The Startups Building on Nvidia Compute

Key Concepts

• Compute Constraints: The limited availability of high-performance hardware (specifically Nvidia H100s) required for training and running AI models.
• Frontier Performance: The necessity for startups to use the latest, most powerful hardware to experiment and achieve state-of-the-art results.
• Long-Horizon Agents: AI systems capable of executing complex, multi-step tasks over extended periods, which are currently driving significant revenue growth.
• Post-training: The process of refining smaller models after initial training to optimize for cost and specific task performance.
• Total Addressable Market (TAM): The projected $26.5 trillion opportunity for AI, specifically within the enterprise sector.

---

1. The Compute Shortage and Startup Strategy

Sarah Guo, founder of Conviction, highlights that compute is the primary constraint for AI startups.

• Direct Procurement: To mitigate supply risks, Conviction took the proactive step of purchasing H100 nodes directly for their portfolio companies, acting as a buffer against market volatility.
• The Lifecycle of Compute: Startups typically begin by utilizing "frontier" hardware (Nvidia chips) to experiment. As they mature, they shift focus toward cost-efficiency by "post-training" smaller models, which allows for higher token usage on specific tasks, thereby improving the user experience.
• Market Stress: Guo notes that the ecosystem has faced two quarters of increasing stress regarding supply access. She describes the unprecedented difficulty of attempting to secure $100 million compute commitments, a scenario she had not encountered previously.

2. The "Parabolic" Demand for AI

Guo validates Jensen Huang’s (CEO of Nvidia) assertion that demand for compute is "parabolic."

• Evidence of Demand: The rapid revenue growth in AI-driven cloud services is largely attributed to the success of "long-horizon agents." These agents automate complex tasks in the knowledge economy, moving beyond simple text generation to functional automation.
• The Human Factor: Guo suggests that while humans struggle to grasp the scale of this growth, the shift toward automating existing enterprise tasks—as described by Andrej Karpathy—is the primary driver of this massive market expansion.

3. Enterprise AI and the SpaceX S1

The discussion touches on the recent SpaceX S1 filing, which identifies a $26.5 trillion TAM for AI, with a heavy emphasis on enterprise applications.

• Strategic Intent: Guo interprets the inclusion of "Enterprise AI" in the SpaceX S1 as a deliberate strategic commitment by Elon Musk.
• Automation as the Core: The fundamental value proposition of enterprise AI is the automation of existing human tasks. By providing models with the necessary tools and "harnesses," companies can effectively offload labor-intensive processes to AI.

4. Value Distribution: Infrastructure vs. Application

A central debate presented is where the value in the AI stack will ultimately reside:

• Infrastructure vs. Models vs. Applications: Guo argues that while infrastructure (compute and hardware) is currently "extraordinarily valuable," the long-term success of companies like SpaceX/xAI depends on whether they can successfully capture value at the model and application layers.
• The "Winner" Perspective: Guo posits that regardless of the specific layer, companies with the infrastructure and the capability to build are positioned to be profitable. The open question remains whether owning the entire stack (infrastructure + model + application) is a requirement for long-term dominance.

---

Synthesis and Conclusion

The current AI landscape is defined by a severe supply-demand imbalance, where compute availability acts as the primary bottleneck for innovation. Startups are forced to navigate a complex path from high-cost frontier experimentation to optimized, smaller-model deployment. However, the underlying driver—the automation of the knowledge economy through long-horizon agents—is creating a massive, multi-trillion-dollar opportunity. The ultimate winners will likely be those who can bridge the gap between raw infrastructure and high-value enterprise applications, a challenge that major players like Elon Musk are now explicitly targeting.