INSIGHTUniversities & research · UC-02
On most campuses the binding constraint on AI research is not the GPU budget — it is the machine room. A modular pod beats a retrofit when a group has funded hardware, a campus electrical service with spare capacity, and no room that can take a liquid-cooled rack. A retrofit wins when the existing hall already has the power, the floor, and the staff. Here are the criteria that separate the two, and the cases where neither is the answer.
PUBLISHED LAST VERIFIED BY JOSEF ELIMELECHREVIEWED PODOS AI ENGINEERING
When a pod beats a retrofit
An award bought accelerators. The feeders, the chilled-water branch, and the structural work sit on a different budget line, on a different calendar.
Spare capacity on the campus loop for the continuous IT load plus cooling and losses. Without it, neither option proceeds — the load study comes first.
If the machine room already has the power, the floor loading, the facility water, and the staff, adding racks there is the lower-risk path.
The funding shape
University compute does not arrive on an annual refresh curve. It arrives in discrete awards, and the shape of those awards decides what can be bought. NSF's Major Research Instrumentation program is the clearest example: up to $4 million for the acquisition or development of a multi-user research instrument, an acquisition award of up to three years, and at least 70% of total project cost on the equipment line of the budget.[1]
That budget rule explains most of the friction. Instrumentation money is built to buy instruments. A machine-room renovation — feeders, a chilled-water branch, structural work, abatement — is a construction cost, funded through a separate capital process on the university's calendar rather than the award's. Departments routinely end up holding hardware they cannot energize because the facility half never got its own line. Whether a self-contained unit qualifies as equipment is a question for sponsored programs, not for a vendor, but the contrast between buying an instrument and renovating a building is not ambiguous. The award period compounds it: three years sounds generous until eighteen months go to design and capital review, and the depreciation clock on accelerators starts at purchase, not at plug-in.
Physical envelope
Three physical limits decide the answer, usually discovered in this order.
Most academic machine rooms were laid out for single-digit kilowatt racks on raised floor. Uptime Institute's 2025 survey of more than 800 operators shows fleet-wide densities climbing into the 10–30 kW band[2], and rack-scale AI systems sit far above that: NVIDIA ships the GB200 NVL72 as 72 GPUs and 36 CPUs in one liquid-cooled rack, with no air-cooled equivalent on offer.[3]
Campus distribution is a shared resource with its own queue: a load study, likely new feeders and switchgear, sometimes a transformer with a lead time measured in quarters. The competition is not local either — LBNL's congressionally mandated assessment put US data-center electricity at about 4.4% of national demand in 2023 and projected 6.7–12% by 2028.[4]
Once the racks are liquid-cooled, the building needs a facility water loop with a defined supply temperature — ASHRAE names liquid-cooling facility water classes alongside the air classes[5] — plus heat rejection and pipe routing to the room. In an occupied building, the routing is often harder than the plant.
A room that cannot cool one such rack cannot be tuned into one that can.
Constraint by constraint
The comparison a research-computing director and a facilities engineer tend to run together. Neither column wins every row.
| # | Constraint | Machine-room retrofit | Modular pod on campus land |
|---|---|---|---|
| UR-01 | How the money arrives | A construction cost on the capital-planning calendar, funded separately from the award that bought the hardware. | Procured as equipment — though whether it qualifies on an equipment budget line is an institutional question. |
| UR-02 | Schedule against the award period | Design, bid, abatement, and construction inside an occupied building, often with the hall offline. | Fabrication runs off-site in parallel with site works; ICC/MBI 1205 splits in-plant and on-site inspection with the local jurisdiction.[8] |
| UR-03 | Rack density ceiling | Bounded by the existing airflow design — most academic machine rooms assume single-digit kilowatt racks. | A property of the enclosure, not the building; specified around liquid-cooled racks from the start. |
| UR-04 | Cooling loop | New facility water, pipe routing through occupied floors, and a heat-rejection plant the building may lack. | Closed loop and heat-rejection interfaces integrated and tested before shipment; ASHRAE facility water classes still govern the interface.[5] |
| UR-05 | Electrical service | New feeders, switchgear, often a transformer upgrade — the long-lead item in most retrofits. | Still needs campus capacity; distribution inside the boundary is factory-built. |
| UR-06 | Physical access | Freight elevators, door widths, and floor loading in a building designed for none of it. | Constrained by the road: federal limits fix National Network width at 102 in and interstate gross weight at 80,000 lb.[9] |
| UR-07 | Waste heat | Usually rejected to atmosphere; capturing it means touching the building hydronic system. | A warm-water loop returns fluid at a temperature a campus heating network can accept.[7] |
| UR-08 | Who operates it | Absorbed by existing facilities and research-computing staff, who inherit the new liquid loop. | Unchanged. A pod does not create an operations team; coolant chemistry is new either way. |
The campus advantage
Universities are one of the few settings where reject heat has a customer next door. Campuses run their own heating networks, and a warm-water direct-to-chip loop returns fluid at a temperature those networks can use. NREL's Energy Systems Integration Facility is the reference implementation: warm-water cooling enables chiller-free operation and energy recovery[6], and captured HPC waste heat reheats the building's office and lab space, at a reported facility PUE of roughly 1.04.[7] That turns an operating cost into a supply. The mechanics are in our heat recovery and direct-to-chip liquid cooling pages.
HONEST LIMITS
Several common campus situations are better served by something else.
In the product
PODOS integrates power, cooling, racks, and networking into a factory-built unit rather than a field-built room. Each PODOS Pod is designed as a standardized 1 MW building block and designed for 128 GPUs — one funded increment, one deliverable, one commissioning event. Because subsystems are integrated and tested before shipment, PODOS targets a 90-day window from order to commissioning for a standard unit. These are design targets, not measured campus results.
The sequence for a research-computing office: confirm campus electrical headroom, confirm siting and the delivery route, then decide whether the capacity joins the existing queue or stands beside it. Our readiness checklist orders the site questions, the power architecture covers the other side of the service connection, and deployment covers site works. If the campus question is construction versus manufacturing, start with modular versus traditional data centers. Other verticals are profiled on the use-case hub; terms are defined in the AI infrastructure glossary.
PODOS AI is an early-stage company. Nothing above describes a completed campus deployment, a customer, or a certified product.
QUESTIONS
That depends on the award and on how your institution classifies the unit — sponsored programs decide it, not the vendor. The structural contrast is clear, though: NSF's Major Research Instrumentation solicitation requires at least 70% of an acquisition's total project cost to sit on the equipment budget line, while a renovation is a construction cost.
Enough spare capacity on the campus loop to carry the continuous IT load plus cooling and losses, with the transformer, protection, and metering to match. A campus electrical study comes before any hardware decision, because the answer sets whether a retrofit is even possible.
Not for bursty work. Shared allocations exist so a group needing a few large runs a year does not have to own, power, and staff a cluster. A dedicated unit makes sense when demand is sustained, when data cannot leave the institution, or when queue waits are the bottleneck.
No. It moves most of the inspection burden off-site rather than removing it. ICC/MBI 1205 sets out how in-plant and third-party inspection relate to the local authority having jurisdiction, which still approves siting, foundations, and utility connections.
Bring the campus electrical study, the siting constraint, and the award period. The configurator walks the same variables a facilities review would.