AI-Services Startup Baseten Nabs $13 Billion Valuation

Key Concepts

• AI Inference: The process of running a trained AI model to make predictions or generate content.
• Open Source Models: AI models with publicly available weights, allowing for customization, self-hosting, and reduced dependency on proprietary providers.
• Frontier Models: The most advanced, large-scale proprietary AI models (e.g., from OpenAI, Anthropic, Google) that represent the current state-of-the-art in reasoning and intelligence.
• Post-Training/Fine-Tuning: The process of taking a base model and further training it on specific datasets to optimize it for particular tasks or domains.
• Compute Diversification: The strategy of sourcing hardware (GPUs) from multiple cloud providers and clusters to mitigate supply chain constraints.
• Unit Economics: The profitability of AI applications, specifically focusing on the cost-per-token versus the value generated.

---

1. Main Topics and Key Points

• Funding and Valuation: BaseTen has raised $1.5 billion across two tranches, reaching valuations of $11 billion and $13 billion. The capital is earmarked for procuring compute capacity and hiring specialized infrastructure and research engineers.
• The Shift to Inference: The market is moving from simple Q&A interfaces to complex, multi-agent workflows. Modern AI applications now involve retrieving, analyzing, synthesizing, and verifying data, which triggers thousands of inference requests per user interaction.
• The "Three C's" of Open Source: Apurv Agrawal (Altimeter Capital) argues that enterprises choose open source for three reasons:
  • Capability: Open source models are reaching parity with frontier models.
  • Control: Enterprises retain their data and intelligence rather than "renting" it from a third-party API.
  • Cost: Open source allows for better unit economics, which is essential for scaling AI profitably.

2. Real-World Applications and Case Studies

• Customer Examples: BaseTen supports companies like Cursor (AI code editor), Abridge (medical documentation), Open Evidence (medical research), and Harvey (legal AI).
• Harvey Case Study: Harvey achieved "frontier-level capabilities" by post-training an open-source model, allowing them to maintain high performance while gaining better cost control and data sovereignty compared to using a closed-source API.

3. Methodologies and Frameworks

• The "Five-Layer Cake": A framework (attributed to Jensen Huang) describing the AI stack. BaseTen operates at the layer that enables companies to harness model power and combine it with proprietary enterprise workflows.
• Compute Strategy: To overcome supply constraints, BaseTen does not rely on a single provider. They utilize 18 different cloud providers across 90 distinct clusters to ensure flexibility and availability.
• The Profitability Curve: Companies often start by aggressively deploying AI, but eventually face a "profitability wall." BaseTen helps these companies transition to optimized, post-trained open-source models to make their unit economics sustainable.

4. Key Arguments and Perspectives

• The "Game of Emperors": Apurv Agrawal notes that the model layer is a "knife fight" between a few massive players (Anthropic, OpenAI, Google, etc.). Most enterprises should not try to build their own foundation models but should instead focus on the application layer using open-source models.
• The "35 Companies" Theory: Citing Jonathan Ross, the speakers suggest that while there are thousands of AI companies, roughly 35 companies drive 99% of all inference demand. BaseTen focuses on the "next 30" and the "next 1,000" companies that need to own their intelligence rather than rent it.
• Seasonal Model Dominance: The speakers argue that the "best" model changes constantly (e.g., DeepMind, GLM, Llama). BaseTen’s platform is designed to be model-agnostic, allowing customers to swap or combine models as the landscape evolves.

5. Notable Quotes

• Tuhin Srivastava (CEO, BaseTen): "You cannot be reliant on one compute source... you have to diversify."
• Apurv Agrawal (Altimeter Capital): "Some truths cannot be said enough times. And one of those is that inference is going to be one of the largest, if not the largest, markets not in AI, in the world."
• Apurv Agrawal: "You've got to compound your unique advantages as an enterprise—your data, your knowledge, your know-how—in a way that you're not giving away your intelligence to rent."

6. Synthesis and Conclusion

BaseTen is positioning itself as the critical infrastructure layer for the "next wave" of AI adoption. By enabling enterprises to move away from expensive, opaque, and rented proprietary models toward optimized, self-hosted, and fine-tuned open-source models, they are solving the primary bottleneck for AI scaling: unit economics. The company’s massive funding reflects the industry's shift from the "discovery" phase of AI to the "operational" phase, where control, cost-efficiency, and the ability to integrate proprietary data into AI workflows are the primary drivers of competitive advantage.