Stanford MS&E435: Baseten and how unit economics reshapes strategy (Category #AI)
What I liked about Apoorv Agrawal's conversation with Baseten co-founder Tuhin Srivastava in Stanford's Economics of the AI Supercycle course was how business growth forces companies to reconsider architecture. A company starts with ready-made APIs, then develops its own models. Meanwhile, its infrastructure provider starts considering GPUs of its own. Both gradually change their answer to the same question: what can we delegate?
The initial logic is clear: run fast. Take a leading provider's model, build a product, and test demand. Before you know whether anyone needs your service, building a model operations team is expensive. Finding out what people will pay for is much more useful.
Then come users, recurring tasks, and the inference bill. This is where run effective begins: adapting an open-weight model to your particular use case can make sense. Every request pays for a large model's general capabilities, even though the product needs only some of them.
Srivastava estimates that open models can run 70–90% cheaper than leading closed models. This is the speaker's estimate, not a promise of those savings for every product. You have to calculate the cost of a successfully completed task, including training, quality evaluations, retries, and the team's work. At sufficient volume, the savings begin to pay for all that effort.
But there is an entire profession between downloading weights and being responsible for a working service. You need people to prepare data, adapt the model, deploy it, achieve the required latency, and investigate why everything failed overnight. This is where Baseten comes in: as Srivastava describes it, the platform handles training and inference infrastructure. The customer brings data and defines a good outcome. Understanding your own product is harder to delegate.
Competition adds another reason. Here is a clear example from outside the lecture. Anthropic supplies models to application developers while also developing Claude Code, which serves the same audience as Cursor. A supplier may decide to sell the end product itself. Cursor's custom models actually predate Claude Code: Tab has used a custom model since March 2024. The company later introduced Composer 2, based on Kimi K2.5, with continued pretraining and reinforcement learning. Incidentally, Baseten does work with Cursor, although that does not mean it serves all of Cursor's inference.
I would separate this competition from the fear that a provider will train on your data. Data use is governed by the terms of the relationship; Anthropic already excludes commercial API data from training by default. But that restriction does not prevent a supplier from independently entering your market. Custom models provide more control over costs and product development.
Then the same story repeats one layer down.
Baseten builds a software platform on top of other companies' clouds: optimizing inference, aggregating available capacity, and providing fault tolerance. According to Srivastava, GPU scarcity is pushing the company toward owning some hardware to secure resources for customers' future demand. Hardware brings acquisition costs and the risks of underutilization and obsolescence. During a shortage, an intermediary that can find capacity across suppliers becomes especially useful. But if you promise customers a reliable service, eventually you may need to take control of the supply of a critical resource. Through long-term leases, owned hardware, or both.
There is a symmetry here: Cursor increases control over models, Baseten over compute. Scale changes the boundary between what is convenient to buy as a service and what you need to take responsibility for yourself. That boundary is worth revisiting alongside the infrastructure bill.
#AI #Architecture #PlatformEngineering #Product #Management