work, startups can strengthen their competitive edge and drive innovation in a fiercely competitive landscape.
One way for founders to stay ahead of their competition in the eyes of VCs is to amplify their proprietary data and technology with modern generative networks and, once product-market fit is proved, to look further into an in-house LLM.
How did SOMIN apply this?
For example, when building our competitor analysis suite, "SoMonitor", and its LLM-powered analytics assistant, "SoDa", technology defensibility was one of our top priorities.
At the core, we set a vast trove of anonymised data covering five years of advertising campaigns run on Meta's advertising platform for our clients. Using proprietary AI technology we built, this data is used to accurately predict the click-through rates (CTRs) and conversion rates (CRs) of advertising banners, along with a heatmap that highlights the visual components driving customer engagement and product purchases.
By integrating the visual and textual data we hold, combining our in-house multimodal models for ad creative performance prediction with the power of LLMs, we made the system generate actionable recommendations for clients. They cover competitor brand archetypes, audience outreach, content and promotional snapshots, delivered as a real-time digest. This lets our clients refine their strategies, improve creative assets and stay ahead in a fiercely competitive market.
Combining LLMs with in-house AI in this way let us build a new application in the martech industry that would not have been possible without LLMs. These capabilities also cannot easily be replicated. We use LLMs to describe the digital marketing industry comprehensively, and the combination of proprietary data and CTR prediction algorithms makes the system highly distinctive and hard to copy, setting it apart from competitors.
How should an LLM-backed startup pitch to VCs?
Whether to invest in LLMs remains a subjective decision, and investors weigh it differently. One piece of advice stays relevant, though: when pitching a startup, back up your claims with proprietary technology, valuable IP and a skilled team committed to success, especially at an early stage.
Going beyond the surface is crucial. If the team shows promising expertise and a vision that extends beyond a mere API call, VCs will certainly want to explore your company more deeply.
In practice, that means a pitch should answer four questions plainly: what data you own that others cannot get, what you have built on top of the model, what protects that work, and why your team is the one to do it. Our own approach to the first two is described in peer-reviewed papers on the SOMIN research page, and the same category data grounds SOMIN's AI marketing agents.
Related reading
- How can businesses use large language models beyond chatbots?
- Agency proprietary data: why category data beats a better prompt
- What is creative clickability, and can you predict it before spend?
- SOMIN research: peer-reviewed papers and talks
- SOMIN's AI marketing agents
FAQ
Can a startup built on a third-party LLM be defensible?
Yes, if it owns something the model provider and rivals cannot copy. Proprietary data, protected IP, in-house models and domain expertise all extend the advantage beyond the underlying LLM.
Why does proprietary data matter so much to VCs?
Unique data deters competitors who have more funding or stronger technology teams, because they cannot simply copy it. It also gives the startup insight into patterns, customer behaviour and market trends.
Should an LLM-backed startup build its own language model?
Not at the start. A practical path is to amplify your proprietary data and technology with existing generative models, then look into an in-house LLM once product-market fit is proven.
What do VCs want to see in an LLM startup's pitch?
Proprietary technology, valuable IP and a skilled, committed team, especially at an early stage. Above all, they look for expertise and a vision that go beyond a mere API call.
