Ex-Siri product manager Ivan Lee on why Datasaur runs open weights models inside client servers, and why most enterprise tasks don't need frontier models.
Enterprises can use AI without sending sensitive data to OpenAI or Anthropic by running open weights models on their own servers, and for most business tasks those models are now good enough. That is the case Ivan Lee, co-founder and CEO of Datasaur, made on Episode 56 of Built This Week. Ivan built AI products at Yahoo and Apple, where he worked on Siri, and Datasaur is his second startup.
Ivan Lee says Datasaur "fundamentally builds private AI for enterprise." The AI is deployed inside the client's own servers and infrastructure, so sensitive data such as clinical patient records or social security numbers is never sent to Anthropic or OpenAI. With new security concerns, and labs that need more data to train their next models, he says clients can rest assured their data never reaches them.
The way Datasaur does it is by deploying open weights models directly onto the client's servers, along with the harness, the context and everything else the agents need to work. Its clients are mostly in healthcare, finance, legal and the public sector: the organizations that care most about privacy and security. Ivan has been working on Datasaur for eight years.
Watch Ivan Lee on open weights vs frontier models in Episode 56 (starts at 5:46) · Episode 56 notes and transcript
Sam asked which industries surprised him. Ivan's answer was that healthcare and legal, two fields known for moving slowly, have been adopting LLMs and AI agents faster than perhaps any technology in history, cloud and mobile included. The use case is that compelling. They still have to deal with HIPAA, SOC 2 and cost, which is how the private AI conversation started.
What changed this year, he says, is that companies started to see their domain expertise as their moat. He pointed to Anthropic partnering with companies and then launching products in the same categories, naming Claude Design and Claude Legal. That made firms ask whether they want to teach a model lab exactly how their complex processes work.
Pair that alongside the fact that there's been a whole new wave of open weight models this year, and suddenly, there is both motive as well as opportunity to build your own AI and turn that into a strategic asset.
— Ivan Lee, Episode 56
We have covered how AI is changing care delivery in other episodes, collected on our AI in healthcare page.
Ivan says this is the number one question clients ask. For the first couple of years after the technology arrived there was a huge gap, and OpenAI looked impossible to catch. By March of this year, he says, people in Silicon Valley questioned your sanity if you weren't using Anthropic's Claude. In his view the models are now largely commoditized, even among the proprietary ones, and a wave of open weights releases has followed, including models from China, a new release from Meta, and Google's Gemma 4, which Datasaur is a big fan of.
He does not claim open models win everywhere. If a client were working on curing cancer, he would want them on the top frontier models. But for everyday enterprise work, his view is blunt:
But if you're using this to summarize your emails, if you're using this to recap your meeting notes, you don't need Fable for that. That is absolutely overkill.
— Ivan Lee, Episode 56
He backed this up with a chart from the benchmarking site Artificial Analysis. His line in the sand is Claude 4.5: agents, he says, didn't really start working until December 2025. Every open weights model on his second chart cleared that bar by a significant margin. He estimates only 5% to 10% of tasks need the top frontier models, and says we could pause model development for a decade and enterprises would still find a lot of gains using open weights models alone.
Jordan asked how hardware shortages affect private deployments. Ivan pointed to Meta's new 100 billion parameter model, which Meta says is built and tested to run on two M5 Mac Studios. That is expensive for a home, but for an average company it is low five figures to own a model that size. Some Datasaur clients have bought millions of dollars of hardware for their private AI, and Ivan expects that cost to keep falling as stronger models run on more commoditized hardware. We went deeper on the economics of running models in Episode 37.
Jordan raised the Jevons paradox: if models get cheaper, does usage explode? Ivan thinks there are many more tokens to be consumed in the enterprise, and gave the numbers:
Your average chatbot query is 200 to 2,000 tokens. Your average agentic task is 3,000,000 tokens. It's over three orders of magnitude more than your ChatGPT query.
— Ivan Lee, Episode 56
He used the crossing the chasm adoption curve to explain the second half of his argument. Chatbots, with a billion monthly active users across ChatGPT and Gemini, have crossed into the early majority. Agents are still with early adopters. When they cross, he expects token usage to skyrocket as companies move beyond their current chatbot tools, and the GPU crunch Silicon Valley felt earlier this year to hit the rest of the world.
Jordan agreed and added that agents also collapse software spend: instead of paying for a dedicated outbound sales tool, he can have his agent do the work, shifting spend from software to tokens.
Ivan calls it AI native services. To the client it feels like white glove consulting: they describe the workflow they run or want, and Datasaur recommends a model and explains how much work it will take. Sometimes it is as simple as deploying a model and connecting it to the client's Dropbox or OneDrive. Sometimes it is a more complicated agent.
Behind the scenes, Datasaur runs what Ivan calls a proprietary model factory. It takes the requirements, workflow, benchmarks and historic data, suggests the best model, fine tunes it and handles the harness engineering. To the client it looks like fast consulting. Internally it is a productized platform built privacy first.
In the news segment we talked about Meta's Muse and Anthropic's new Duet. Ivan, who worked on Siri at Apple, says enterprise agents charged ahead first, starting with coding agents, then customer support, marketing and sales. Personal agents lagged until Muse, which he says found a UI and used your own context to take the first real pass at it. He sees consumer adoption as good for Datasaur: clients who used ChatGPT at home wanted it at work, and he expects the same with agents.
Jordan thinks Google is being underestimated because it is already connected to people's email and data, and expects Microsoft to answer with an updated Copilot. Ivan closed with a hot take:
I think agents are just universally a better fit for the enterprise than they are for the consumer world at this point.
— Ivan Lee, Episode 56
His reasoning: companies already have operational playbooks written to train the next human employee, while consumers all want different airports and seats. His own favorite personal use is an MCP server he built for his Whoop band. Claude reads how he slept and his daily activity and gives him a training plan that adapts when he is busy or at a conference.
Ivan covers the benchmark charts, the Mac Studio math and the Datasaur model factory in more detail on camera. Watch Episode 56 of Built This Week, where the full transcript is also on the page.
