Running unknown compute is no longer an option
Which model answered, on whose machine, in which country? Running unknown compute is no longer an option. The standard is measurement, not trust.
When you send work to an AI today, you almost never know what actually ran it. Which model? Which version? On whose machine, in which country, with what else loaded alongside it? You get an answer back and a promise.
That was tolerable when AI drafted emails. It isn’t tolerable when AI approves payments, files documents, reads patient records or acts for your company with partners.
The industry has spent years calling this “trust”. I think that’s the wrong word. Hardware has been able to measure what it’s running, and sign that measurement, for a long time. Trust is a decision you make afterwards, on top of that evidence. So the standard shouldn’t be “trust us”. It should be: here is what was measured, here is who signed it, you decide.
Measurement has depths, and an honest system says which one it reached:
- The connection: this answer came from this provider’s service, in this session, at this time.
- The machine: this ran on hardware that measured its own state and signed it.
- The model: the specific weights that were loaded were measured, and bound to the company accountable for them.
And every measurement raises the question I’ve wrestled with for most of my career: who signs the reference? A measurement only means something when it’s compared with a known-good value that someone accountable has signed. That brings you straight back to the corporate root.
The machinery is already there. Nearly every business PC and server shipped in the last decade has a TPM that can measure it, and the newest data-centre GPUs can attest to their own state. Most companies never use it.
A shallow measurement isn’t a fake one. It’s a true statement of how far the evidence goes, and it gets deeper as the hardware does. What doesn’t work any more is no measurement at all: compute nobody can describe, producing work nobody can check.
The buyers who matter, in regulated industries and in government, are already asking where their AI ran. Soon their agents will ask too, on every request, and they’ll route the work to compute that can answer.
Measured, not trusted.
Third in a short series on how I see AI working from here. Rootz is building, partnering and investing in this future every day.