Quantum today is AI in 2017.

Where the stack eventually earns revenue from F500 companies and governments, which trickles it down through the layers of the stack. Advanced, but not commercially ready: the computers it relies on have been theorised, not manufactured.

Here we look for companies with a deep understanding of the problem as they are already providing solutions and have a roadmap for quantum integration.

Thin revenue from selling NISQ machine time. Squeezed from both sides: hyperscalers sell access per shot, differentiation migrates down into control and up into compilation.

Most investment, highest valuations, longest horizon. Billions spent making one chip bigger and cleaner.

Transistor density stays theoretical until readout and wiring stop being built one qubit at a time.

The layer that attacks today’s actual bottlenecks. Room-sized 1969 moon computer to the phone in your pocket. Every modality buys from here regardless of which architecture wins.

Few qualified sources, long qualification cycles, and demand that grows with every machine built. A programme can clear every technical milestone and still be rate-limited by a component with one foreign source.

$600k for a standard unit, several million for a 1,000-qubit-class system. Seven tonnes, a week to cool. Smaller systems or on-chip cooling are poised to slash the cost of these systems.

A heat and wiring problem. New approaches to control electronics may accelerate the development of multiple modalities.

Fast and noiseless entanglement over fiber across meters and kilometers would unlock quantum supercomputing and the quantum internet.

Thousands of stable beams have to reach thousands of atoms; today that means a room of optical tables. In neutral atoms the laser is the gate, so laser noise is the fidelity ceiling. Moving beam delivery onto photonic chips would shrink the largest and most expensive part of an atom machine and raise the ceiling at the same time.

Microwave to optical opens cross-modality computing. Potential to transform superconducting readout by connecting the qubit cavity to a functional electro-optic transducer, which would eliminate much of the modality’s scaling problems.

Room-temperature detectors are too noisy. The good ones are nanowires at 2 kelvin. A higher-temperature noiseless detector would reduce the limitation of photonics.

Nascent. If loading the data costs what the classical answer costs, there is no speedup.

Helium-3 comes almost exclusively from tritium decay in nuclear weapons stockpiles at about twenty thousand litres a year worldwide. Currently there is no substitute. Civilian demand is expected to grow from 1,200 to 16,000 liters/year in the next decade.

Quantum-grade diamond, enriched silicon-28, tantalum, niobium, sapphire, UV optics. Few qualified sources, long qualification cycles, and demand that grows with every machine built.

Anthropic, OpenAI, MSFT, GOOGL, META, PLTR, APP

Differentiation migrated down into hardware and up into applications. No basket.

NVDA, AMD, AVGO, MRVL

TSM, MU, SK Hynix, ANET, Coherent

ASML, AMAT, LRCX, KLAC, Advantest