Quantum Academy is now open. Enroll in live online training – see the schedule.
Skip to content
Browse courses

Quantum-Inspired, Quantum-Enhanced, Quantum-Powered: Telling the Labels Apart

Market Reality

Quantum Academy7 min read

Quantum-inspired, quantum-enhanced and quantum-powered describe three different kinds of product. Quantum-inspired software is classical: it runs on ordinary CPUs and GPUs, using algorithms from quantum research. A quantum-enhanced product is classical apart from one quantum component, such as a quantum random number generator or a quantum sensor. A quantum-powered product, which makes the strongest of the three claims, depends on a quantum processor for at least one step of its results. Three questions sort them: what hardware runs it, what’s the classical baseline, and what happens without the quantum part?

The quantum threat to public-key cryptography is real. NIST’s draft transition plan, NIST IR 8547 (November 2024), states that RSA and elliptic-curve signature algorithms “are vulnerable to Shor’s Algorithm on a cryptographically relevant quantum computer”, meaning one large enough to break real keys. It schedules quantum-vulnerable algorithms to be disallowed after 2035. None of the labels says whether a product addresses that threat.

Quantum-inspired algorithms

A quantum-inspired algorithm is a classical algorithm that borrows a technique or an insight from quantum research. Two families have results you can check.

Tensor networks

A tensor is an array of numbers with any number of indices; a matrix is a tensor with two. Physicists write the state of many interacting quantum particles, which would otherwise need an impossibly large array, as a network of small, linked tensors. A tensor network can describe the low-energy states of real materials efficiently. Those states belong to “a tiny, exponentially small, corner” of all possible states.

They work on ordinary data too. In 2015, researchers rewrote an image-recognition network’s fully connected layers as tensor networks, compressing one layer’s weight matrix up to 200,000 times and the whole network up to 7 times. In 2022, a team used tensor networks to simulate turbulent flow in a computational space more than ten times smaller than direct numerical simulation needs.

Researchers have also used tensor networks to challenge the classical cost estimates behind early quantum supremacy claims. In a 2022 paper, one team generated a million approximate samples from the 53-qubit circuit of a 2019 supremacy experiment in about 15 hours on 512 GPUs.

Dequantized algorithms

In 2016, Iordanis Kerenidis and Anupam Prakash published a quantum algorithm for recommendation systems, the software behind product suggestions. It was considered one of the strongest candidates for an exponential quantum speedup in machine learning. In 2018, Ewin Tang published a classical algorithm that was only polynomially slower, given the same fast access to the data that the quantum algorithm assumed, so the exponential gap disappeared. Building a classical counterpart this way is called dequantization. In 2019, Tang and five co-authors extended it to principal component analysis, support vector machines and regression.

Dequantized algorithms are useful when the data suits them. A 2020 benchmarking study found they perform well only on very large data with a simple, well-behaved structure: “low rank, low condition number, and very large dimension of the input matrix”. They also set the baseline a quantum machine learning claim has to beat.

For a buyer, a quantum-inspired product is classical software. Benchmark it on your own data against the best classical method for the same problem.

Quantum-enhanced products

Two kinds of quantum component show the pattern.

Random number generators

A quantum random number generator (QRNG) measures a quantum process, such as which way a single photon goes at a beam splitter, to produce random bits. In a product, the QRNG is the entropy source, which supplies the unpredictable input for key generation.

Under its Entropy Source Validation program, NIST certifies entropy sources against SP 800-90B and records each one’s noise source, the process that produces its randomness, in one of two categories. A physical noise source uses dedicated hardware; a non-physical one uses system data or human input. There’s no quantum category. Swap in a validated classical entropy source and the product generates keys the same way, with the same algorithms.

So a device that agrees keys with elliptic-curve Diffie–Hellman (ECDH) is as exposed to Shor’s algorithm with a QRNG as without one. Protection comes from post-quantum algorithms such as ML-KEM (Kyber), standardized in FIPS 203.

