The method
Adaptive Current Modulation makes current a controlled variable.
Power systems have always reacted. ACM anticipates. One mathematical method shapes current across 22 orders of magnitude, from the drift of a national grid to the switching of a transistor.
What changes
Reactive regulation becomes predictive control.
Conventional power delivery waits for a transient, then corrects it. ACM models the workload signature ahead of the event and shapes the current draw before the excursion arrives. The compute does not slow down. The power envelope stops dictating the plan.
How it runs
Four steps, run continuously on the rail.
Sense, predict, determine, control. The reference numerals match the published method figures.
- 110
Sense
Read the rail as a signal rather than as an average.
- 120
Predict
Anticipate the next demand from the shape already present.
- 130
Determine
Choose the correction that holds the envelope.
- 140
Control
Apply it, then measure the result and repeat.
Where it sits
The method moves from the chip into the rack, facility, grid and beyond.
The sequence plays outward from the die to orbit in six scales; pause it, pick a scale, or set a site power at the rack and facility steps to read the capacity released.
Six line drawings step outward along the power path: a processor package with the rail into its die, one rack of eight compute units on a power spine, a data hall from above with power reaching every rack, a substation and transmission lines feeding several facilities, a wireframe globe of linked sites, and an orbital platform with solar arrays above the Earth. Each drawing marks the power path in green, and the panel beside it names the ATHLAZ product acting at that scale: iPMM at the chip, ADAPT at the rack and the facility, and the ATHLAZ Power Network from the grid outward.
Scale 1 of 6
Control begins at the die.
Intelligent Power Management Module brings programmable power into the silicon power path and the rail that feeds it.
iPMM at this scale+2.1 MW
Released
+26
80 kW racks
The mechanism
Workload fingerprinting: know the workload, then give it exactly the power it needs.
Think of recognising a song from a few seconds of sound. Here the song is a power draw. Every workload has a signature in the frequency domain, and once the signature is known, the workload is served from memory.
Read the live draw
Every workload draws power in a pattern as distinctive as a signature. ADAPT reads that pattern in the frequency domain rather than one averaged number.
Recognise the signature
The pattern is matched against the fingerprint library. A workload seen once is handled from memory the next time, so the fleet gets leaner the longer it runs.
Apply the operating point
The workload gets exactly the power it needs, ahead of the excursion rather than after it. The same signature also flags hardware that should not be there.
One more concept belongs to the recognising side of the loop. The runtime takes a reading of what a machine is doing from its own power draw: how hard a part is working, what kind of work it is doing, and when the work changes. We call that reading the workload fingerprint. It is one of the inventions inside the method, and it reads the power, clock and utilisation counters the platform already exposes.
Conventional, reactive
Measure, then cap.
The ceiling is sized for the worst case, so every rack carries power it never uses. Capping that margin after the fact costs you throughput.
Illustrative, not to scale
Shaded band: paid for, never used
ADAPT, predictive
Read, then modulate.
The draw is resolved before the excursion arrives, so the saving costs nothing. In plain terms the engine was revving at the traffic lights. ADAPT gives it only the fuel it needs.
Illustrative, not to scale
Delivered power tracks the live workload
The inventions
Five inventions around one method.
This is the shape of what is owned, at the level it is published: five inventions that work as one method.
The family describes five inventions around one method. The difference the examiner recognised is between reading a power signal spectrally and computing the correction there. Reading is measurement. Correcting in that domain is control, and that is the whole difference between reacting to power and programming it.
Adaptive Current Modulation
The method itself: sense the power signal, predict what the load is about to ask for, and deliver an updated parameter instead of correcting after the event.
AI-driven waveform synthesis
The shape of the delivered power becomes an output of the control, not a fixed property of the supply.
The multi-objective engine
Power, throughput and thermal limits are held together at once, so a gain in one is not quietly paid for by another.
Predictive load management
The operating point moves ahead of demand rather than behind it, which is why throughput is protected rather than traded away.
Electrical fingerprinting
Recognising what a machine is doing from its power draw alone, so a workload seen once is served from memory the next time.
Examined by the UK Intellectual Property Office, which found the electrical-signature and frequency-domain claims novel and inventive. Filed, not granted. The algorithms, fingerprinting methodology and control logic are held as trade secrets.
Published applications: WO 2026/167363 A1 (international) and GB2704473 A (United Kingdom).
22 orders of magnitude
One method, every band that carries power.
The mathematics does not care whether the current is drifting across a national grid or switching inside a transistor. Only the power stage changes.
Grid
Sub millihertz to hertz
Frequency response, demand shaping, multi site coordination.
Facility
Hertz to kilohertz
Envelope control, thermal headroom, switchgear relief.
Rail
Kilohertz to megahertz
Predictive regulation inside the power delivery network.
Silicon
Megahertz to terahertz
Spectral control at the die and in lithography.
Compute is measured today, on NVIDIA H100 NVL. The other bands are designed, and each is open to a feasibility programme on your equipment.
Evidence
What is measured, and where.
- Measured architecture
- NVIDIA H100 NVL.
- Sustained power reduction
- Up to 21% of GPU die power over a continuous 48 hour window on eight H100 NVL GPUs, with GPUs 0 to 5 managed.
- Baseline arm
- The run held a baseline arm under an equal power cap on the same eight GPU node. We re-derived the result in house, and it came within 0.4 percentage points of the nearest archived reference. The cap value is shared with the assessment.
- Tokens per watt
- Up to +22%, derived from the measured run.
- Throughput
- No throughput change in the measured run.
- Projected reach
- Up to 30-40% die power reduction projected as management scope widens across the fleet and the method is tuned per workload family. A design estimate, not a measurement; validated on your own fleet before it enters any agreement.
- Other architectures
- Part of 118 days of testing across six accelerator architectures. On yours, a three week baseline run on your own fleet sets the figure, and until then site arithmetic runs on your inputs.