Monitoring reads what is happening now. We work out what happens next, and we use whichever technique gives the most accurate what, when and why for that class of asset — physics-informed models where physics earns its place, statistical models where it does not.
A data-center operator or a battery offtaker is not buying an algorithm. They are buying an accurate answer to when this asset degrades, why, and what it costs to wait — and the confidence that we will say so when we do not know.
Where the physics earns its place — predicting outside the range an asset has run before — we encode the governing equations rather than borrowing their shape.
We run the same honest comparison for each asset class rather than assuming one method transfers. Across three classes the answer is not the same every time.
On stale or incomplete data the system abstains rather than guessing. In testing it rejected about 70% of degraded rows; error on the rows it accepted fell between roughly 1.5× and 10×.
This is a sampling problem before it is an intelligence problem. The physics that degrades power electronics happens orders of magnitude faster than the systems watching it.
You cannot infer what you never measured. That is why prediction starts with the data path, not the model — and why the Alliance exists.
The visibility gap has a consequence people miss. Sampling fast enough to see degradation produces far more data than any site is going to ship off-premises — so the prediction has to run where the signal is made.
Sampling at kilohertz rates across many channels generates continuous high-volume waveform data. Streaming that to a cloud is not a cost problem, it is a bandwidth impossibility — particularly at the remote and constrained sites where this equipment often lives.
Because inference happens on-site, the raw operating data stays inside your perimeter. What crosses the boundary is a forecast and a record, not the telemetry itself — which is usually the difference between a security review that takes a week and one that takes a quarter.
A site with intermittent or deliberately limited connectivity still gets predictions, because nothing about the forecast depends on reaching a data centre. The link carries results out; it is not in the critical path.
The models run on a compact NVIDIA Jetson Orin NX module — the same class of device that ships with a deployment. It is a mature, industrially available part with a long supply life, not a bespoke board, which matters when the asset it is bolted to has a twenty-year life of its own.
We benchmarked two physics models running concurrently on one module for four hours, then measured what was left over on the GPU.
Measured on an NVIDIA Jetson Orin NX during an edge-deployment benchmark, May 2026 (SOW-EDGE-5). These are deployability measurements — latency, endurance and GPU headroom — not accuracy claims. CPU ran near its ceiling at peak, so how many assets one box can carry is a separate validation.
Four independent evaluations across three asset classes. In three of them a simpler model won or the gain traced to something other than the physics terms. We report each as found.
Physics is the reason we can say anything about conditions an asset has not yet seen.
It is a capability we apply, not an identity we wear, and we hold that distinction exactly so the message survives the next result whichever way it goes. We do not claim anywhere that physics improves accuracy as a blanket rule — four independent evaluations do not support that.
The proprietary methodology behind the models is not disclosed. Patents-pending strategy in development.
A read-only assessment proves the forecast on your own data before any hardware conversation. You see the record; you decide what it is worth.
