Praecise Metal: the machines underneath
Buying racks, accelerators, power and cooling is the easy part of running AI on owned hardware; the operating capability is what takes years. Praecise Metal is that capability, measured from instruments on the machine, from core clusters to edge cells.
The agent era will be decided as much by watts and degrees as by models. Every agent that answers a customer or reconciles a ledger runs on a machine that draws power, makes heat and costs money whether or not it is working. The organisations that operate those machines well will set the price and the reliability of everything above them. The ones that guess will pay for their guesses.
Praecise Metal is owned machines, placed, powered and cooled to measured targets. It is measurement, forecasting, placement, power and thermal control, built on Praecise Furcate. Praecise runs its own machines on it and builds the same capability on metal its customers own.
The hardware is the easy part
Demand is not in doubt. The International Energy Agency's Energy and AI report projects that global data centre electricity use will roughly double by 2030, with AI the most important driver. Racks, accelerators, power and cooling can be bought. What cannot be bought off a shelf is the operating software and expertise that make the hardware do the work it was bought for. That is what takes years to build, and it is what most operators putting AI on their own hardware lack.
Praecise Metal is that capability, delivered to operators who have the metal and cannot wait years to learn how to run it.
Measured, not declared
The governing principle is that figures are read from instruments on the machine, not inferred from specification sheets. The difference is not academic. On one fleet, once it was instrumented, the declared multiplier used to convert processor draw into wall draw was wrong by between four and nine times, in a direction that flattered the report. Processor counters cover only the package. A metered PDU, or a UPS that reports its power, is the highest-value hardware in an efficiency programme.
Measurement changes the conclusions as well as the numbers. On the machines Praecise studied, idle draw, not the power cap, was the binding constraint, so a standard power-cap controller would have solved a problem the hardware did not have.
An instrument that refuses to guess
Metal is built to return nothing rather than a guess. Without a facility meter, PUE is recorded as unaccounted, not as 1.0. Without a declared import origin, no consumption carbon intensity is given. Deferral is refused when the gap between cheap and peak prices is too small to justify it. Minutes outside any tariff period are not priced. On-site water and source water are kept separate and never summed.
The carbon case shows why. On one hydro-heavy grid that generates at about 25 gCO2eq/kWh, roughly 32% of consumption is imported from coal-heavy neighbours, and consumption intensity comes out at 203 gCO2eq/kWh. A report that hard-codes the clean generation figure shows its best numbers in the worst weeks. For an executive signing a sustainability disclosure, that is the difference between a figure that holds and one that does not. Reporting indicators are produced from the same instruments the control loop reads.
From measurement to control
With a trustworthy baseline in place, Metal acts on it. Telemetry reads draw, temperature, airflow and utilisation at the supply and the device. Forecasting produces quantile forecasts backtested on the site's own data. Placement consolidates work and suspends and resumes machines across failure domains, pricing the cost of waking against the idle power saved. Power control uses performance ceilings and turbo state rather than blunt governor changes. Thermal control tracks headroom per domain and treats fan work as a measured cost. Carbon and tariff logic moves deferrable work only where the spread justifies it. Connectivity covers operator-owned fabric, routing and reach, and site design covers layout, power distribution, failure domains and growth increments for new sites and retrofits.
Everything you own, at every scale
Metal reads ownership across two kinds of estate. Core clusters are machines in a hall, measured and operated end to end: placement, power, thermal and cost. Edge cells are a home, a shop, a plant or a district run as a cluster of its own on Praecise Furcate, with its own control, still working when its link is cut. Each cell holds its key in hardware and sends summaries upward, never its raw state.
The ownership view brings inventory, health, power and cost for every machine and every cell together, read from the operator's own chain of keys. Metal reads ownership; it never reaches into a cell to control it. The first edge cell runs today. Nested cells, scoped managers and cell roll-ups in the ownership view are coming.
Who it is for
Praecise Metal is for operators putting AI on their own hardware, from a first cluster to infrastructure at scale. For a CTO it is a metered baseline and the control loops built on it. For an executive it is cost, power and carbon figures that survive scrutiny because they were measured. Infrastructure owners who would rather not run their own can take GPU capacity on Praecise Metal.
Engagements follow the evidence. A survey instruments what you have and produces a metered baseline you own. A programme prices each efficiency lever against that baseline, each one reversible. Software builds the control loops, forecasting, placement, telemetry and reporting into operations on Praecise Furcate. Design covers layout, power, failure domains, cooling and growth for new or retrofitted sites.
How it connects
Metal is the bottom layer of the Praecise stack, and everything above runs on it. Its software is Praecise Furcate, built on the open Furcate operating system, which gives each machine and each cell its identity, policy and ledger. Praecise Cloud runs agents, MCP servers, models and data on the machines Metal operates. The automation layer, Praecise Inbox, Ipnops and Skerry, runs on Cloud, so every agent that answers mail, runs operations or makes a payment is ultimately drawing power from metal that Praecise measured.
The floor everything stands on
The cost, reliability and footprint of agents will be set by the machines underneath them, and those machines can only be run well if they are measured honestly. Praecise Metal measures first, controls second and refuses to guess. Read more on the Praecise Metal page.
