The rapid rise of capable open-weight models such as DeepSeek and Kimi suggests AI is following the same trajectory. Why pay a premium for third-party services like Claude or ChatGPT when you can deploy a powerful model on your own infrastructure? A growing number of companies have already reached that conclusion. Airbnb has adopted Alibaba's Qwen model for customer service, Apple had Google build a custom model for Siri AI, and DoorDash CTO Andy Fang has described using multiple open-weight models for their business, including Moonshot's Kimi. The examples are piling up because the economics are becoming difficult to ignore.
Higher education now appears to be approaching the same inflection point. Universities have many of the same incentives that are driving industry toward self-hosted AI—lower long-term costs, greater control over sensitive institutional data, and opportunities for students to experiment with production-scale AI systems. Understanding why those incentives matter, and how they could reshape campus technology, requires a closer look.
The Cost of AI on Campus
Campuses are currently faced with two distinct pricing models for AI. First, there are enterprise site licenses—a flat per-user subscription that gives students and staff direct access to chat models and agents (like ChatGPT Edu or Microsoft Copilot). Second, there is token-based billing where institutions pay only for the exact amount of data processed to power custom, backend tasks. These could include automated 24/7 student-advising chatbots, AI-driven grading assistance for TAs, adaptive learning tools inside the LMS, and large-scale data processing for faculty research.
For unrestricted institutional deployment, however, neither model is likely to be financially sustainable. A Microsoft Copilot enterprise license, for example, ranges from $18 to $32 per user per month depending on education or business pricing. At the lowest published rate of $18 per user, a mid-sized university with 7,500 licensed users faces a bill of $135,000 a month—or $1.62 million a year. Even assuming a substantial negotiated discount that cuts those costs in half, the institution would still spend approximately $800,000 each year.
As for the token-based route, Uber recently made headlines when it burned through its entire annual AI budget in just four months due to surging token usage costs. While higher education operates on a vastly different scale, the takeaway is clear: uncapped token consumption is a financial wildcard that most universities cannot afford.
Compounding these inflated costs is a widespread lack of model discernment. Most universities currently over-provision their AI—using expensive frontier models for basic, low-stakes tasks. A campus chatbot designed to answer questions about library hours, parking permits, or drop-add deadlines doesn't require a trillion-parameter model like GPT-5. These routine queries can easily be handled by smaller, specialized, or open-weight models at a tiny fraction of the compute cost. Simply matching the model to the task—a practice known as right-sizing—can dramatically reduce AI spending without sacrificing performance.
Commoditizing AI on Your Campus
What would it actually cost for a university to own, rather than rent, its AI infrastructure? Let’s consider two plausible deployment strategies based on current enterprise hardware pricing and the computing requirements of today's leading open-weight models.
A teaching-focused institution of around 7,500 users could deploy a modest on-premises GPU cluster capable of supporting routine administrative workflows, campus chatbots, learning management system integrations, and most student experimentation for an initial investment on the order of a few hundred thousand dollars. Once the hardware is in place, ongoing operating costs fall dramatically, with future expenditures largely limited to electricity, maintenance, and periodic upgrades. The institution also retains flexibility: models can be replaced, tuned, or redeployed as the AI landscape changes.
Research-intensive universities may choose a different path. A substantially larger investment—on the order of several million dollars—would support a multi-node AI cluster capable of running today's frontier open-weight models while enabling advanced faculty research, graduate education, and large-scale experimentation. Although the upfront cost is considerably higher, it remains competitive with the cumulative expense of enterprise AI subscriptions over a typical three- to five-year hardware lifecycle.
The comparison is not perfect. Most commercial AI platforms bundle managed hosting, software updates, security, and vendor support into their subscription fees. Yet subscriptions remain an operating expense that must be renewed indefinitely. Hardware, by contrast, is a capital investment. It becomes part of the university's research infrastructure, much like a scientific instrument, a high-performance computing cluster, or a campus recording studio. When the depreciation cycle ends, the institution still owns an asset that can support teaching, attract external funding, and evolve alongside the next generation of open-weight models.
