“Selling coffee beans to Starbucks” – How the AI ​​boom can leave behind the biggest AI companies

How important is the basic model?

While that may seem like a ridiculous question, conversations with AI startups are becoming increasingly comfortable for businesses that have been dismissed as “GPT rappers” and for companies that build interfaces on existing AI models such as ChatGPT. Recently, startup teams have focused on customizing AI models for specific tasks and interface tasks, and see the underlying model as a product that can be replaced or taken out as needed. This approach was featured in particular at last week’s Boxworks Conference. This appears to be completely dedicated to user-friendly software built on top of AI models.

Part of what drives this is that the benefits of pre-training scaling (the original process of teaching AI models using large datasets, the only domain of the foundation model, has slowed down. That doesn’t mean that AI has stopped progress, but the early benefits of hyperskulled foundation models collided with reduced revenues, and focused attention on post-training learning and reinforcement learning as a source of future advances. If you want to create a better AI coding tool, it’s better to work on tweaking and interface design rather than spending billions more dollars worth of server time before training. As Anthropic’s Claude Code success shows, Foundation Model Companies are also very good in these other areas, but not as durable as they used to be.

In short, the competitive landscape of AI is changing in ways that undermine the benefits of the biggest AI lab. Instead of all powerful AGI races that can be tailored or surpassed by human capabilities across all cognitive tasks, the immediate future looks like a surge in individual businesses, including software development, enterprise data management, image generation, and more. Aside from the benefits of first-movers, it is not clear that building a foundation model will give all the benefits to those businesses. Worse, the wealth of open source alternatives means that if you lose competition in the application layer, your foundation model may not have price leverage. This will turn companies like Openai and humanity into back-end suppliers of low-margin product businesses.

It is difficult to exaggerate how dramatic this is for the AI ​​business. Through the modern boom, AI’s success has been inexplicable from the success of companies building foundation models, especially Openai, humanity, and Google. Being bullish on AI meant that we believe the transformational impact of AI would make these a generational crucial enterprise. While we could debate which companies would come to the top, it was clear that some foundation model companies would acquire keys to the kingdom.

At the time, there were many reasons to think this was true. For years, development of foundation models has been the only AI business. And Silicon Valley has always had a deep love for the platform advantage. But even if AI models end up making money, the Lion’s share of profits would return to the foundation model companies, where replicating the most difficult task.

Over the past year, the story has become more complicated. There are many successful third-party AI services, but they tend to use the basic models interchangeably. For startups, it doesn’t matter whether the product is on top of GPT-5, Claude, or Gemini, and we expect that end users will be able to switch models in Mid-Release without realizing the difference. The foundation model continues to make real progress, but it seems no longer reasonable for one company to maintain a great advantage enough to dominate the industry.

There are already many indications, indicating that the benefits of a first me bar are not very good. As A16Z venture capitalist Martin Casado pointed out in a recent podcast, Openai was the first lab to publish code models and image and video generation models, only losing all three categories to its competitors. “As far as we know, there is no moat inherent to the AI ​​technology stack,” Cassad concluded.

Of course, we should not count foundation model companies yet. There are still plenty of durable benefits, including brand recognition, infrastructure and unthinkable cash reserves. Openai’s consumer business may be more difficult to replicate than the coding business, with other benefits likely to emerge as the sector matures. Given the fast pace of AI development, current interest after training could easily reverse courses in the next six months. Most uncertain, the competition for general intelligence could be rewarded with new breakthroughs in medicine and materials science, fundamentally changing ideas about what would make AI models valuable.

In the meantime, however, the strategy of building an unprecedented foundational model appears to be far less attractive than last year. And Meta’s billions of dollars are beginning to be in great danger.