January 27, 2025. (Happy birthday, mom!)

When we set out to start our AI company, one of the earliest challenges we faced was a sobering reality: the AI landscape was littered with examples of companies rendered obsolete overnight by the emergence of new foundation models. These models, built on vast amounts of data and computational power, can reset the competitive game in an instant. For example, OpenAI’s GPT-3 upended countless startups working on natural language tasks, while advances in image generation models like Stable Diffusion reshaped the landscape of creative tools. Companies that had spent years refining niche AI solutions were suddenly irrelevant.

This environment fueled a common hypothesis among venture capitalists: foundation models were the only defensible game in town. Why build on shaky ground when you could own the bedrock? However, over the past year, it’s become increasingly clear that this thesis doesn’t hold water. The AI world is evolving in unexpected ways, and we’re seeing compelling counterexamples that show there’s more to AI than just the strength of the underlying model.

Drawing of von Kempelen's Mechanical Turk (Villatoro 2013)

Cracks in the Foundation Model Monopoly

Take the photo generation AI space, for instance. Smaller, nimble companies have carved out impressive niches, demonstrating that agility, creativity, and a deep understanding of specific customer needs can trump the brute force of a large foundation model. Meanwhile, an internal memo at Google captured the growing skepticism about foundation models as the ultimate moat, famously declaring, “We have No Moat, And Neither Does OpenAI” Even OpenAI itself has shifted focus—investing heavily in consumer-facing product experiences rather than relying solely on the strength of its models.

The latest release from DeepSeek underscores this point further: foundation language models are in a race to the bottom. As competition heats up, these models are rapidly commoditized, losing money at an alarming pace.

Finding Value Beyond the Obvious

This harsh reality forced us to rethink our approach. Foundation models were out. AI apps that could be eaten by a newer, shinier AI? Also out. So, what’s left?

Our investment thesis rests on a simple yet profound idea: there is immense value in AI applications that cannot be replaced by foundation models. But what does this mean in practice? It means targeting value chains with entrenched barriers where AI alone isn’t enough. These are areas where human challenges—politics, process, regulation, and integration across adversarial systems—create friction that technology cannot easily erase.

The Path to Enterprise and Beyond

This insight naturally ruled out consumer-focused applications. Consumer markets digital pipelines are typically fully automated, making them highly accessible to AI solutions that can quickly disrupt or absorb existing processes. Instead, we turned our focus to the enterprise world, particularly mid- to low-tech industries ripe with inefficiencies. These sectors often feature:

• Non-digitized processes stuck in the analog era.
• Silos of data and workflows fragmented by entrenched political and organizational barriers.
• Complex regulatory environments where compliance is non-negotiable.

These are challenges that can’t be solved by an AI model in isolation. They require deep integration into workflows, domain expertise, and a nuanced understanding of human dynamics.