Xanadu and Rolls-Royce cut the estimated cost of loading data into a quantum computer, on paper
Xanadu and Rolls-Royce built a compiler that loads structured data into a quantum computer far more cheaply than a generic method. For a 20 qubit fluid flow problem, their estimate of the whole workflow falls from about 2 × 1020 T gates to about 9 × 1015. Nothing ran on a quantum computer, and 9 × 1015 is still far beyond any machine that exists.
What the paper claims
Many promised quantum speedups quietly assume the input data is already sitting inside the machine. Getting it there can cost so much that the speedup disappears. This paper, which passed peer review and was published in Quantum Science and Technology on October 26, 2026 1, tackles that loading step for engineering problems where the data has structure, such as smooth flow fields.
This is a theory and resource estimate paper. The authors count gates for an idealized error corrected machine. There was no hardware run.
Where the savings come from
The trick is to stop treating data as an arbitrary list of numbers. Smooth functions can be compressed into a chain of small matrices called a matrix product state. The authors show that keeping each link in that chain to 32 values holds the loading error under one in ten thousand for states from 11 to 20 qubits, and for one 11 qubit state at a looser one in a thousand, two values per link are enough.
Cost is counted in T gatesThe expensive operation in error corrected quantum computers, so counting T gates is a standard way to price an algorithm.. For one benchmark state in their tables, their method needs about 6,240, against about 111,000 for a standard general purpose method and about 11,500 for a lookup table approach.
The whole workflow
The headline case is a fluid dynamics problem on 20 qubits, which means solving a system with a matrix of about a million by a million entries. To invert that matrix, the algorithm builds a polynomial of degree 100 million, so the loading step is called about a hundred million times in every run. That call count is set by the math of the inversion and is the same in both versions. What changes is the price of each call.
In the authors’ generic approach, which assumes nothing about the data, one call of the loading step costs about 90 million T gates. Their structure-aware version costs about 300,000, roughly 300 times less. Over a hundred million calls, one run drops from about 9 × 1015 T gates to about 30 trillion.
The rest of the gain comes from reading the answer out. Pulling a solution out of a quantum state takes many repeated runs. Using a different final transform, their estimate cuts the required runs from about 20,000 shots to about 300 shots, roughly 70 times fewer. Multiply it through and the whole workflow falls from about 2 × 1020 T gates to about 9 × 1015. The authors put the savings at “a factor of over ~104 compared to an exact approach.”
So the 9 × 1015 that remains is mostly still the loading step, now cheap per call but repeated a hundred million times in each of about 300 runs.
About 300 times cheaper per call, times about 70 times fewer runs, still leaves about 9 × 1015 T gates. Resource estimate only, against the authors' own generic baseline; nothing ran on hardware. Redrawn from Alonso-Linaje et al., Quantum Science and Technology (2026), Fig. 7.
What this result does not say
The comparison point is their own generic baseline, not the best classical solver. A 20 qubit problem is something an ordinary computer handles easily, so this is not a quantum advantage claim. And 9 × 1015 T gates would still need a large, fully error corrected machine running for a long time. No such machine exists, and error correction is far from solved.
What I take from it
For anyone following quantum computing as an investment, I read this as a cost line moving, not a new capability. A step that used to dominate the bill for engineering simulations now looks like something careful compilation can shrink by orders of magnitude, and that kind of saving carries over to any future machine. It does not bring a useful fluid simulation any closer to running today, because the remaining bill is still far beyond any machine that exists, and it says nothing about beating the best classical solvers. The paper also shows that the price of a workload depends heavily on software choices, so a headline gate count is only as good as the baseline it is measured against. I would treat Rolls-Royce’s involvement as a sign of genuine industrial interest rather than as a commercial deal. The next thing I would look for is the same accounting at problem sizes classical solvers find hard, set against the best classical method.
