Tenant power density is changing Your visibility hasn't
GPU tenants are pushing rack power well beyond the design spec your facility was commissioned against. Verdigris gives you per-tenant-zone visibility into actual electrical capacity. So you can sign tenant SLAs grounded in measured headroom, not assumptions.
Goldman Sachs projects $250–650 billion in AI infrastructure capital expenditure by 2026. Total data center capex is forecast to reach $3 trillion by 2030, growing at 14% annually. A growing share of that money is going into colocation. And the tenants writing those checks are pushing power density well beyond what most existing white-space designs were sized for.
A colocation facility commissioned in 2018 for 8 kW racks now hosts tenants deploying 120–140 kW Blackwell racks in the same zones. The original PDU sizing, transformer K-factor, harmonic profiles, and protection coordination were all calibrated for a different load class. Your monitoring still reports averaged 1 Hz data. Which means you're signing tenant SLAs against a capacity model that doesn't match the physics of what's actually running.
$250B–$650B
AI infrastructure capex projection through 2026 (Goldman Sachs)
$3T
Total data center capex by 2030. 14% CAGR pulled by AI workloads
20–40%
White-space capacity stranded by derating without validated headroom
The capacity-vs-SLA problem
Most colo operators confront the same compounding bind:
- Design specs were written for a different load class. A 200 A PDU rated for steady-state server load has different effective capacity under a GPU load with elevated harmonic content. Heating from 5th and 7th harmonic THD reduces real headroom. Your spec sheet doesn't capture that.
- Mixed-tenant zones share infrastructure. One zone has legacy CPU racks, the next has a GPU tenant pushing 120 kW racks. The transformers, switchgear, and paralleling between them see a load profile no one designed for. Harmonic stress accumulates on equipment rated for clean sinusoidal current.
- You sign tenant SLAs against an averaged dashboard. Per-zone capacity numbers from your monitoring don't reflect the shared-infrastructure penalty above. When the tenant deploys and a protection event happens, you're explaining a downtime credit to a customer with competitive alternatives a phone call away.
The conservative answer is to derate harder. Leave 30–40% margin. That strands capacity you've already paid to build. The aggressive answer is to sign SLAs against the dashboard. That occasionally costs you a tenant.
Spectra computed at runtime via weighted superposition of canonical UPS-rectifier and GPU-PSU presets per IEEE 1459. K-class per IEEE C57.110. Per-rack power: 8 kW (typical 2018-era colo design) → 130 kW (Blackwell NVL72 nominal, EVD-110). The transformer is stressed differently at each extreme: K-class limit at 100% legacy (harmonic heating), amperage limit at 100% GPU (fundamental current ×15).
Per-zone capacity, validated continuously
Verdigris gives you a measured operating limit per circuit, refreshed continuously as the load mix in each tenant zone shifts.
8 kHz waveform capture on the circuits feeding each zone resolves harmonic content through the 25th order. The Verdigris Signals layer translates that into the actual safe capacity for each circuit given the load it's currently carrying. Not the nameplate. Not the historical assumption. The validated electrical limit, today.
That gives you a defensible tenant SLA. When a GPU tenant wants to expand into a zone, you can say "here's the measured headroom on this PDU under your current load profile, and here's how it changes when you double the rack count". Backed by physics, not spreadsheet projections.
Why incumbent BMS doesn't close the gap
Schneider, Eaton, Vertiv, and ABB all market predictive maintenance for colocation. Their data sources are still 1 Hz BMS polling and threshold-based alarms. The failure modes that take down GPU-dense tenant zones (sub-cycle transients, harmonic stress on transformers and rectifiers) live below their measurement floor. Adding more sensors at the same sample rate doesn't close the gap.
What changes for colo operators
Recover stranded white-space capacity
Replace conservative derating with per-zone validated limits. Free up the 20–40% margin you're holding back to cover loads you can't measure.
Defendable tenant SLAs
Sign capacity commitments grounded in measured headroom. When a tenant pushes up against the limit, you see it well before trip-time, with planning-window lead time instead of an after-the-fact protection event.
Mixed-tenant zone visibility
Per-circuit electrical fingerprinting lets you separate signal contributions from each tenant. Investigate degradation without guessing whose load caused it.
Capacity planning that matches your physics
When you're sizing a new deployment, base it on measured limits from circuits already running similar load profiles. Not on the design spec from a different decade.
Read More
Three Approaches to Predictive Maintenance in AI Data Centers
Schedule, condition, and physics-based comparison. Where each approach detects relative to a UPS-rectifier failure event.
Read more
When training jobs die and the GPU logs don't explain it
The sampling-gap failure mode tenants encounter. How 1 Hz monitoring smooths over the events that actually trip rack-level protection.
Read more
What is 8 kHz monitoring?
Why sample rate is the bottleneck on capacity validation, and how 8 kHz waveform analysis grounds tenant SLAs in measured headroom.
Read more
The High Cost of Power Guesswork in Data Centers
When operators lack real-time visibility into their power infrastructure, they don't only lose control; they lose money.
Read morePricing & SKUs
Per-MW for Cloud and Sentinel. Cloud Light per-gateway for brownfield pilots. Perpetual license + maintenance for Forge.
Read moreValidate your white-space capacity
Replace conservative tenant-zone derating with measured operating limits. Sign SLAs you can defend with physics.