Adiabatic Machines builds superconducting AI processors that use adiabatic logic to compute at the physical limit of energy efficiency, consuming 10-100x less power than conventional chips, cooling included. The chips are fabricated in existing commercial foundries, run standard AI models without retraining, and operate at 4 kelvin inside commodity cryocoolers. The company provides a full-stack digital twin, calibrated against fabricated circuits, to demonstrate performance from PyTorch down to Josephson junction dynamics.
Funding
Funding not disclosed
Founders
Product
Problem
Modern AI compute racks draw as much power as small commercial buildings, and conventional transistor-based processors waste roughly 100,000 times more energy than physics requires. Every switch in a traditional chip dumps its driving energy as heat, making AI inference increasingly power-bound and unsustainable at scale.
Solution
Adiabatic Machines builds superconducting AI processors that compute using adiabatic logic, reshaping each gate's energy landscape with an AC clock so switching is nearly reversible and dissipates almost no heat. The chips are superconducting end to end, eliminating interconnect resistance, and run at 4 kelvin inside closed-cycle cryocoolers—the same commodity refrigeration used in MRI machines and quantum computers. The company measures efficiency at the system level, cooling included, and claims 10-100x more compute per watt than conventional hardware. Processors are fabricated in existing commercial foundries using standard lithography, and they run standard AI models out of the box without retraining, making adoption straightforward for existing software toolchains.
Target Audience
Primary customers are AI infrastructure providers, hyperscale data center operators, and enterprises running large-scale AI inference workloads that are constrained by power availability and energy costs.
Features
- Superconducting adiabatic logic gates that switch in picoseconds and operate at gigahertz clock rates while remaining energetically reversible
- End-to-end superconducting interconnect with zero resistance, eliminating the interconnect power that dominates conventional processors
- System-level efficiency of 10-100x more compute per watt, with cryogenic cooling included in the measurement
- Fabrication in existing commercial foundries with no exotic processes, using commodity cryogenics from mature industries
- Full-stack digital twin calibrated against fabricated superconducting circuits, demonstrating execution of real language model matrix arithmetic from PyTorch down to device-level Josephson junction dynamics
- Purpose-built for AI inference, the fastest-growing compute workload, with no model retraining required