Every few months there’s a new headline about quantum computers getting more qubits. More qubits, more power, and quantum AI is just around the corner. Or so it seems.
But the more I read, the more I think we’re watching the wrong number. The biggest blocker to quantum AI isn’t the number of qubits. It’s that there’s no efficient way to get our data into a quantum computer in the first place.
The input problem#
Machine learning runs on data: images, text, numbers, millions of them. That data is classical. It lives in ordinary memory as ordinary bits.
To use it in a quantum algorithm, it first has to be encoded into a quantum state. On paper this looks amazing. With amplitude encoding, 2ⁿ numbers fit into just n qubits, so a million values fit in about 20 qubits.
The catch is that preparing that state isn’t free. Someone still has to feed in every single one of those numbers.
Where the speedup goes#
A recent paper, The Input Problem: A Permanent Bottleneck for Quantum Machine Learning, puts actual numbers on this. Loading just 256 values (8 qubits) took 247 CNOT gates with an optimized library, and the count doubles with every qubit you add. On top of that, the classical computer has to read the entire input just to work out how to load it.
So the quantum algorithm may be lightning fast, but the loading step eats up the speedup before the algorithm even starts. It’s like building one of the fastest engines in the world and connecting it to a tiny straw as the fuel line.
Is QRAM the answer?#
The usual proposed fix is QRAM, quantum random-access memory, which would let a quantum computer query classical data efficiently. Without scalable QRAM, these algorithms keep hitting the same data bottleneck.
The paper goes further, though. It argues this is a basic counting limit, not an engineering problem better hardware will fix. If you have N numbers, someone has to touch all N of them. Its advice is to focus on problems where the data is already quantum, such as simulating molecules or working with output from quantum devices, instead of forcing classical data through the straw.
Why it matters#
I think this is the question to watch in quantum AI, more than qubit counts. If someone finds a practical way around the input problem, it could change computing in a big way, and the current race to build bigger data centers and more GPUs could start to look very different.
Until then, quantum AI will probably shine on quantum-native problems first, and GPUs aren’t going anywhere just yet.
Do you think the input problem can be solved, or will quantum AI always be limited to problems that start out quantum?