Investors
One method, filed once, licensable twice.
For 140 years electricity was fixed. ATHLAZ makes it programmable, and licenses the method rather than manufacturing anything.
The shape of it
What is bought here is a licence to a method, not a product to manufacture.
The asset
One method, filed and examined, with the algorithms and control logic held as trade secrets.
The evidence
Up to 21% less GPU die power, measured over 48 hours on NVIDIA H100 NVL under paired arms and an equal power cap.
The model
The same method licensed twice: software per GPU today, silicon design in and royalty next.
What compounds
The fingerprint library only grows on fleets that run it, so the fleet that starts first stays ahead.
The constraint
Power, not chips, is the ceiling on AI.
Demand for AI compute is rising faster than grid capacity can be built. A connection takes years and switchgear takes the better part of a year. You cannot build your way out of that, but you can stop wasting what is already installed.
Up to 21%
Less GPU die power, measured
Sustained across a continuous 48 hour window.
0%
Throughput change in the run, derived
Every service level intact, on up to 21% less GPU die power.
Up to +22%
Tokens per watt, derived, balanced mode
The same tokens on less energy.
118 days
Of testing
Across six accelerator architectures.
Measured on production NVIDIA hardware, eight H100 NVL GPUs with managed and baseline control arms, under signed per GPU energy records.
The milestones
Three milestones, each one testable.
The first is already how a customer starts.
- Fleets run it
- A three week validation on the operator's own hardware, then production licences.
- Results extend
- Each further accelerator and workload shows its result on its own paired baseline, the same design as the H100 NVL run.
- One part commits
- A manufacturer designs the method into a product programme, and royalties begin.
The asset
The method is the product. Everything else is the design that makes it usable.
Adaptive Current Modulation is a method, not a device. It covers reading, predicting and reshaping a power signal in the frequency domain.
Filed as a UK application and its divisional, a Taiwan filing, and an international application published as WO 2026/167363, opening the national phase.
The UK Intellectual Property Office's search and examination found claims 5 and 15, the electrical-signature and frequency-domain claims, novel and inventive. The application is published as GB2704473 A.
Published applications: WO 2026/167363 A1 (international) and GB2704473 A (United Kingdom).
Programme memberships and advisers. The method is architecture agnostic and runs across a mixed fleet.
The moat
A library that only grows on fleets that run it.
Workload fingerprinting compounds. A workload seen once is handled from memory the next time, so the fleet that starts first stays ahead, and the advantage is not something a competitor can buy.
The road
Prove, scale, licence, platform.
ADAPT on customer racks
Software revenue on fleets already installed, licensed per GPU per month, with results measured on the customer's own hardware.
Bench, then test silicon
The method into reconfigurable hardware against instrumented loads, then to test silicon. A centre of excellence around the engineering.
First design in
A manufacturer commits the part to a product programme. Production royalties follow, in the pattern established by processor IP.
A standard across domains
The same method, a different power stage, one domain at a time. Chip, rack, facility, grid.
Hardware follows a gate the software business opens: break even held for a quarter, twelve months of runway and the second close received. ADAPT carries the plan, and iPMM is what it earns the right to build.
The model
Licensed like an instruction set. Never manufactured.
One method, licensed twice
- Software today
- Licensed per GPU per month. It starts on a three week validation on the customer's own hardware, and the fingerprint library it builds only grows on fleets that run it.
- Silicon next
- Design in licence, then a royalty on every part shipped. Designed in, it cannot be uninstalled without redesigning the board.
- Who pays
- Software today: GPU cloud and neocloud operators, AI factory and data centre owners, and public bodies running funded compute, per GPU per month. Silicon next: silicon and accelerator vendors, power and regulator vendors, system OEMs and board builders, and manufacturers beyond compute, as design in and royalty.
- Capital shape
- No fabrication capital and no inventory. Design houses, packaging partners and foundries are who we pay, not who pays us.
- Where we differ
- An instruction set licence is paid per unit. This one is paid against a measurable reduction in the energy a part consumes.
- Entity
- ATHLAZ Limited, England and Wales, company 16357048. A UK based asset.
Not an offer of securities. Forward looking statements are intent, not forecasts.