The umbrella category. Hardware where the physics of the substrate, not arithmetic on a digital processor, does the computation. The user encodes a problem into the substrate’s energy landscape, applies an input, and reads the answer out after the substrate has relaxed (or sampled). No iteration in software. No CUDA-equivalent. The chip is the algorithm.
Substrate-class taxonomy
Four families sit under this umbrella. Each implements energy-minimisation (or distribution-sampling) in a different physical medium.
Adjacent substrate-classes (different physics, similar architectural inversion):
| Family | Dynamics | KB concept |
|---|---|---|
| Memristive analog matmul | In-memory matrix multiplication via conductance crossbar | Memristors |
| Reversible / adiabatic | Computation without energy dissipation (Landauer limit) | Reversible Computing |
| Photonic analog | Optical interference for matmul or Ising | (in Analog Computing) |
| Quantum annealing | Tunneling through energy barriers in superconducting qubits | (in quantum-computing/ folder) |
What unites them
What distinguishes them from digital AI accelerators
Cerebras, Groq, H100 still iterate arithmetically. They run the same algorithm faster. Physics-native compute runs a different algorithm — physics does the work. Different category. The architectural inversion is the moat (or the failure mode, depending on whether the calibration loop dominates the speed budget).
Pre-seed and seed deals in this category are still rare enough that:
- Substrate-class first-mover gets a category-defining position
- The mature-foundry path (90nm, 130nm, 65nm, GF 22FDX) is uncrowded — see Mature Foundry Positioning
- The compiler-build risk is structural across the category; teams that solve it first win disproportionately
- Strategic acquirers exist at each substrate-class transition (Lightmatter for analog photonic, NTT for CIM, Cerebras / Groq for AI inference primitive, Cadence / Synopsys for EDA-side optimisation)