Prime Performance Group · Engineering

Cloud vs. Hybrid Compute — Z University

A back-of-envelope model for whether the media + always-on-AI workload justifies owning GPUs on campus. Every input is editable — change any number and the verdict updates. All unit costs are mid-2026 estimates; verify at engagement.

Assumptions

Defaults reflect a plausible "AI-native + media program" load. Edit freely.

AI text usage (ops + students)
#
#
#
tok
d
B
$/M
Media / video generation
#
min
wk
$/min
Owned GPU node (hybrid)
$
yr
$
%
Verdict
—

—

Pure cloud — annual
—
—
Hybrid — annual
—
—

Where the money goes

Pure cloud—
Hybrid (owned node + cloud burst)—
LLM tokens (cloud) Media on cloud Media cloud burst (hybrid) Owned GPU node
Annual savings (hybrid)
—
Pure cloud minus hybrid
Node payback
—
On avoided media-cloud spend
—
How to read this. The point isn't the exact dollar figure — it's the shape. Text/LLM cost is real but controllable with routing, caching, and per-user budgets. Media generation on per-second cloud APIs is the runaway, and it's exactly the workload an owned GPU node absorbs at the cost of electricity. So the school stays cloud-first for frontier + burst, and moves the predictable media/bulk base load on-prem once the telemetry proves the utilization — the same power + cooling + GPU footprint the campus is already weighing for the compute-hosting cost-offset. Cold-climate free cooling makes owning cheaper here than almost anywhere.

Unit-cost anchors (mid-2026 estimates): LLM tokens ~$0.10–0.50/M (small) to ~$15–30/M (frontier) — blended default $1.50/M assumes a cheap-model-first router with prompt caching · AI video ~$0.07–0.40/sec (~$4–24/min) · H100 ~$2/hr rent, ~$300k for an 8×H100 node, ~$1–2k/mo power, ~7–8 mo payback at sustained use. Sources: BenchLM, FluxNote, IntuitionLabs.