‘selling coffee beans to starbucks’ — how the ai boom could leave ai’s biggest companies behind

The Evolving Significance of Foundation Models in AI

The importance of foundation models is a question that has been frequently raised, particularly in discussions with AI-focused startups. These businesses are demonstrating increasing acceptance of companies initially considered merely “GPT wrappers” – those that construct user interfaces leveraging pre-existing AI models such as ChatGPT.

A Shift in Focus for AI Startups

Currently, startup teams are prioritizing the customization of AI models for specialized applications and refining user interfaces. They view the foundation model itself as a replaceable component, adaptable as needed. This trend was particularly evident at the recent BoxWorks conference, which largely showcased software built upon AI models.

Diminishing Returns from Pre-Training

A key factor driving this change is the slowing of scalability benefits derived from pre-training – the initial process of training AI models using extensive datasets, a domain exclusively held by foundation models. While AI development continues, the substantial early gains from hyperscaled foundational models are experiencing diminishing returns.

Consequently, attention is shifting towards post-training techniques and reinforcement learning as avenues for future advancements. Improving an AI coding tool, for example, is more effectively achieved through fine-tuning and interface design rather than further investment in pre-training server time.

The Changing Competitive Landscape

As demonstrated by Anthropic’s Claude Code, companies specializing in foundation models are proficient in these alternative areas as well. However, this expertise no longer provides as significant a competitive edge as it once did. The AI competitive landscape is evolving, potentially lessening the advantages held by the largest AI labs.

Instead of a relentless pursuit of all-encompassing Artificial General Intelligence (AGI), capable of matching or surpassing human cognitive abilities, the near future appears to be characterized by a proliferation of specialized businesses – encompassing software development, enterprise data management, image generation, and more.

Potential Commoditization of Foundation Models

Beyond an initial first-mover advantage, it’s unclear whether constructing a foundation model offers any substantial benefit within these specific business areas. Furthermore, the growing availability of open-source alternatives suggests that foundation models may lack pricing power if they fail to compete effectively at the application level.

This scenario could relegate companies like OpenAI and Anthropic to the role of back-end suppliers in a low-margin commodity market – akin to providing raw materials, as one founder described it, “like selling coffee beans to Starbucks.”

A Dramatic Shift in the AI Business Model

Such a transformation would represent a significant departure from the current state of the AI industry. Throughout the recent boom, the success of AI has been intrinsically linked to the success of companies building foundation models – notably OpenAI, Anthropic, and Google.

Optimism regarding AI has historically implied that these companies would become generationally important entities, with the potential to shape the future. Debate centered on which company would ultimately prevail, but it was widely accepted that a foundation model company would ultimately control the core infrastructure.

The Rise of Interchangeable Models

Previously, foundation model development was the sole AI business, and the rapid pace of progress seemed to solidify their leading position. Silicon Valley’s preference for platform advantages further reinforced the assumption that the majority of benefits would accrue to these foundational companies, having undertaken the most challenging work.

However, the past year has introduced complexities. Numerous successful third-party AI services have emerged, often utilizing foundation models interchangeably. Startups are increasingly indifferent to whether their product relies on GPT-5, Claude, or Gemini, anticipating seamless model switching without user disruption.

Foundation models continue to advance, but maintaining a dominant advantage for any single company appears increasingly improbable.

Limited First-Mover Advantage

Evidence suggests a limited first-mover advantage. Venture capitalist Martin Casado of a16z recently noted on a podcast that OpenAI was the first to release coding, image, and video generative models, only to subsequently lose leadership in all three categories. “As far as we can tell, there is no inherent moat in the technology stack for AI,” Casado concluded.

Remaining Advantages and Future Uncertainties

Despite this, it’s premature to dismiss foundation model companies. They retain considerable advantages, including brand recognition, established infrastructure, and substantial financial resources. OpenAI’s consumer-facing business may prove more resilient than its coding offerings, and further advantages could emerge as the sector evolves.

The rapid pace of AI development means that the current focus on post-training could easily shift in the coming months. Moreover, the pursuit of general intelligence could yield breakthroughs in fields like pharmaceuticals or materials science, fundamentally altering our understanding of AI model value.

However, for the time being, the strategy of continually expanding foundation model size appears less compelling than it did previously, making Meta’s substantial investment seem increasingly risky.