0
There are 0 items in your cart: Cart Subtotal: $ 0.00
A Community College CIO Says Colleges Should Own Their AI Instead of Renting It, and the Cost Argument Deserves a Hard Look
The University Marketplace
0
An Arapahoe Community College executive argues colleges should own their AI, not license it. Here are the cost claims, the caveats, and the questions presidents and CIOs should ask first.
A Community College CIO Says Colleges Should Own Their AI Instead of Renting It, and the Cost Argument Deserves a Hard Look
An Arapahoe Community College executive argues that open-weight models running on hardware a college controls could cut costs and protect student data. It is advocacy from one early adopter, and the questions it raises are worth every president's time.
Key Takeaways
• In an October 13 Community College Daily commentary, Jeff Nesheim, vice president for strategy and CIO at Arapahoe Community College in Colorado, argues that community colleges should run open-weight AI models on hardware they control instead of licensing commercial products.
• His illustrative estimate for a 7,000-student college compares about $1.18 million a year in commercial software with about $168,000 in hardware and about $40,000 a year to operate, with five-year savings of about $5.4 million.
• The piece is advocacy from a practitioner whose college is piloting the approach. The cost figures are the author's estimates, and the claim that open models match commercial quality is asserted, not benchmarked.
• The underlying questions are real for every college: who controls student data, how new AI costs get funded, and whether institutions should build technical capacity or buy it.
The Argument in One Sentence
Community college leaders are being asked to adopt artificial intelligence quickly, often with budgets that were set before anyone knew what it would cost. In a commentary published October 13 by Community College Daily, Jeff Nesheim, vice president for strategy and chief information officer at Arapahoe Community College in Colorado, offers a bold answer: own it, do not rent it. His argument is that colleges should run open-weight AI models on hardware they control, instead of licensing commercial products or building on vendor interfaces.
Before going further, a note on what this is. The piece is labeled commentary, written by a practitioner who describes his own college's work. It is not independent reporting, and several of its numbers are estimates. That makes it more useful as a prompt for discussion than as a verdict, and the discussion is overdue.
Every AI decision a college makes is also a decision about who controls student data and who pays when usage grows.
Why the Question Is Arising Now
Nesheim points to several pressures. A 2025 EDUCAUSE study he cites found that only 2 percent of respondents were covering new AI costs with new funding sources, which means most colleges are paying for AI by cutting somewhere else. He notes that California State University signed a $17 million annual OpenAI contract in early 2025 covering 22 campuses, a reminder of how large the bills can get. And he reports that software-as-a-service costs rose an average of 34 percent in the first half of 2026.
The common thread is unpredictability. A college that licenses AI per seat or per use may find costs rising with every new student, every new tool, and every new vendor price change. A college with a fixed budget and variable usage has a problem.
What Open-Weight Means
Open-weight models are AI systems whose trained parameters are published, so an organization can download and run them on its own machines. Nesheim names models such as Qwen3, Llama 4, Gemma, and Mistral in the 32 billion to 70 billion parameter range, and argues they are adequate for advising, knowledge retrieval, and institutional research. He cites a Mac Studio with 96 GB of unified memory, which costs about $5,100, as the kind of hardware that can run them.
Running a model locally means student questions and records never leave campus systems, which he says supports data privacy. He also says it ties access to institutional capacity rather than annual budgets. Both are promising arguments, and both need verification.
The Cost Estimate
The article's most eye-catching claim is an illustrative comparison for a 7,000-student college. Commercial software, he estimates, could run roughly $1.18 million a year. A cluster of 30 Mac Studio units serving about 600 concurrent users would cost about $168,000 in capital and about $40,000 a year to operate. He estimates five-year savings of about $5.4 million and a payback period of about 51 days of avoided fees.
A savings estimate is only as strong as its assumptions, and the assumptions are where the debate should start.
The Questions a Careful Reader Should Ask
Several caveats matter. The cost comparison depends on the author's assumptions about how many people use the system at once, and the piece gives little detail on staffing, maintenance, security, or model updates. Someone has to run the hardware, patch it, monitor it, and replace it, and those costs can be significant. The claim that open models match commercial quality for these uses is asserted, not benchmarked. The author acknowledges that owned infrastructure needs capital, technical skill, and sustained commitment.
The data privacy argument also needs care. Nesheim's point about FERPA compliance is his interpretation and does not address the full legal picture. A chief information security officer will want to examine access controls, logging, and incident response before agreeing that local means safer. And the author leads a college that is piloting this approach, so the piece is advocacy informed by one institution's early experience.
One citation also deserves a second look: the Gartner figure about rising inference costs in agentic workflows is linked to a press release whose title refers to a different prediction, so the source for that specific number is unclear. That does not sink the argument, but it is a reminder to verify before repeating numbers in a board presentation.
What Arapahoe Is Actually Doing
The most grounded part of the piece is the description of Arapahoe's pilots. The college runs open-weight models on Mac Studio hardware it already uses. Its pilots include a tool for navigating internal policy documents, one for synthesizing institutional research data, and one supporting IT operations. The college describes a three-year plan toward AI access for every enrolled student.
