Copyright forbes
CAMBRIDGE, MASSACHUSETTS - JULY 08: A view of Harvard Yard on the campus of Harvard University on July 08, 2020 in Cambridge, Massachusetts. Harvard and Massachusetts Institute of Technology have sued the Trump administration for its decision to strip international college students of their visas if all of their courses are held online. (Photo by Maddie Meyer/Getty Images) Getty Images Higher ed is racing to “do something with AI,” but the institutions pulling ahead aren’t shopping for apps—they’re making infrastructure decisions. The gap is visible in the data. With 90% of students using AI and 77% of educators feeling unprepared for AI—a clear signal that adoption is outpacing campus-level enablement and governance. And while AI governance is maturing, it’s still early; the International Association of Privacy Professionals’ 2025 report describes organizations "building the plane while flying it,” with formal AI oversight spreading but far from universal in education. Against that backdrop, two universities offer a useful contrast in how infrastructure drives strategy. Two Strategic Approaches to AI Infrastructure Old Dominion University has taken a cloud-first path, standing up a centralized AI hub on Google Cloud to accelerate research and student-facing tools, announced in an October 29 press release. In an interview and press materials shared ahead of today's announcement, ODU described MonarchSphere—an AI incubator built with Google Public Sector—to connect researchers, instructional designers, and operations teams to a common stack for compute, models, and data. The approach builds on ODU's existing MonarchMind platform (a secure, self-hosted interface that routes multiple models, including Gemini, ChatGPT, and Llama, through an enterprise layer with campus controls). "We weren't looking to buy something off the shelf," President Brian O. Hemphill told me. "We wanted a partner willing to innovate… and build something that has not been done before… We're not waiting for the future—we're building it." On the industry side, Google Public Sector's Matthew Schneider emphasized that what began as a research conversation "quickly moved… to the whole of campus," tying workforce development and community use cases (Hampton Roads coastal resilience) into the same backbone that powers course assistants and research workflows. The timing also aligns with Google's broader higher-ed push. This summer the company announced a $1 billion multi-year initiative for AI training, cloud credits and tools in U.S. universities, alongside a new AI for Education Accelerator that offers no-cost AI training, Google Career Certificates and protected access to Gemini/NotebookLM. Rensselaer Polytechnic Institute demonstrates a complementary route: Put high-performance compute on campus and give students and faculty direct, low-latency access. "We have a quantum computer and a supercomputer on campus. Everybody has access to that… The compute is not an issue," said Liad Wagman, dean of the Lally School of Management. RPI's AiMOS system—an IBM Power9/NVIDIA GPU supercomputer that debuted as the most powerful at any private university—anchors research from AI and life sciences to environmental modeling, while the Jefferson Project at Lake George applies sensing + analytics to freshwater quality and harmful algal blooms, with lessons transferable beyond the region. These are two different answers to the same question: Where will AI actually run—and under what controls? While this article highlights two well-resourced institutions, these approaches can be scaled appropriately for smaller colleges and universities through regional partnerships, shared services, and strategic investments that match institutional priorities and budget constraints. What follows are seven infrastructure choices that, in practice, separate pilots from full systems change. For longer-term investment, look beyond your balance sheet to regional coalitions and federal funding streams like the NSF Regional Innovation Engines, which are designed to build shared, cross-institution innovation capacity in priority tech areas. The through-line is simple: Rent elasticity, pool what you can, and spend leadership attention on governance and faculty development—because capability, not hardware, is what compounds. Seven Infrastructure Decisions: A Roadmap, Not a Checklist Rather than a checklist, read these as directional guardrails you can adapt as higher education leaders. 1. Decide your workload placement Not every campus can underwrite a seven-figure GPU build—and they don't have to. The pragmatic play is access over ownership: Tap community contracts and shared services (e.g., Internet2's NET+ programs that give higher ed negotiated access to major clouds and tooling), then add cloud "bursting" for research spikes while you stand up a lightweight enterprise access layer and faculty skilling. If your research needs burstable GPU capacity, cloud compresses queues and time-to-results; it also makes it easier to wire assistants into the LMS at scale, as ODU is doing alongside its AI hub. If you need immediacy (minutes, not days) for iterative labs or sensitive datasets, an on-prem enclave like AiMOS can be decisive. Codify a short workload placement policy that states what runs where and why. Cost considerations: Cloud-based solutions typically require lower initial investment but ongoing subscription costs, while on-premises solutions demand substantial upfront capital but may offer better economics for compute-intensive programs over a 5+ year horizon. 