The Truth About The AI Bubble

Key Concepts

• AI Economy Stabilization: The AI landscape has matured, with defined layers (infrastructure, model, application) and established playbooks for building AI-native companies.
• LLM Preference Shift: Anthropic has surpassed OpenAI as the preferred Large Language Model (LLM) API among Y Combinator (YC) companies, driven by performance in coding tasks.
• Model Commoditization: The increasing capabilities of models like Gemini are leading to a commoditization effect, shifting value towards the application layer.
• Vibe Coding & Agent Development: A significant growth area in AI, with models like Anthropic excelling in coding-related tasks.
• Infrastructure Constraints: Challenges in power generation and data center capacity are driving innovation in areas like space-based computing.
• Fine-tuning ROI: The value of extensive fine-tuning of LLMs is being questioned, with some companies finding limited benefit.
• Human Resistance to Change: Societal and organizational inertia are slowing the widespread adoption of AI, providing a buffer against rapid disruption.

The Evolving AI Landscape in 2025

The discussion centers around key observations from 2025 regarding the development and stabilization of the AI economy. A significant shift has occurred, moving away from the volatile, rapidly changing environment of late 2024. The AI ecosystem now appears structured into three distinct layers: infrastructure (providing the computing power), model (the LLMs themselves), and application (companies building on top of the models). This structure suggests a more predictable path to profitability for companies operating within each layer. The speakers note a feeling that “everyone is going to make a lot of money” and that a “relative playbook” for building AI-native companies has emerged. This contrasts with the earlier period where survival and waiting for the next major announcement were key strategies.

The Changing LLM Landscape: Anthropic’s Rise

A surprising trend observed by Y Combinator (YC) during their Winter 26 selection cycle is the rise of Anthropic as the preferred LLM API. Previously, OpenAI held a dominant position, reaching over 90% usage in earlier batches. However, Anthropic’s share has grown dramatically, exceeding OpenAI’s in the most recent cycle. This growth, a “hockey stick” pattern occurring in the last 3-6 months, is attributed to Anthropic’s strong performance in “vibe coding” tools and coding agents. Tom Brown, during a conversation with the podcast hosts, revealed that excelling in coding was a deliberate internal evaluation goal for Anthropic, influencing their model design. While coding isn’t the primary use case for most applications, familiarity with Claude’s personality and capabilities appears to influence founder choices even for non-coding projects.

Gemini is also gaining traction, climbing from single-digit percentages to approximately 23% usage in Winter 26, with the hosts personally impressed by Gemini 3.0’s quality.

Model Personalities and User Preference

The speakers discuss the distinct “personalities” of different LLMs. OpenAI is described as having a “black cat energy,” while Anthropic is characterized as a “happy, go-lucky golden retriever.” Gemini is positioned as somewhere in between. One host has switched to Gemini as their primary model, citing its superior reasoning abilities and grounding in Google’s search index, providing more accurate real-time information compared to tools like Perplexity. However, they acknowledge the difficulty in isolating the impact of the model versus the tooling used to access it. Another host remains loyal to ChatGPT, valuing its “sticky” memory and understanding of their individual preferences and context. They believe memory is becoming a significant competitive advantage for consumer-facing AI applications.

The Rise of Orchestration Layers and Model Agnosticism

A key trend is the emergence of companies building “orchestration layers” that abstract away the underlying LLM. These companies, often at the Series B level, are no longer loyal to a single model provider. Instead, they dynamically swap models based on performance for specific tasks. For example, one startup uses Gemini 3 for context engineering and then feeds that output into OpenAI for execution. This approach allows them to leverage the strengths of different models and adapt to new releases. This is driven by proprietary evaluation frameworks and the availability of specialized datasets within regulated industries. The speakers compare this to the evolution of the CPU market with Intel and AMD, where users can choose the best processor for their needs.

Infrastructure Challenges and Innovative Solutions

The discussion highlights growing infrastructure constraints, particularly in power generation and data center capacity. This is driving innovation in unconventional areas like space-based computing. Companies like StarCloud (initially met with skepticism) and Google are now exploring data centers in space. The speakers note that even companies like Boom Supersonic are pivoting to address the power supply shortage, demonstrating the severity of the issue. YC has several companies tackling different aspects of the infrastructure problem, including fusion energy (Zephr Fusion) and space-based data centers.

Fine-tuning, ROI, and the Commoditization of Models

The value of extensive fine-tuning of LLMs is being questioned. The speakers point out that some companies have wasted significant capital on fine-tuning without achieving a substantial advantage. This is particularly true as newer, more powerful models are released. The commoditization of models, driven by improvements in Gemini and other alternatives, is shifting the focus back to the application layer and the ability of startups to deliver unique value.

The AI Bubble Debate and Long-Term Prospects

The speakers address concerns about an “AI bubble,” comparing it to the telecom bubble of the 1990s. They argue that the current situation is fundamentally different, as the abundance of compute power (similar to the excess bandwidth of the telecom era) creates opportunities for innovation. They emphasize that the glut of resources benefits startups, as it lowers costs and increases competition. They also point to the inherent resistance to change within organizations as a factor that will slow down the disruptive potential of AI, providing time for society to adapt. They reference economist Carlo Perez’s work on technology revolutions, which identifies two phases: installation (heavy capex investment) and deployment (widespread adoption). They believe we are currently transitioning from the installation phase to the deployment phase, which is good news for startups.

The Return to Normal Startup Dynamics

The speakers observe a shift away from the unusually easy startup idea generation of 2024, where major AI announcements frequently created new opportunities. Idea generation is returning to “normal levels of difficulty.” They also note that the initial wave of AI-native companies (like Harvey) are facing increased competition from newer entrants (like Lora and Giga). While early companies may have achieved rapid growth with minimal hiring, the current trend suggests that scaling requires building larger teams to meet growing customer expectations. The “reverse flex” of companies boasting high revenue with small teams (like Gamma) is seen as a positive sign.

Conclusion

The AI landscape in 2025 is characterized by stabilization, increased competition, and a growing focus on practical applications. While infrastructure challenges remain, innovation is driving solutions in areas like space-based computing. The commoditization of models is shifting value towards the application layer, creating opportunities for startups to build unique and valuable products. The initial hype surrounding AI has subsided, replaced by a more pragmatic and sustainable approach to development and deployment. The era of the single-person, trillion-dollar AI company is not yet here, but the potential for highly efficient and impactful AI-driven businesses remains strong.