From the Campus Operations edition of October 14, 2026
Arapahoe Community College's CIO argues for owned open-weight AI infrastructure while piloting policy, institutional-research and IT-operations agents on campus hardware
Arapahoe Community College (Colorado); commentary published by Community College Daily · Community college IT infrastructure and administrative AI · Colorado, United States
- Publisher
- Community College Daily
- Original publication
- Community College Daily, 13 October 2026
- Source retrieved
- 2026-10-15
What happened
Jeff Nesheim, vice president for strategy and CIO at Arapahoe Community College, writes that ACC's IT team runs open-weight language models with open-source inference software on Mac Studio hardware already in use for other purposes. The first institutional agents being piloted help staff navigate internal policy documents, synthesize institutional research data and support IT operations. They run entirely on college-owned hardware against data that stays on the college network. He gives a cost model: commercial SaaS AI access for a 7,000-student college costs about $1.18 million a year. A 30-unit Mac Studio cluster serving about 600 concurrent users would cost $168,000 in capital and about $40,000 a year to operate. On those figures he estimates $5.4 million in five-year savings and a payback of roughly 51 days of avoided fees. He also cites a 34% average SaaS cost rise in the first half of 2026 and a Gartner projection that agentic inference costs per workflow will rise five- to 30-fold through 2028. He acknowledges that owned infrastructure needs capital, technical capacity and institutional commitment. The college says it is just starting a three-year plan.
Why it matters
This is a build-versus-buy argument for administrative and IT AI from a resource-constrained public institution. It is relevant to campus CIOs and CFOs weighing per-seat or token-based AI licences against institution-run inference, especially for agentic workflows where usage-based costs can grow. Relevance to large research universities and to regulated data (HR, health) is limited until security and support costs are shown.
Evidence and measured results
The author's description of his own college's early pilots, plus his own cost model. The source gives no pilot usage, accuracy, staff time or user-satisfaction data. Cost inputs (SaaS price basis, staffing, power, replacement cycle) are not itemized. Third-party figures (the SaaS increase, the Gartner projection, the 2025 EDUCAUSE AI Landscape finding that only 2% fund AI from new sources) are cited but were not independently checked in this run.
Limitations and uncertainty
Opinion piece by an interested practitioner. The savings are projections, not results. The 600-concurrent-user capacity and the $40,000 operating figure are not explained, and it is unclear whether staff time is included. No pilot outcomes are reported. Comparisons with commercial frontier models are asserted, not tested. Commercial pricing varies widely by contract.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-10-14; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Colleges with flat budgets face AI licence costs that scale per user or per token, with agentic workflows expected to multiply consumption.
- Stakeholders
- CIO / VP for strategy
- CFO or budget office
- Institutional research director
- President's cabinet
- Information security officer
- Discovery questions
- What does current or quoted AI licensing cost per year, and how is it priced?
- Which internal use cases (policy lookup, IR synthesis, IT support) would run first?
- Who would operate and support local inference after the pilot?
- How will the institution compare local model quality with current commercial tools?
- Value hypothesis
- Interpretation: for high-volume internal use cases, owned inference may cap marginal cost and keep data on campus, if quality and support costs are acceptable.
- Potential engagement
- A bounded total-cost and capacity assessment comparing a small local inference pilot with current or quoted SaaS pricing for two or three internal use cases.
- Unsupported claims
- Guaranteed $5.4 million five-year savings or 51-day payback
- Parity with frontier commercial models on all campus tasks
- That on-premises hosting alone satisfies FERPA or security obligations
Pre-sales engineering
Role takeaway
- Fit
- Fits internal retrieval and summarization workloads (policy documents, IR data, IT runbooks) that do not need frontier-model capability.
- Architecture and integration
- Open-source inference runtime on owned Apple-silicon hardware, retrieval over internal document stores and IR data, fronted by campus identity.
- Prerequisites
- Hardware capacity and placement
- Inventory and classification of source documents and IR datasets
- SSO integration
- Model licence review
- Constraints
- Fixed concurrency ceiling versus elastic cloud
- Staff skills to operate inference
- State data-handling rules and FERPA obligations
- Security
- Not described in the source; confirm host hardening, network segmentation, audit logging, access control on IR data and prompt-injection defences for agents.
- Proof of value
- Load test with realistic concurrent prompts
- Side-by-side accuracy test against current tool on sampled policy and IR questions
- Measured monthly run cost including staff time
Delivery
Role takeaway
- Work
- Productionize the pilot agents with named owners, service levels, monitoring, and a documented cost and quality evaluation.
- Dependencies
- IT infrastructure team
- Institutional research office
- Policy owners for source documents
- Security office
- Ownership
- IT (CIO office) owns the platform; business units own each agent's content and accuracy.
- Skills and adoption
- IT staff need model-serving and monitoring skills; end users need guidance on checking outputs against source policy.
- Governance checkpoints
- Agent intake and risk review before access to IR data
- Model licence approval
- Quarterly accuracy and cost review
- Acceptance criteria
- Proposed: answer accuracy on a sampled policy question set meets an agreed threshold
- Proposed: response latency within agreed limits at peak load
- Proposed: documented fully loaded annual cost compared with SaaS quote
- Risks
- Hidden staffing and maintenance cost
- Model quality shortfall driving users back to unsanctioned tools
- Single-site hardware failure
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
An on-premises inference tier using 32–70B open-weight models (the author names Qwen3, Llama 4, Gemma and Mistral) on Apple-silicon workstations, with retrieval over internal policy and institutional research data. Institutions would need to plan for concurrency limits, model and runtime updates, redundancy, monitoring, and integration with identity and data warehouses. Usage-based costs are swapped for fixed capacity limits.
Governance
Who approves, reviews and stays accountable for outcomes?
Local hosting does not remove the need for an agent approval process, model licence review, data classification for IR and HR data, output validation, and a named owner for each agent. The author frames the choice as a mission decision for presidents and CIOs. That implies decisions at board or cabinet level rather than within IT alone.
Security and privacy
What data, permissions and controls need testing?
Keeping data on the campus network reduces third-party processing and model-training exposure, which the author ties to FERPA liability. It also moves all security work to the college: host hardening, access control to institutional research data, logging, patching of inference software, and prompt-injection controls for agents. The source does not describe any of these controls.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
The author argues that running inference builds staff expertise, but the operating load falls on a small IT team (his illustrative college has an IT team of 15 people). No workforce impact or accessibility testing is reported.
Procurement
What should contracts, pricing and exit terms secure?
Changes AI procurement from licences to hardware capital plus open-model licence review. It reduces vendor lock-in and contract-renewal risk, but leaves the college with no vendor support and responsible for keeping the stack updated.
Operating model
Which teams own the service once it runs?
IT becomes the AI service operator rather than a licence manager. Business units still own content and accuracy, and a multi-year plan is needed to move from internal agents to learner-facing tools.
Publication history
- 2026-10-14Campus Operations · Issue 374 resources
Stable resource ID: arapahoe-cc-owned-open-weight-ai-infrastructure-20261013