Stanford MS&E435 Economics of the AI Supercycle | Spring 2026 | Applications, Applied AI

Key Concepts

• Inference: The process of running a trained AI model to make predictions or generate outputs.
• Post-training: The process of refining a base model (often open-source) with specific data and utility functions to optimize it for a particular use case.
• Frontier Models: The most advanced, large-scale, closed-source AI models (e.g., GPT-4, Claude).
• Heterogeneous Compute: The use of diverse hardware architectures (Nvidia GPUs, TPUs, custom silicon) to handle different parts of the AI stack.
• Modular Data Centers: A proposed framework to standardize the physical infrastructure of compute, similar to how shipping containers standardized global trade.
• CUDA: Nvidia’s parallel computing platform and programming model, which currently serves as the industry standard for AI development.

---

1. The Business of Base10

Base10 is an infrastructure company focused on production inference. The company’s core thesis is that while 95% of current AI spending goes toward frontier models, the future of defensible, profitable AI companies lies in custom models. Base10 provides the software stack, reliability, and multi-cloud orchestration that allows companies to run these custom models efficiently.

• Key Customers: Companies like Whisper Flow (speech-to-text) and Abridge (healthcare ambient scribing). These companies use dozens of specialized models that require high reliability and low latency, which Base10 optimizes.
• Value Proposition: Base10 abstracts the complexity of managing inference across multiple clouds, providing better performance, reliability, and observability than raw cloud providers (AWS, GCP, Azure).

2. The "Custom Model" Argument

Tuhin argues that relying solely on frontier models is a strategic risk.

• Economic Viability: Open-source models are currently about 90 days behind frontier models in capability but can be 70–90% cheaper to run.
• Defensibility: Using frontier models risks "leaking" proprietary user signals and workflows to the model providers. By owning their own intelligence through post-trained open-source models, companies can protect their unique value proposition.
• Scale: As companies grow, the cost of "token-trading" with frontier models becomes unsustainable. Shifting to post-trained models is often an existential requirement for achieving positive gross margins.

3. The Compute Ecosystem and Infrastructure

Base10 operates on a "rent-first" strategy but is moving toward ownership due to extreme compute scarcity.

• Compute Scarcity: Tuhin notes that the backlog for high-end GPUs (like Nvidia B200s) is currently 12–15 months. He describes the current compute market as a "drug market" with significant price volatility and slippage.
• Heterogeneous Compute: While Nvidia currently dominates due to its supply chain and the CUDA ecosystem, Tuhin expects the future to be heterogeneous. He suggests that architectures will eventually separate "prefill" (memory-bound) and "decode" (compute-bound) tasks onto different types of chips.
• The "Modular" Vision: Tuhin proposes that the next major innovation in infrastructure will be modular data centers. By standardizing the physical unit of compute, the industry could industrialize the build-out of data centers, effectively creating an "API for compute."

4. Strategic Perspectives

• The "East India Company" Analogy: Tuhin likens frontier labs to the East India Company—powerful entities that form partnerships while ultimately seeking to control the underlying "territory" (the intelligence layer).
• Open Source as National Security: Tuhin argues that the U.S. must foster a robust open-source AI ecosystem. He notes that currently, some of the best open-source models are emerging from China, and he believes it is a strategic necessity for the U.S. to maintain a competitive open-source alternative to prevent a duopoly of two or three companies controlling all global intelligence.
• Inference as the "Last Market": Tuhin posits that if AI continues to advance toward AGI, inference will eventually become the only remaining market, as the "intelligence" itself becomes the primary product.

5. Notable Quotes

• "Inference is about to go up a billionx."
• "If you're just trading tokens, it starts becoming very, very expensive... you need that path to profitability."
• "We're the rebellion. We're trying to arm the rebellion." (Regarding Base10's role in enabling custom models against the dominance of frontier labs).
• "I don't think [compute scarcity] is ever going to normalize... it just keeps compounding."

Synthesis

Base10 is positioning itself as the essential infrastructure layer for the "rebellion" against centralized AI control. By enabling companies to move from expensive, generic frontier models to cheaper, specialized, post-trained models, Base10 is betting on a future where intelligence is decentralized and owned by the application layer. The company’s transition from renting to owning compute reflects the harsh reality of the current hardware market, where access to silicon is the ultimate strategic advantage.