UC San Diego Built Its Own AI — and Now Sells It to Berkeley

UC San Diego and the University of Delaware refused to hand their data to AI vendors — and the decision sequence that let them stay in control without giving up commercial models

Every cabinet I talk to frames the AI infrastructure decision as a binary: build our own, or buy from a vendor. Build means control and cost and risk; buy means speed and convenience and dependence. Pick one. I covered the two-path version of this in Forbes last fall — cloud-first at Old Dominion, on-prem compute at RPI — and the framing held up. But the more I look at the institutions actually pulling ahead, the more the binary dissolves. The ones getting it right are not choosing build or buy. They are building the layer that controls their data and buying the horsepower that runs on top of it, and the sequence in which they did those two things is the whole lesson.

Two universities show the pattern most clearly. UC San Diego built TritonGPT on its own supercomputer — and now licenses it to UC Berkeley and other partner institutions. The University of Delaware built PATHS Engine to keep faculty intellectual property out of vendor hands. Neither rejected commercial AI. Both refused to let a vendor own the part that matters — the data, and the decisions about it.

UC San Diego: The Build That Started at the Service Desk

TritonGPT’s origin is the detail every leader should sit with, because it is the opposite of how these decisions are usually imagined. There was no master plan. In 2023, the chancellor asked IT Services whether the campus could use generative AI to make administration more efficient — deliberately starting with administration, because the academic senate held teaching and learning. UC San Diego had no contract at the time with a major AI provider. What it had was the San Diego Supercomputer Center, GPUs already running instructional and research workloads, and smart people.

The breakthrough came from a student. A staffer at the IT service desk — which UC San Diego staffs with students — downloaded an open-source orchestration framework called Onyx and used it to crawl the university’s Confluence knowledge base so he could answer support calls faster. He brought it to an internal AI community of practice, and the team realized the thing solving one student’s call-center problem could work enterprise-wide. Onyx, it turned out, was founded by two UC San Diego alumni. As Brett Pollak, the IT Services executive director who leads the AI program, put it to me: their solution let the model “draw information from our knowledge base” to answer questions. UC San Diego became the company’s first enterprise customer.

That is the build half. UC San Diego bought hardware, ran open-source models on its own infrastructure at SDSC, and wired in connectors to its enterprise systems and data warehouse through a retrieval-augmented architecture. Development began June 2023; a roughly 400-person administrative pilot ran by that October; the platform launched to all staff and faculty in spring 2024, with student access following in 2025. It now reaches 38,000 faculty and staff, with over 20 different agents in the ecosystem with a Socratic student tutor that integrates with their Canvas LMS and other repositories faculty use to teach courses.

The Part That Dissolves the Binary

Here is where build-versus-buy stops being a binary. UC San Diego did not refuse commercial models. It offers them via API— OpenAI through Azure, Anthropic’s models through AWS— but only inside contracted, single-tenant enclaves where, in Pollak’s framing, none of the data leaves. Open-source models run on-premises at SDSC for complete data control, while commercial models are accessed through agreements that keep UC San Diego data contained. Pollak’s reasoning—and the thesis of this piece—is simple: enterprise data should stay under UC San Diego’s control, without sacrificing users’ ability to choose the model that fits the task.

The other thing UC San Diego refused was a sealed product. It chose Onyx over competitors like Glean specifically because it wanted to build around the platform rather than inherit a black box. That single preference — flexibility over convenience — is what later let UC San Diego layer custom agents on top, and it is the difference between owning your AI strategy and renting it.

The proof that the build paid off: UC San Diego now sells it. The university licenses TritonGPT to peer institutions—Berkeley runs it as BearGPT, with UC Merced, UC Agriculture & Natural Resources, and Fairleigh Dickinson also on board—as a revenue-generating service that funds its own development team. The institution that built to control its data ended up with an asset it could sell to peers who couldn’t build it themselves. That’s what taking the power back looks like when it works: not just independence from vendors, but becoming the thing other institutions would otherwise have to buy.

Delaware: Buying Would Have Meant Surrendering Faculty IP

The University of Delaware reached the same sovereignty conclusion from the teaching side rather than the administrative one. PATHS Engine takes a faculty member’s own course materials — slides, PDFs, lecture transcripts — and turns them into structured learning objects the faculty member approves before any student sees them. The design principle, as the UD team described it in our May Use Case Lab, was a walled garden: faculty feel safe using AI and students feel permitted to use it, because both are working from the faculty member’s own data rather than a frontier model’s training set.

The reason this had to be built rather than bought is the same reason UC San Diego kept data at SDSC. To give students AI tools mapped to their actual course, UD would otherwise have had to send faculty intellectual property to a vendor — scraped, ingested, surrendered to a black box. UD’s answer was no scraping, an audit trail tracing every generated object back to its source, and faculty authority over what the system becomes. The newest phase, a federated learning pilot across Delaware institutions, extends the principle: institutions share how their models improved, never the underlying data. Same conviction as UC San Diego, different campus, different starting point — the institution keeps the data, and the vendor relationship is bounded to the parts that do not compromise it.

Data sovereignty is surrendered at a single moment: the day a cabinet leader signs a vendor contract without the language that specifies where the institution’s data goes and who may touch it — and by the time anyone notices, it cannot be recovered. What follows is not more of this argument; it is the diagnostic and operational apparatus that keeps that signature from being a mistake. It includes the contract language that forces a vendor to state exactly where data resides and who can access it; the data classification that determines which tier of institutional data may touch which model; the standing-committee structure that makes sovereignty a required question before any purchase clears; and the AI-Ready Institution Scorecard rubrics for the governance and operational-redesign pillars where these decisions are won or lost. Alongside them sits the Use Case Lab archive — the documented record of how real institutions actually made these calls, the precedent you will want in the room when it is your turn to decide.