AI on-prem is becoming
a building project.
The board approved an AI initiative. The GPU servers are about to ship. Or your manufacturer rep — APC, Vertiv, or another partner — pointed you here for an assessment. Whatever brought you, the conversation is the same: the building has to change before the workload arrives, and most existing infrastructure can't carry it. This is the assessment that gets you decision-grade data before the hardware lands on the dock.
The conversation we keep having.
We get a version of the same call almost every week now. Sometimes it's the IT director who just got handed a board mandate to "have an AI plan." Sometimes it's the facilities lead whose CIO came back from a conference saying GPU servers are coming. Sometimes it's a manufacturer account executive who needs the building's power and cooling validated before quoting a high-density rack.
What's consistent across all of them is the gap. The decision to deploy AI on-prem is almost always made upstairs — in the C-suite, in the board room, by clinical leadership at hospitals, by district leadership at school systems. The decision about whether the building can actually run the AI workload is a separate conversation that almost never happens at the same table.
Three scenarios we've seen recently:
"We ordered four DGX servers for a research initiative. They're showing up in three weeks. Can you tell us what we need to do?"
"Our clinical team wants to run inference on radiology images. The vendor sent a power spec. It's not what we have."
"Our manufacturer wants to quote us a high-density rack. They asked if we'd had a readiness assessment done. We hadn't."
The right time to have the building conversation is before any of those scenarios — ideally when AI is still on a slide deck, before purchase orders are cut. That's what an AI Readiness Assessment is for.
What AI infrastructure actually demands.
A modern AI server is not a rack-and-stack file server. It's a dense, power-hungry specialized appliance designed to run GPUs at maximum sustained draw for hours at a time. The numbers have moved fast, and the gap between what AI demands and what existing infrastructure was designed to carry isn't incremental — it's a different category of build. A single AI chassis can draw 5–15× what a traditional enterprise rack was sized for. The highest-density NVIDIA reference platforms reach 30× or more.
These aren't theoretical numbers — they're what NVIDIA publishes in their installation guides, and what manufacturer cooling sizing references for AI deployments call out. A single DGX H200 at full draw is roughly 5× the load a typical enterprise rack carries today. A single 125 A single-phase 208V circuit, after NEC's 80% continuous-load derate, will deliver one DGX. Add a second DGX and you're past the line.
Our deep-dive article on AI on-prem electrical math walks through the full calculation — single-phase versus three-phase capacity, the NEC 210.19(A)(1) derate, why 208Y/120V is the typical answer for mid-enterprise deployments and 480Y/277V is the path for hyperscale. The short version: AI on-prem at any meaningful density requires a three-phase upgrade plus new feeder, new subpanel, and new permits. The infrastructure conversation isn't about adding capacity — it's about building a new electrical and cooling regime that didn't need to exist before AI showed up.
Where most buildings break.
The electrical service can't carry it.
Single-phase 208V circuits that comfortably ran traditional servers can't deliver AI density. The fix is real electrical work: new feeder from the main panel, new three-phase subpanel, new permits, new outage windows. Not a rack PDU swap — a building project.
The cooling can't move the heat.
Spaces designed for traditional compute are being asked to dissipate 5–10× the heat they were sized for. The cooling math doesn't bend that far — at AI densities, traditional CRAC or perimeter cooling can't keep up. Above ~50 kW per rack, air cooling stops working entirely; direct-to-chip or rear-door heat exchangers become necessary, not optional. That's a different cooling project, and it can't be retrofitted in a weekend.
The UPS can't hold the load.
VRLA UPS systems sized for the previous generation go into overload the moment the new switch and GPU servers come online. The right answer is a properly-sized lithium-ion UPS — but the UPS conversation has its own electrical, fire-code, and runtime considerations that need to be planned, not improvised.
The KAT-5 + Genoa approach.
AI Readiness is delivered through our partnership with our sister company Genoa PCM, which holds the methodology platform behind the assessment. Genoa builds the engineering rigor — the field capture system, the rules engine, the analysis methodology. KAT-5 is the value-added reseller delivery channel: we walk the building, run the assessment, and turn the findings into the actual quote and installation work.
The product is AIRAS™ — the AI Readiness Assessment Service — which is Genoa's PIIA™ (Physical Infrastructure Impact Assessment) methodology calibrated specifically for the questions that matter when AI workloads are on the way. Power density headroom. Three-phase readiness. Cooling thresholds. Fire code implications of lithium UPS at the scale AI demands. Rack-level density planning.
