Stranded Capacity in Data Centers: The Power You Paid For but Cannot Use
Stranded capacity is power infrastructure you provisioned and paid for but cannot safely utilize. Verdigris helps operators recover it with validated headroom data.
I have spent most of my career thinking about the gap between what infrastructure can do and what people believe it can do. In semiconductors, that gap was measured in clock cycles and thermal envelopes. In data centers, it is measured in megawatts.
The number that keeps coming up in our deployments is this: most facilities are running at 60-70% of their design capacity. Not because the infrastructure failed. Because nobody can prove the remaining 30-40% is safe to use.
That gap has a name. Stranded capacity. And for operators trying to meet GPU-driven demand without a two-year construction timeline, it represents the most immediate path to more compute.
The compounding problem
Stranded capacity is not the result of bad engineering. It is the result of good engineering practiced without adequate measurement.
An engineer designs a facility with safety margins. Operations adds derating for observed issues. Thermal constraints reduce available power below the electrical ratings. Equipment ages. Nobody re-validates the original assumptions, because there has been no practical way to do so continuously.
Take a facility rated at 10 MW. After a decade of conservative assumptions layering on top of each other, it can only safely deliver 7 MW. The 3 MW gap is not broken infrastructure. It is infrastructure that nobody has the data to trust. Every year that passes without validation, the assumptions get more conservative and the stranded gap gets wider.
We saw this pattern at T-Mobile. Stranded capacity across their portfolio sat at 15-20% before Verdigris was deployed. After, it dropped below 5%. The infrastructure did not change. The measurement did.
Why BMS cannot solve this
Building management systems track consumption. They tell you how much power you are using. They do not tell you how much power you can safely deliver. Those are fundamentally different questions, and they require fundamentally different instrumentation.
Validated headroom means understanding the actual electrical condition of every component in the power path: transformers, switchgear, UPS systems, PDUs, panelboards. Has any component degraded in a way that reduces the safe operating limit below the design specification? Is the standby UPS actually healthy enough to take the load during a switchover?
At 1 Hz polling, you cannot answer those questions. You can track watts. You cannot validate capacity. This is the measurement problem at the center of every stranded capacity conversation.
What validated headroom looks like
At Verdigris, we instrument the power chain at 8 kHz, continuously assessing each component for its actual operating condition rather than its nameplate rating. The output is a validated operating limit for each circuit and each path through the distribution hierarchy.
Where the design says "10 MW," we tell you whether the infrastructure can actually deliver 10 MW safely today. When the validated limit exceeds the current load, the difference is recoverable capacity, not based on nameplate ratings, but validated by continuous electrical measurement.
This distinction matters more than it sounds. Nameplate capacity is a static number from the day the equipment was commissioned. Validated capacity is a living number that changes as equipment ages, as loads shift, as thermal conditions evolve. One is a design document. The other is operational truth.
The T-Mobile story
When T-Mobile deployed Verdigris across their data center portfolio, the capacity recovery numbers were significant. Across 100-150 MW of monitored power, the team reclaimed 10-15 MW of stranded capacity, projecting roughly $100M in deferred CapEx. John Coster, their Sr. Manager of Innovation, Planning and Strategy, noted in a published case study that Verdigris "has been instrumental in refining our capital planning processes, enabling us to make more informed and strategic investment decisions across our facilities."
But the story I keep coming back to is what we found underneath the capacity data. 4% of the UPS rectifier fleet was exhibiting electrical degradation that standard monitoring could not detect. Zero alarms had fired. Every parameter looked normal.
This is the part that makes stranded capacity recovery more than a financial exercise. You cannot responsibly recover headroom above equipment that is trending toward failure. The capacity question and the reliability question are the same question, answered by the same measurement. At T-Mobile, that combined answer delivered a 6:1 ROI with a 5-month payback.
The real cost of new construction
For operators facing GPU workload demand, the alternative to capacity recovery is building new. Here's what that looks like:
| Approach | Timeline | Cost | Risk |
|---|---|---|---|
| New construction | 18-36 months | $7M-$25M per MW | Permitting, supply chain, construction delays |
| Validated capacity recovery | 30-90 days | Measurement infrastructure | Requires validated data to confirm safety |
| Unvalidated capacity recovery | Immediate | Zero | Deploying load above equipment of unknown health |
The third row is what happens when operators eyeball headroom based on nameplate ratings. I understand the temptation. GPU demand is urgent. Construction timelines are long. But deploying additional load above degrading infrastructure creates the conditions for cascading failure. The urgency of AI demand does not change the physics.
Validated recovery sits in between: results in weeks rather than years, at a fraction of the cost of new construction, with the safety confidence that nameplate assumptions alone cannot provide.
Capacity planning as a continuous capability
Stranded capacity recovery is not a one-time audit. Equipment conditions change. Load profiles shift. What is safe headroom today may not be safe headroom in six months.
The Verdigris Signals platform provides continuous validation. As equipment health changes, the validated operating limits update automatically. Operators see how much capacity is available and whether any component in the power path is constraining it.
This changes capacity planning from a periodic engineering study into a continuous operational capability. At T-Mobile, site audits that used to take weeks now take hours. The data is always current. The headroom numbers are always validated.
Why this matters beyond any single facility
The entire AI buildout is being erected on a grid that was not designed for it. The industry is spending over $1 trillion projected for global data center infrastructure by 2030 (Goldman Sachs, 2024). Some of that spending is unavoidable. But a meaningful portion of it is compensating for capacity that already exists, stranded behind conservative assumptions that nobody has the measurement to challenge.
Every megawatt of validated, recovered capacity is a megawatt that does not need to be built, permitted, wired, and cooled from scratch. In a world where power availability is the binding constraint on AI progress, the fastest new megawatt is the one you already have.
Talk to an engineer about what your facilities can actually deliver.