Baseten CEO Tuhin Srivastava isn't just seeing growth; he's riding an AI inference rocket. In the last 24 months, his company has surged 30x, on track for over $1 billion in revenue this year. The reason? A widespread realization that “you can put AI everywhere,” Srivastava says. But here's the kicker: this isn't happening with off-the-shelf models. The real money and real utility come from custom AI.

Key Takeaways

  • Baseten, an AI inference company, is on track for over $1 billion in revenue this year, experiencing 30x growth driven by the broad realization that "you can put AI everywhere."
  • The market is rapidly shifting from generic AI solutions to "longtail models"—highly specialized, post-trained models optimized for specific use cases.
  • Over 95% of Baseten's production deployments involve custom models, where customers modify open-source weights with their own data for both quality and performance.
  • Simply running vanilla open-source models is becoming obsolete; real-world success and competitive advantage depend on customizing for specific application needs.

The AI Gold Rush is Custom-Built, Not Off-the-Shelf

Forget the hype about plugging in a generic, open-source model and calling it a day. Tuhin Srivastava sees something far more specific driving Baseten's staggering expansion. He claims that for almost all of Baseten's users, “it is all custom; it's basically so like 95% plus” of their production models. This isn't just a preference; it's a necessity for real-world performance.

When customers bring AI into production, they're not just accepting vanilla weights. They're making modifications with their own data, specializing models for their unique use cases. “No one is just running the vanilla open source weights,” Srivastava states plainly. The goal isn't just to make the model a little better, but to optimize it for specific quality and performance metrics that generic models simply can't hit. This shift towards highly customized, 'longtail' models is where the competitive edge now lies.

Stop Chasing Generic: Your Product Needs its Own AI Brain

Why does this matter? Because if your product's AI relies on a model everyone else can download, your AI isn't a differentiator; it's a commodity. Srivastava's insight pushes past the idea that AI is just another API call. He's seeing companies invest heavily in making AI their own, tailoring it to internal data and specific application demands. This could mean compiling models in different ways or fine-tuning them to reduce latency by milliseconds in a critical user flow. These micro-optimizations, invisible from the outside, translate directly to better user experience and better business outcomes.

This isn't just about tinkering; it's about building defensibility. When a company truly customizes its AI, it creates an intelligence layer that is deeply intertwined with its product and user base. “I think once you do that and then you kind of see what is possible then comes the whole custom model adoption,” Srivastava says. “I think that is all that is ahead of us today.” This implies a future where product success hinges on proprietary intelligence, built on top of—but far beyond—generic AI foundations.

What to Do With This

If your product relies on a vanilla open-source model, assume it's a short-term patch, not a long-term solution. Identify your core AI-powered feature. This week, task your engineering lead to prototype a custom fine-tuning approach using your unique user data, even if it's just a small dataset to start. Focus on how a specialized model could deliver a 2-5x improvement in a key metric (e.g., latency, accuracy for a specific query type, cost per inference) that a generic model can't touch. This is about building defensibility, not just features.