Quantum sensors

A 2022 US strategy defines quantum sensors as devices that use quantum properties, “such as atomic energy levels”, for measurement. GPS already depends on atomic clocks, one of the strategy’s examples. In a navigation unit with an atomic clock inside, the baseline is the part the clock replaces, such as a quartz oscillator. Ask how far the unit’s time or position drifts when satellite signals are lost, with and without the clock.

Physicists use “quantum-enhanced” more narrowly. In Quantum-Enhanced Advanced LIGO Detectors, the LIGO team reported in 2019 that squeezed light, a quantum state with less noise in one property, raised its two detectors’ expected detection rates by 40% and 50%. That’s a named component, a baseline and a measured difference.

Quantum-powered products

A quantum-powered product depends on a quantum processing unit (QPU), the chip that contains the qubits, for at least one step of its results. Ask which processor, with its technology and number of physical qubits, the hardware count before any error correction. Ask which step it runs, and what the product returns when the processor is unavailable. Variational quantum algorithms “use a classical optimizer to train a parametrized quantum circuit”. The QPU executes that circuit, and classical computers run the optimizer and everything else.

A 20-asset example

Suppose a datasheet says a QPU picks the best subset of 20 assets for a portfolio. Each asset is a yes-or-no decision encoded on one qubit, so there are 2²⁰ = 1,048,576 possible portfolios, few enough for an ordinary computer to check every one.

The circuit can be simulated exactly too. A state-vector simulator stores the full state of n qubits as 2ⁿ complex numbers, at 16 bytes each in double precision:

  • 20 qubits: about 16 megabytes
  • 30 qubits: about 16 gigabytes, within a workstation’s memory
  • 45 qubits: about half a petabyte, which took 8,192 supercomputer nodes in a 2017 simulation

Qubit count alone doesn’t settle it: researchers have simulated larger circuits with tensor networks, such as the 53-qubit sampling above. So ask for the classical baseline in writing: the same workflow with a classical solver or a simulator in place of the QPU, and both sets of results. If the results don’t change, the quantum step isn’t contributing to what you’re buying.

Quantum-inspired encryption

“Quantum-inspired encryption” has no definition in NIST’s cryptographic standards, and it describes neither of the two established approaches. Post-quantum cryptography is classical mathematics designed to resist quantum attack, such as ML-KEM. Quantum key distribution (QKD) establishes keys using the physical states of photons. A cipher described as quantum-inspired runs on classical hardware, so the evidence for its security has to come from public cryptanalysis, as it does for AES.

In research papers, the nearest meaning is a classical cipher built around a simulated quantum process. A 2015 image-encryption scheme generated its keys from a quantum walk simulated on a laptop, and its authors assessed it with statistical tests. NIST’s test suite for random number generators warns that “statistical testing cannot serve as a substitute for cryptanalysis”.

So ask for the algorithm’s name. If it’s ML-KEM or AES, the label is inaccurate but the product can be assessed. If it’s proprietary, ask what public cryptanalysis it has had; A Field Guide to Misleading Quantum Marketing covers the questions that follow.

Three questions for any label

Question Quantum-inspired Quantum-enhanced Quantum-powered
What hardware runs it? CPUs and GPUs Classical hardware plus one quantum component A QPU for at least one step, classical computers for the rest
What’s the classical baseline? The best classical method for the same problem The classical part the quantum component replaces The same workflow with a classical solver or simulator in place of the QPU
What happens without the quantum part? Nothing, because there isn’t one The product still works, at the classical part’s performance The results should change, and the vendor should show by how much

If you assess technical claims in pitch decks or earnings calls, Quantum Technology for Investors covers the maturity of each technology, supply chain dependencies and how to evaluate those claims. For products sold as quantum AI, Evaluating Quantum AI and Machine Learning covers current reality versus marketing.

Before your next vendor meeting, add the table’s three questions to your list and ask for written answers that name the processor or component. For a quantum-powered claim, also ask for one run with the quantum step removed.