The State of AI: Growth, Fragmentation, and the Next Wave

Market Scale and Velocity

• AI companies are growing larger and faster than anticipated, with frontier model labs (e.g., OpenAI, Anthropic) surpassing early revenue ramps of top SaaS companies and hyperscalers.
• Model inference costs have decreased 10x year-over-year, aligning market conditions for rapid application growth.
• The market is fragmenting rather than consolidating, with distinct subspaces (language models, diffusion models, apps, tooling) requiring unique strategies rather than a single "zero-sum" approach.

App Layer Dynamics vs. Incumbents

• Specialized AI-native apps are outperforming traditional SaaS by delivering 10x+ customer experience improvements, compared to the incremental 25-50% gains typical of SaaS 2.0.
• AI-native companies are achieving time-to-$100M ARR faster than SaaS counterparts, driven by replacing services budgets and high ROI rather than just software licensing.
• Traditional SaaS firms face an "innovator's dilemma," where existing revenue-generating products and remote work cultures hinder their ability to pivot effectively despite superior distribution.
• The term "GPT wrapper" is rejected as a distinct category; building software on top of models involves significant complexity, integrations, and value creation akin to traditional cloud software.

Defensibility and Customer Retention

• Foundational models alone do not provide inherent defensibility due to commoditization; winning companies must layer complex workflows, deep data integrations, and two-sided marketplaces to secure retention.
• Brand effects are re-emerging as a critical moat, similar to the early internet (e.g., Google, Amazon), where name recognition drives adoption even when functional parity exists with competitors.
• Consumer and prosumer usage is acting as a primary distribution channel, creating a pipeline for enterprise adoption with higher-than-normal enterprise deal flow (e.g., new AE closing $1M on day one).
• While some high-growth apps exhibit lower gross dollar retention (GDR) compared to 2010s SaaS leaders, the firm is exercising caution regarding valuations that demand stickiness without proven enterprise conversion.

Performance Metrics and ROI

• Cursor users report productivity gains ranging from 30-50% generally, with specific cases of 10x individual team productivity and 90% of company code being AI-generated.
• Decagon customers report customer support cost reductions of up to 80%, with deflection rates jumping from 30% to 60-80% and customer satisfaction scores doubling.
• Enterprise purchasing has shifted from "vibe-based" experimentation to tangible ROI focus, requiring hard data on productivity and cost savings.

Investment Strategy and Pitfalls

• Foundation Models: The firm avoids early-stage bets on state-of-the-art model development due to intense competition and heavy subsidization by incumbents (Meta, Google); instead, they back premium, proven teams capable of raising capital (e.g., investment in Ilya Sutskever's ventures).
• Avoidance Factors: The firm rejects "researcher vagaries" (excessive academic focus over product) and avoids "tell the story" companies that raise massive rounds without traction, as high capital raises create intense pressure to deliver immediate performance.
• Capital Constraints: Early, aggressive investing in non-winning concepts creates conflicts of interest, preventing funds from deploying capital into proven winners.
• China Factor: Chinese competition is viewed as a mixed blessing; while they offer strong open-source models and cheap data access, they historically struggle to build software for prosumer/enterprise markets outside China.

Forward-Looking Statements and Conclusion

• The firm advises that "heat cannot be confused with momentum" and that the stakes in AI are higher than in previous software cycles.
• Being "in the field" is mandatory; firms that are disengaged are becoming irrelevant.
• Success requires a "smarter" approach to betting than ever before, acknowledging that the market is larger and more complex than previously modeled.