What you get back is what we call decision-grade data — structured, field-verified, vendor-neutral findings that you own. Not a sales pitch dressed up as a report. Not a vendor's spec sheet with your logo on it. Real per-rack data, real environmental measurements, real lifecycle analysis on the equipment that's already in place, and a prioritized roadmap of what has to change for the AI deployment to land cleanly.
Walk the building.
Field-tech assessment of every IDF, MDF, and data center space involved in the AI deployment path. Per-rack capture: nameplates, environmental conditions, power consumption, cabling, cooling, lifecycle status.
Run the math.
Genoa's rules engine analyzes the captured data against the AI workload requirements: power load math, three-phase capacity, cooling thresholds, UPS sizing, fire code implications. Outputs the gap.
Quote the gap.
The findings translate into a prioritized scope of work — what has to happen first, what can be sequenced, what's optional. Vendor-neutral, but tied to specific manufacturer products where it matters.
Deliver the work.
KAT-5 carries the install, with Genoa's OWL™ Lifecycle providing the project management overlay. Coordinated with electrical contractors, cooling installers, and your internal team. Documented end-to-end.
The output is yours regardless. If you take the findings to a different reseller for the implementation, that's fine — the data is structured, exportable, and not locked behind a sales process. The assessment is the product. The hardware quote is the next conversation.
What we work with.
AI infrastructure isn't a single product — it's a system. Power, cooling, racks, networking, and the GPUs themselves all have to fit together. Below are the manufacturer relationships that anchor most of our AI deployment work, with the products we're quoting most often right now.
APC by Schneider Electric & Vertiv
Both APC and Vertiv bring deep, well-developed product portfolios to AI infrastructure. Their lines cover what AI-class deployments require: in-row and rear-door cooling, direct-to-chip liquid cooling for the highest densities, three-phase UPS systems, lithium-ion battery technology, and the monitoring platforms to manage all of it at scale.
We work with both manufacturers across our customer base. Customers often arrive with a brand preference based on existing infrastructure standards, channel relationships, or operations team familiarity. We respect those preferences and deliver against them. Where customers are open, we help match the right product set to the specific deployment — density, footprint, runtime requirements, and existing rack standards all factor in.
Common deployments: APC Smart-UPS Lithium-Ion · APC Galaxy VS / VL · APC NetShelter racks · Vertiv Liebert DSE / PEX4 · Vertiv direct-to-chip liquid cooling · Vertiv high-density rack solutions · in-row and rear-door cooling · iCOM and EcoStruxure controls
NVIDIA
The H100, H200, and B200 platforms set the power, cooling, and electrical specs that drive every other decision in an AI deployment. We work to NVIDIA's published installation guidance and integrate with the broader NVIDIA ecosystem partners on actual hardware delivery.
Reference platforms: DGX H100 · H200 · B200 · GB200 NVL72
Multi-vendor
AI deployments demand correctly-sized cabling (Cat 6A minimum for any 802.3bt PoE we're feeding), high-density fiber for inter-rack runs, and switching that can deliver the bandwidth GPU servers actually need. Our switching work is most often Cisco Catalyst, Juniper EX, or HPE Aruba CX, depending on the customer's existing standard.
Common deployments: Cisco Catalyst 9300/9400 · Juniper EX4400/EX4650 · HPE Aruba CX 6300/6400 · CommScope & Panduit cabling
Xtreme Power Conversion
XPC Connect (XPCC) is our specialty lithium UPS line for situations where the major-vendor catalogs don't fit cleanly — closets running hot, shallow racks, TAA-compliant procurement for government and education, or three-phase deployments where the modular form factor wins. LiFePO₄ chemistry across the full portfolio, 50°C operating range, 15-year battery life, and a 6-year warranty.
Their M90C modular three-phase platform (5–60 kVA) is particularly relevant for AI-adjacent deployments where scalability and hot-swap serviceability matter. For shallow-depth closets and edge IT, the J60C and J90 series fit places where standard rack UPS won't.
Products we deploy: J60C / J90 1U lithium · P91 Li 2U rackmount · M90C modular three-phase · Li90 three-phase · SPDU intelligent PDUs · TAA-compliant configurations
Get the building conversation right — before the GPU servers ship.
If AI is anywhere on your roadmap — committed, planned, or still being scoped — that's the moment to run the assessment. Decision-grade data, vendor-neutral findings, and a prioritized roadmap that you own.