These are modest, practical uses, and they are the kind where an internal tool can make sense: questions about the college's own documents, data that should not leave the institution, and tasks that staff repeat daily.
Mission and Equity
Nesheim frames AI access as a mission and equity question. Community colleges serve working, first-generation, and part-time students, many of whom do not have the money for premium tools. If access depends on whether a college can afford a vendor contract, students at the least funded institutions get less. If a college can provide access from its own infrastructure, the argument goes, it can offer every student the same tools.
That framing connects to a pattern we have seen in recent coverage of workforce demand, such as the waitlists at Utah technical colleges. Colleges are being asked to do more with limited capacity, and technology choices can either widen or narrow the gap.
A Sensible Way to Start
Nesheim's recommendations are straightforward. Start small: set up a local inference environment, load a model, and identify a problem to address. Invest in institutional technical capacity and accept that building takes longer than buying. Treat AI access as a mission and equity question, not only a procurement decision. And presidents and CIOs should test the objections to the approach rather than dismiss them.
A college that wants to follow that advice can run a low-risk test. Choose one internal use case, such as policy search for advisors. Run a commercial tool and a local model side by side for a semester. Track accuracy, staff time, user satisfaction, and total cost, including staffing. Report the results to the cabinet. The goal is a decision based on evidence, whichever way it points.
Who Decides
AI purchases at a community college rarely sit with one person. The president and the vice president for strategy set direction. The chief information officer and the chief information security officer evaluate technology and risk. The chief financial officer weighs costs. The vice president of academic affairs and the vice president of student services own the use cases in classrooms and advising. Institutional research directors often have the best view of data quality and the clearest sense of what would help.
For vendors, that means a single email to a single title is unlikely to move a decision. Understanding the whole committee, and speaking to each member's concern, matters more than ever.
A Middle Path: Buy Smart, Build Selectively
The choice is not only between renting everything and owning everything. Many colleges will land in the middle: licensing commercial tools for broad uses such as writing support and productivity, while running local models for sensitive internal work. A hybrid approach lets a college hold the line on student data in its most sensitive systems without taking on the full burden of running a platform. The decision for each use case is simple to state: what data does it touch, how many people use it, and what happens if the vendor changes its price or its terms?
Contract terms matter as much as technology. Colleges that do license commercial AI can negotiate limits on how student data is stored and used, caps on annual price increases, clear exit rights, and the ability to export their own prompts and records. These protections cost little at signing and a lot later.
The Staffing Question
The most underestimated cost in any build-versus-buy decision is people. Running local models requires someone who understands hardware, security, and model behavior, and community college IT departments are often stretched thin. A college that pursues this path should name an owner, budget for training or a new hire, and set aside time for monitoring. It should also decide what happens if that person leaves. Institutional knowledge that sits in one head is a risk.
Some colleges may find shared infrastructure more realistic: a regional consortium, a state system, or a partnership with a nearby university that already runs computing resources. Sharing spreads the staffing cost and gives smaller colleges access to expertise they could not hire alone.
What Students Will Notice
Whatever a college chooses, students will judge it by the experience. Does the advising assistant give accurate answers about financial aid deadlines? Does it work on a phone? Does it explain what it cannot do? Is there a person to turn to when it gets something wrong? Accuracy and trust matter more than where the servers sit, and any pilot should measure them directly through student feedback and error tracking.
Signals to Watch
Several developments are worth following: whether other colleges publish results from local AI pilots, whether independent benchmarks compare open and commercial models on advising and research tasks, how software prices move over the next year, and whether states or consortia offer shared AI infrastructure to community colleges. Each will show whether owning AI is a niche choice or a mainstream option.
College Leads maintains college email lists, a higher education contact list, and university mailing lists built around the technology and budget buying committee, including a community college president email list, a chief information officer email list, a vice president for strategy email list, a vice president of academic affairs email list, an institutional research director email list, a chief information security officer email list, a vice president of student services email list, and a chief financial officer email list. These are the higher ed decision makers who will shape technology choices this year. If you are ready to buy college email list data, or want a higher education marketing database by role and institution type, build your list here while colleges are planning budgets for the next year.
Frequently Asked Questions
Should community colleges build their own AI infrastructure?
One CIO argues yes. In an October 13 Community College Daily commentary, Arapahoe Community College's Jeff Nesheim says colleges should run open-weight models on hardware they control. The cost figures are his estimates, and the piece is advocacy from a college piloting the approach.
What are open-weight AI models?
Open-weight models are AI systems whose trained parameters are published, so an organization can run them on its own machines. Models named in the commentary include Qwen3, Llama 4, Gemma, and Mistral, in the 32 billion to 70 billion parameter range.
Who decides AI purchases at a community college?
Decisions usually involve the president, the chief information officer, the chief information security officer, the chief financial officer, and the vice presidents of academic affairs and student services, with institutional research directors advising on data and use cases.