Seen in this light, the question is no longer whether universities can afford to build AI infrastructure. It is whether they can afford to keep renting it indefinitely.
Data Sovereignty
Cost, however, is only part of the equation. The way universities deploy AI also determines who ultimately controls institutional knowledge, research data, and the infrastructure that powers campus operations.
Many universities treat AI as a service to be rented rather than infrastructure to be owned. Every chatbot interaction, workflow, research query, and custom application is routed through a third-party provider. While convenient, this approach creates long-term dependence. As more institutional processes become intertwined with proprietary AI platforms, switching providers becomes increasingly difficult. If a vendor raises prices, changes its licensing model, retires a model, or alters its terms of service, the university has little leverage. Data sovereignty ultimately means more than protecting information—it means maintaining control over the technology stack that powers your institution.
This isn't a hypothetical concern. Recently, Anthropic changed its enterprise pricing structure, moving from a subscription-only model to one that combines a base fee with token-based usage charges. An institution that had built internal workflows around predictable subscription pricing could suddenly face substantially higher operating costs. When core academic processes depend on infrastructure owned by someone else, pricing changes become institutional risks rather than simple purchasing decisions.
Trustee Priorities and Practical Pedagogy
Many university boards include technology executives and business leaders who recognize that AI literacy is becoming a workforce expectation. They increasingly want graduates who can engage AI critically and creatively—not simply use it as a chatbot. Achieving that level of fluency means moving beyond prompt engineering into areas such as fine-tuning models, building retrieval-augmented generation (RAG) systems, and orchestrating agentic workflows. Educating students in this way requires a laboratory where students can experiment, fail, iterate, and build.
Experimenting solely through rented black-box models cannot provide the same educational depth as working directly with open-weight models on owned infrastructure. Operating on-premises hardware capable of running frontier open weights—like DeepSeek or Kimi—would allow students to look beneath the interface. They can inspect model behavior, fine-tune weights, benchmark performance, build evaluation pipelines, and deploy systems end-to-end. Moreover, open-weight models remove the financial friction that might discourage creative experimentation. Students are free to iterate hundreds of times without every failed idea generating another cloud bill.
Simply put, the strongest AI programs will combine robust on-premises compute with state-of-the-art open-weight models. Such an environment transforms AI from a subscription students consume into infrastructure they understand, modify, and improve. It enables richer coursework, attracts externally funded research, and prepares graduates to evaluate and create with AI tools. This should not be limited to computer science programs. Students in business, the humanities, the sciences, the arts, and professional programs will all benefit from opportunities to experiment with AI within their own disciplines.
Commoditization is Inevitable
AI is commoditizing. As this commoditization accelerates, capital will eventually pull back from the market, inflated valuations will normalize, and the infrastructure supporting today's AI boom will become dramatically more affordable. This pattern has played out repeatedly in technology markets. Fiber-optic networks, server infrastructure, and computing hardware all followed similar trajectories after previous technology bubbles. When that happens, the financial case for owning local hardware and leveraging open-weight models will become even more overwhelming.
We are already seeing early cracks in the assumption that bigger always means better. The prevailing narrative as recently as last year was that raw scale was everything—that whoever accumulated the most compute would build exponentially superior AI systems. That assumption is now being challenged. Companies that invested billions in AI infrastructure, like Meta and SpaceX, are now looking for ways to monetize excess compute capacity. At the same time, the open-weight ecosystem continues to close the gap with expensive proprietary models. The implication is significant: AI capability is becoming less concentrated and more accessible. Compute is evolving from a competitive moat into a commodity resource.
To prepare, higher education institutions need to stop blindly writing checks for recurring licenses and start asking fundamental questions: What specific tasks do we actually need AI to perform, and what are the right-sized models for those needs? Should we delay broad enterprise subscriptions and instead build a capital fund over the next couple of years to acquire high-performance hardware and run models locally? Do we want our university to be a consumer of AI technology or a participant in creating and leveraging it? By choosing ownership over endless leasing, universities can protect their budgets, retain their data sovereignty, and build permanent infrastructure that serves their students and faculty for years to come.
