AI startups are rethinking where value lies. Foundation models once looked like the crown jewel, but now many see them as a commodity. Instead of building ever-bigger models, startups are focusing on fine-tuning and user interfaces.
At the recent Boxworks conference, the spotlight was not on the models themselves but on the software built on top of them.
The Slowdown in Scale
For years, training larger models on massive datasets gave big labs an edge. Companies like OpenAI, Google, and Anthropic seemed untouchable. But the scaling benefits are fading. Adding more data and compute no longer delivers the same leaps in performance.
Progress now comes from fine-tuning, reinforcement learning, and better product design. That means startups can compete without billions in training costs. Claude Code’s success shows that foundation model makers still excel here, but it is not the moat it used to be.
The Market Shifts
AI is splitting into many niches: coding tools, enterprise data, creative apps, and more. Owning the biggest model doesn’t guarantee leadership in these areas.
Open-source alternatives make things tougher. They give developers capable, low-cost models to build on. If customers can swap models without losing quality, large labs risk becoming back-end providers in a low-margin business. As one founder put it, “like selling coffee beans to Starbucks.”
Why It Matters
For most of the boom, foundation model companies looked like the clear winners. If AI was going to transform the world, then firms like OpenAI or Anthropic would hold the keys. Their platform advantage seemed strong.
The past year challenges that idea. Startups now mix and match models freely, often without users even noticing. A16z’s Martin Casado points out that OpenAI launched early models in coding, image, and video but lost ground in each case. His conclusion: there is no strong moat in today’s AI stack.
The Giants Still Have Strength
That doesn’t mean the big players are finished. They still control huge infrastructure, global brands, and vast cash reserves. OpenAI’s consumer products, for example, may prove hard to copy. And if research delivers breakthroughs in science or medicine, the advantage of foundation models could return quickly.
Still, the strategy of “bigger is always better” looks weaker than it did a year ago. Spending billions on scale alone is a major risk. Meta’s big bets on training may yet pay off — or prove costly.
Resources:
https://x.com/
https://knowledgenexuses.com/
