Praecise Metal
Bare-metal AI computing

Run AI on your own metal.

Praecise builds the software and infrastructure to operate AI on bare metal you own — measurement, placement, power and thermal control, forecasting, connectivity and cost. The operating capability that otherwise takes years to build, so the hardware you bought does the work you bought it for.

4–9× A declared power multiplier, wrong by this much against draw read at the supply. basis · measured
8.1× Carbon understated by a national generation average where consumption was the figure that counted. basis · computed from measured
228kWh/yr Drawn by two always-on machines at idle, before a single request arrives. basis · measured
pue Reported by a site with no facility meter. The answer is unaccounted. basis · not measured
CH 01The gap

Owning AI compute is one thing. Operating it is another — and that is what takes years.

The hardware can be bought and the building can be built. What does not arrive with them is the software and the operational expertise to run the thing — measurement, placement, power and thermal control, forecasting, and the discipline that keeps the figures defensible. Built from scratch, that is years.

Praecise is that operating capability. It is for operators putting AI on their own hardware — from a first cluster to infrastructure at scale — who hold the metal but not yet the software or the expertise to run it, and cannot wait years to acquire both.

The hardware is the easy part

Racks, accelerators, power and cooling can be bought. The software that turns them into an operated system does not come in the crate.

Bought in weeks; built in years.
At scale, without the catch-up

For those with the mandate and the hardware to stand up AI compute at scale, but not the operating software or the expertise, and no years to spend building it.

The capability, delivered.
Measured, not modelled

Every figure a Praecise site reports is read from an instrument on the machine, not inferred from a spec sheet. That is where operating it well begins.

The denominator is real.
CH 02Measurement

Optimising a number nobody has read is theatre.

Most efficiency work starts with a model of a site. We start with an instrument on it. The gap between the two is not a rounding error: on the first fleet we 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 are free and cover the package. They do not cover disks, fans, conversion loss or anything else in the chassis. A metered PDU or a UPS with power reporting is the highest-value hardware on an efficiency programme, because without it the denominator is a guess and every percentage above it is a guess too.

Package, idle
13 W
Package, inference
29 W
Package cap
140 W
Wall draw
unaccounted
Per token
1.08 J
What the reading changed

The headline technique in this field is a controller that arbitrates a workload against a power cap. It wins where the cap binds. Here the cap sits at 140 W and the workload peaks near 30 W, so there is nothing to arbitrate. What binds is idle. Porting the fashionable controller would have solved a problem the hardware does not have.

CH 03What we build

Software and infrastructure design, in the same practice.

The reason these sit together is that most efficiency lives in the seam between them. A control loop is only as good as the meter it reads, and a meter is only worth installing if something acts on it.

Telemetry

Draw, temperature, airflow and utilisation read at the supply and at the device, with a privileged reader kept small so the agent stays unprivileged.

Per-process attribution that makes waste visible.
Forecasting

Quantile forecasts of load, draw and occupancy, backtested on the site's own data and scored before they are trusted.

Decisions that account for the next six hours rather than the last reading.
Placement

Consolidation, suspend and resume across real failure domains, with the wake cost priced against the idle saved.

Fewer machines awake, and none of them the wrong ones.
Power control

Performance ceilings and turbo state rather than blunt governor changes, tuned to shave peak draw without a cliff.

Demand charges cut without stretching the work.
Thermal and cooling

Headroom per domain, fan work as a measured cost, and cooling strategy chosen against the site's climate rather than a default.

Lower overhead, and a water figure that holds up.
Carbon and tariff

Deferrable work moved to the hours that are actually cheaper or cleaner, and left alone where the spread does not justify it.

Shifting only where shifting pays.
Connectivity

Operator-owned fabric, routing and reach designed so bandwidth cost and latency are properties you set rather than inherit.

Egress that is a line item you control.
Reporting

The indicators regulators now require, produced from the same instruments the control loop reads.

One set of numbers, used for operating and for filing.
Site design

Layout, power distribution, failure domains and growth increments for new sites and retrofits, sized to the building that exists.

Capacity added in steps a site can actually take.
CH 04Refusals

An instrument that guesses is worse than no instrument.

Every system we build returns nothing where it does not know, and says which kind of nothing it is. This is the single most valuable property in the whole practice, because a plausible wrong number propagates into a decision and an absent one does not.

pue_is_reportable

A site with no facility meter upstream of the IT load cannot compute a PUE.