2. Stand up an enterprise AI layer If dollars are tight, treat the "platform" as configuration, not code. Extend the SSO/MFA you already own (Shibboleth/AD/Okta) to gate model access; use LMS LTI and existing data connectors before buying new middleware; turn on basic audit logs in your cloud tenants; and publish a two-page policy mapped to NIST AI RMF (ISO/IEC 42001 can wait). For model access, start model-agnostic. Route to one inexpensive commercial model plus one open-weight model hosted via a small gateway (e.g., a single GPU server or short-lived cloud instance). You’ll get role-based access, logs, and a governed entry point without standing up a full platform team. 3. Put GPUs in students' hands—safely ODU's approach pairs protected, managed notebooks and training with Career Certificates; RPI pairs campus compute with authentic datasets and labs. Whichever route you choose, the pattern is the same: make advanced compute a course primitive, not a special request, and wrap it in policies that keep costs and data use in bounds. If budgets are tight, set aside a modest pool of preemptible/spot GPUs for capstones and research sprints; enforce idle auto-shutdown and per-course spend caps; and prioritize smaller, well-tuned models (7B–13B) for most assignments while reserving a frontier model for a few showcase tasks. Pair usage with a short "responsible use" addendum in syllabi and a one-hour micro-clinic for faculty TAs. 4. Fund for velocity, not just capacity A cloud-first incubator like ODU's newly announced MonarchSphere centralizes pilots so the best ones can scale; use that model as the pattern, not a line-item claim about their budgets. In practice, many campuses keep velocity high by allocating a modest pool of cloud credits to a few high-value use cases (for example, LMS-embedded assistants, transcript-evaluation workflows, or research pipelines), enforcing auto-shutdown by policy, and publishing a one-page "how to buy compute" guide so faculty don't reinvent procurement. When a pilot proves uptake and impact, extend credits; when it doesn’t, reclaim and redeploy. Keep on-prem only where low-latency or specialized hardware clearly beats cloud economics. Even modest allocations of cloud credits can enable meaningful pilots while controlling costs. Successful pilots can later be scaled with dedicated funding. 5. Wire in one public-impact domain ODU’s leadership cited an AI-powered flood prediction and emergency response platform as a top research priority—aligned to a real regional need in Hampton Roads. RPI's Jefferson Project shows how a long-running environmental platform can become a magnet for student projects, grants, and cross-disciplinary work. Pick a domain that matters locally; make it a proving ground for your stack. 6. Treat skills and credentials as part of infrastructure Tooling without skilling stalls out. Even if your institution does not have the budget for large scale infrastructure, it is important to invest in upskilling faculty, students, and staff as part of AI strategy. The common thread: Practice on the platforms employers use, with data protections, ethical training, and educational resources in place. Vendor programs can further defray cost and accelerate fluency: Google’s AI for Education Accelerator pairs no-cost training and Career Certificates with education-protected tools so students and instructors can practice on real platforms without heavy CapEx. 7. Nail authorship, IP, and agent policy early Hemphill stressed that ODU’s efforts sit inside existing IP and data-security policies and respect shared governance. Schneider underscored that the Google stack is open by design to integrate with non-Google models and campus systems. Translate those principles into short, comprehensible rules for classrooms and admin agents (what's permitted, what's logged, what needs a human in the loop), mapped to NIST/ISO so the program survives leadership changes. The Core Infrastructure Question What ties these choices together is clarity about where intelligence lives, who can reach it, and how it's governed. In that sense, ODU's and RPI's paths are less different than they look. One leans on elastic cloud and a campus incubator to turn pilots into products. The other lowers the friction to experiment by putting serious compute undergraduates can actually touch. For institutions with more modest resources, a hybrid approach often makes sense: leverage cloud services for flexibility, pursue regional partnerships to share costs, and make strategic on-premises investments aligned with institutional strengths and priorities. Universities keep asking which AI tool to buy. The better question is how quickly your people can get secure, governed access to the right compute and models—and whether your infrastructure choices increase educational and mission-driven opportunities next semester. The institutions that answer this infrastructure question effectively will be the ones that truly transform education with AI, regardless of their size or resources. Editorial StandardsReprints & Permissions
