One number in, a whole plan out: how many computers the money buys, which colleges get them, how many people end up covered, and what it costs to keep running. Every figure is an estimate for planning — the four-computer validation replaces them with measurements.
How to use this page. Drag the budget, or press one of the four marked stops, and everything below re-solves for that amount. Every control has an (i) beside it that says what the control changes and where its number came from — start there if a word is unfamiliar.
UVU Local AI — budget scenario · printed from the Utlyze simulator · evidence 2026-09-02
Exam week is a stress test about nine times an ordinary day. The plan does not buy machines for it — it queues, and sends overflow to the free tools.
Education is the price UVU would really pay. fact
Previous is the lineup Apple replaced on 25 August, bought refurbished: cheaper, sold one unit at a time, about a tenth slower. unknown
Paid catalysts in each college, awards for real classroom work, drop-in clinics. Below about a tenth of the budget, machines go unused.
An account with a spending limit, in the cloud — not a machine, and not the 512 GB node below.
One Mac Studio with 512 GB of memory. It can run one larger local model — GLM-5.3 — for about two research conversations at a time. It is an instrument for research, not a service for the campus.
One accountable professional plus paid students at $18 an hour — cheaper once the fleet is large. est
Pilot scale
College
People
Needs
Minis
Studios
Carries
Share of the busiest hour it covers
How to read this table. “Needs” and “Carries” are both conversations happening at the same moment in the busiest hour of the day — what the college is modelled to need, and what its machines can actually hold. Over 100% is headroom for growth and exam weeks, not waste. A college with no machines is not cut off: it uses the free tools UVU already licenses (Copilot Chat, Gemini for Education) until the budget reaches it.
Where all of this comes from. How much demand each college has: its Fall 2025 headcount, multiplied by a rate measured at a university that already runs a service like this — 4,488,101 questions asked by 5,721 eligible people in a year, converted into how many conversations are open at the same moment in the busiest hour. That comes to 1.44 conversations per 1,000 people; the measured campus itself came to 1.43, and Berkeley and Michigan agree within the same range. Slower, more thoughtful answers take longer and roughly double the need, which is what the exam-week setting covers. Different subjects use AI at different rates, from a survey of 95,000 students. What the machines cost: Apple's education prices, verified 2 September 2026. What they can carry: measurements on the same chips plus our own test run, which passed its isolation checks. What staff cost: the loaded Provo wage published by the US Bureau of Labor Statistics. How the money is shared out: the college with the highest demand times its strategic weight draws first from the shared pool; Smith Engineering and Science get the heavyweight machines; about one machine in seven is held as a spare; one rack kit per six machines. When to buy more: the plan says add six minis whenever weekly load sits at 70% of capacity or higher for two weeks running. Three years out: expect about 1.9 times the first year's demand (somewhere between 1.3 and 2.4 times) as more people adopt it — the number of eligible people barely grows, about 0.6–1.3% a year. Full method and every source: the guide, §13.
Prepared independently by Utlyze. Not a publication of, endorsed by, or affiliated with Utah Valley University.