Records the overhead as unaccounted. A published 1.0 claims the overhead was measured and was nil.
consumption_intensity

Where imports are material and their origin is undeclared, no carbon intensity is returned.

Treating unknown imports as zero-carbon is how a coal-backed third of a grid disappears from a report while every arithmetic step stays correct.
shifting_pays

Below a 25 per cent spread between cheap and peak hours, deferral is refused.

Returns nothing to gain rather than picking an arbitrary best slot. A scheduler that always defers looks like a bug to whoever is waiting.
price

A minute not covered by any tariff period is not priced.

Charging zero for uncovered hours makes them the cheapest of the day and quietly attracts every deferrable job on the site.
water

On-site and source water are carried separately and never summed.

A closed-loop site is a structural zero on-site and can still sit a long way down the table on the water its power consumed.
no rating

An undeclared limit draws no bar.

No denominator is not zero headroom. A full meter and an empty one are both claims about a machine nobody measured.
CH 05Carbon

The cleanest grid is not the cleanest place to compute.

A site we instrumented sits on a grid generating almost entirely from hydro, around 25 gCO₂eq/kWh, one of the three lowest production intensities in the world. Every instinct says a rack there is about as clean as computing gets.

The answer is on the import meter. About a third of that grid's consumption is imported, across interconnectors from the most coal-dependent grids nearby. Consumption intensity comes out at 203 gCO₂eq/kWh. The standard behind the software carbon intensity metric is location-based by design. It asks what the wire carried, and will not take a national generation average instead.

The gap widens in exactly the wrong season. When reservoirs are low the grid imports more, and it competes for those imports when the neighbouring grids are short and the marginal generator is coal. A system with the clean figure hard-coded reports its best numbers during the worst weeks.

Generated
25 gCO₂eq/kWh
Consumed
203 gCO₂eq/kWh
Imported share
32 %
On-site water
0.00 L/kWh
Source water
contested
CH 06Forecasting

Every power decision was already a forecast with a horizon of zero.

Whether to consolidate depends on whether load is about to return. Whether to shed depends on how long the supply will last. Whether to defer depends on what the next six hours cost. All three were being answered from the present reading.

Quantiles, not points

Provisioning to an expected value under-provisions half the time by construction. The planner sizes against the upper path, and crossing quantiles are detected rather than planned against.

Intervals come from each method's own residuals, because load floors at idle and spikes upward.
The model earns its place

A foundation model never simply replaces the baseline. Both run, both are backtested on rolling held-out windows of the site's own data, and the model is used only while it is measurably ahead.

A 10 per cent margin to be adopted, and merely not losing to be kept.
Scored for the decision

Scale-free error for the median, so watts and occupancy counts are comparable. Pinball loss for the quantiles, which punishes overshooting and undershooting differently, the way a planner's exposure does.

A better median with worse quantiles is refused outright.
Eight windows minimum

One window is an anecdote. Swapping the forecaster on an anecdote is how noise becomes policy.

An idle fleet at 08:50 on a working day is idle in a way that is about to stop being true.
CH 07Reporting

The numbers you file should be the numbers you run on.

Data centres at or above 500 kW of IT power increasingly report annually on energy performance and sustainability indicators. Rating schemes and minimum performance standards are following, with carbon-neutral targets for the sector.

Most operators answer this with a reporting exercise bolted on once a year. We instrument the site so the same readings drive the control loop and populate the return. That costs less and goes wrong less often.

IndicatorWhat it measuresWhat it needs on site
PUETotal facility energy over IT energya facility meter upstream of the IT load
WUELitres consumed per kWh of IT energymake-up water metering, or a declared closed loop
ERFEnergy reused rather than rejectedheat offtake metering at the boundary
REFRenewable share of energy consumedsupply contract and grid mix, consumption-based
Where a figure cannot be produced

We install the instrument or we report the indicator as unaccounted. We do not estimate one into existence, because an estimate filed once becomes a baseline everything afterwards is measured against.

CH 08Work with us

Four ways in, and the first one is always measurement.

Survey

We instrument what you have and report what it draws, what that costs, what it emits and which figures cannot yet be produced.

Ends with a metered baseline you own.
Programme

Efficiency levers priced individually against that baseline, applied in order of return, each one reversible.

Ends with the saving measured, not modelled.
Software

Control loops, forecasting, placement, telemetry and reporting built into your operations rather than sold as a dashboard.

Ends with the site operating itself.
Design

New sites and retrofits: layout, power distribution, failure domains, cooling strategy and growth increments.

Ends with a site that can be added to.