A leak is one word hiding three different failures.
Liquid leaks have entered compute-availability management: NVIDIA Mission Control surfaces a detected leak as a compute-node state alongside node health, not only as a facilities alarm. Most project boundaries have not caught up. They still compress three risks into one checkbox: leak detection installed.
Definition
The three clocks (three clocks of liquid-cooling leaks · 三个时钟) are the three timelines on which a liquid-cooling leak is created: manufacturing (a defect that leaves the factory), commissioning (a joint disturbed or left untested during installation) and service life (hoses, elastomers and fluid chemistry that age out). Each clock has a different owner and leaves different evidence, so a leak sensor reports where liquid arrived, not who created the risk.
Manufacturing
Owner: the factory
What it looks like. A weld or braze that escapes the factory. A void in a cold-plate seam. A crimp outside tolerance.
What survives as evidence. Bounded factory acceptance test (FAT) leak-test records, traceable to serial number. Crimp process parameters. Material certificates for elastomers and wetted metals.
Commissioning
Owner: the integrator and the commissioning agent
What it looks like. A quick disconnect disturbed during install. A field weld that was flushed but never passivated. A fitting that looked dry at handover.
What survives as evidence. Connection records for every joint made or disturbed on site. Flushing and cleanliness results. Pressure tests against written acceptance criteria. Tested compute-and-liquid isolation sequences.
Service life
Owner: the operator
What it looks like. A hose that ages out in year four. An elastomer that hardened. A fluid whose inhibitor package was never topped up.
What survives as evidence. Fluid-quality trends. System-performance trends against commissioning baselines. Service logs for every disturbance of a boundary. Replacement intervals with a basis.
A sensor tells you liquid has arrived. It cannot tell you who created the risk.
Leak detection is necessary and it is late. By the time a rope sensor is wet, the question is no longer whether there is a leak but whether the response sequence works, whether the compute can be isolated from the liquid safely, and what evidence survives the incident. None of that is answered by the sensor. It is answered by a chain that was built before the first drop:
- bounded FAT leak-test records
- connection and flushing records
- fluid-quality and system-performance trends
- tested compute-and-liquid isolation sequences
What gets documented can be priced
Project owners, and increasingly their insurers and lenders, may ask which clock a project has covered. You can price evidence. You cannot price "the work was done carefully". A project that runs well for years and leaves no records has succeeded operationally and failed commercially the first time someone asks for proof.
Reference: NVIDIA Mission Control Systems Administration Guide, leak detection.
We build the three clocks into the scope. The factory clock is covered by what we require of suppliers before a component is released. The commissioning clock is our own work and is recorded joint by joint. The service clock is handed to the operator as trended baselines, not as a binder.
Questions this page answers
How is liquid-cooling leak risk classified, and who is responsible?
By the clock on which the risk was created. Manufacturing leaks (a void in a braze, a crimp out of tolerance) belong to the factory and its acceptance-test records; commissioning leaks (a quick disconnect disturbed during install, a weld never passivated) belong to the integrator and commissioning agent; service-life leaks (aged hoses, hardened elastomers, depleted inhibitor) belong to the operator and its trend data. Ten leak points →
What does a leak sensor actually tell you?
That liquid has arrived at the sensor. It cannot tell you which of the three clocks created the risk, so the evidence for that has to exist before the leak: factory test records traceable to serial numbers, a connection log for every joint disturbed on site, and fluid-quality trends in service. Commissioning and acceptance criteria →
What evidence should exist for each clock?
Clock one: factory leak-test records per serial, crimp process parameters and material certificates. Clock two: a connection log for every joint made or disturbed, flush and passivation records, and a pressure test of the assembled loop. Clock three: fluid-chemistry trends, performance trends against commissioning baselines, and a service log per boundary. What the turnover package contains →