Greylock

From Rented to Owned Intelligence with Baseten

The video introduces Baseten, an AI infrastructure firm that helps businesses move from "rented" AI—pay‑per‑token, shared‑endpoint models—to "owned" intelligence, where firms fine‑tune and host their own models, controlling quality, latency, and expenses. Baseten’s vision is a future populated by many specialized models, each managed through continual‑learning pipelines that keep performance aligned with evolving data. Key insights include a three‑tier product stack: shared open‑source APIs for quick entry, dedicated inference for custom SLAs, and a post‑training suite that bridges training and deployment. Adoption varies by company stage: early startups prototype with closed models, hyper‑growth AI firms blend open‑source and custom models, while enterprises—particularly in healthcare and finance—are beginning to invest heavily as ROI becomes evident. The discussion cites real‑world examples such as Intercom, Decagon, Cursor, and healthcare pilots, illustrating how post‑trained models improve outcomes. Baseten’s recent acquisition of Parsed adds tooling for continual learning, enabling models to evolve in‑context during long‑horizon, agentic tasks. The company positions itself as a customer‑centric inference platform rather than a pure research lab. Implications are significant: businesses gain cost predictability, regulatory compliance, and performance guarantees while avoiding vendor lock‑in. As AI workloads shift toward agentic, tool‑rich workflows, Baseten’s expanding inference stack—combining reliable runtimes with higher‑level primitives—could become a critical layer for scaling next‑generation AI applications.