Hey!
For years, the most common and frankly most reasonable critique of quantum computing has been: "When will it do something useful on real-world problems?"
Not quantum simulation. Not breaking encryption. Something that actually matters for the world we live in, the one dominated by classical data, classical machine learning, and classical AI.
This week, a team from Caltech, Google Quantum AI, MIT, and Oratomic published the most convincing answer yet!
They proved that a quantum computer with fewer than 60 logical qubits can perform machine learning on massive classical datasets using ten thousand to a million times less memory than any classical machine.
Tested on real data. Movie review sentiment analysis. Single-cell RNA sequencing for cell type identification. Not a toy problem or a contrived oracle designed to make quantum look good.
Why this is not just a storage trick
The quantum computer doesn't just store data better. It runs the entire machine learning pipeline.
The biggest unsolved problem in quantum ML has been data loading. Quantum algorithms get their power from superposition, processing many possibilities simultaneously. But classical data arrives one sample at a time. Movie reviews come one by one. Gene expression profiles come one by one. How do you get a million data points into quantum superposition when the world only gives them to you sequentially?
This paper solves it with an algorithm called quantum oracle sketching. It works by streaming. Each classical data sample flows into the quantum processor, applies a small carefully designed quantum rotation, and gets discarded. Gone. Never stored. As more samples flow through, these tiny rotations accumulate coherently. After processing all the data, the quantum state has built up what's called a quantum oracle, a compressed representation of the entire dataset encoded in the amplitudes and phases of a handful of qubits.
Here's why this is hard and why nobody pulled it off before.
The intuition says feeding noisy classical data into a quantum system should destroy coherence. All that randomness, all that noise, it should cause decoherence and blow up the computation. And with naive approaches, it does. The error compounds quadratically.
The clever part: the rotations are designed so that contributions from different data values act on orthogonal subspaces of the Hilbert space. They accumulate coherently instead of destructively interfering. The error ends up scaling linearly (N/M) rather than quadratically (N squared over M). That difference is everything. It's what makes the whole approach feasible.
Once the oracle is built, quantum linear algebra takes over. Algorithms based on quantum singular value transformation run classification or principal component analysis directly on the quantum state. Then a measurement protocol called interferometric classical shadows extracts a compact classical model that can make predictions on new data independently of the quantum computer.
The full cycle: data streams in, quantum rotations build the oracle, quantum ML runs on the oracle, classical predictions come out. No dataset storage at any point.
What this means for AI
This is where it gets practical. Classical AI is hitting the memory wall. Training large models consumes enormous energy, and a huge fraction of that goes to storing and moving data. The Large Hadron Collider generates petabytes of data per hour but researchers discard all but roughly one in a hundred thousand collision events because storage physically cannot keep up. The same bottleneck appears in genomics, astronomy, climate modeling, and every other field where data grows faster than storage.
A quantum ML pipeline that processes streaming data without storing it doesn't offer a marginal improvement. It opens analyses that were physically impossible before because you couldn't afford to keep enough data.
The Caveat
This is a theoretical proof validated through numerical simulations. It has not been demonstrated on actual quantum hardware. That matters and you should know it.
But the qubit requirements are what make this feel different from most theory papers. Fewer than 60 logical qubits is in the range that near-term error-corrected machines are targeting. IBM, Atom Computing, and many others have all laid out roadmaps that hit this qubit count within the next few years. Given the pace of progress in high-rate quantum error correction codes, an experimental demonstration feels like a question of when, not if.
Whether you've been a skeptic or a believer, this work changes the conversation. The evidence for useful quantum advantage is finally here.
Until next time,
References
Paper: arxiv.org/abs/2604.07639