Start with the journey, not the machine

An individual may fly from Schiphol. Irregular equipment may travel by car by road and ferry or through the Eurotunnel. Large volumes of standardised goods may favor shipping. If material needs to flow continuously, perhaps the right solution starts to resemble a pipeline. Moreover, getting to an airport, harbor or terminal can add an extra leg to the journey and a cost of reloading and repackaging of your cargo. It’s not about which single vehicle is fastest, but at what journey duration will a given vehicle create economical value.

Computing works the same way. Central processing units (CPUs) provide car’s flexibility. Graphics processing units (GPUs) provide cargo ship’s enormous throughput but only when work can be organised appropriately. Dedicated accelerators can be spectacularly effective for other tasks. The economically optimal system is therefore usually hybrid: different parts of the workload travel by different computational modes. Quantum computing adds other possible modes, but for what tasks?

In principle, from the point of difficulty there are three forms of computational problems that businesses run into. While blindfolded, finding a needle in a haystack (you stop when you find one), finding the best needle in the stack of needles (you keep looking, cause you never know if you found the best), and finding a needle by following instructions to its known location. We have built an entire civilization on solving problems in the latter category, and making every effort to avoid the other two.

Optimisation is often the needle in the needle-stack type problem. Over decades, businesses have adapted to the limitations of available computing. Difficult problems were simplified, approximated or reformulated until they were reduced to the category of finding the rough location of the best needle, not the needle itself. This created an approximation debt. To close it, businesses turned to acquiring more data to make the problems easier, or customized pre and post-processing or parameter tuning were needed, often with a human in the loop, hence resisting automation. Eventually the workaround became the normal workflow, and the value sacrificed along the way was largely forgotten.

Machine learning is a good example. Many learning problems contain difficult optimization at their core, but we have become very good at reformulating them into versions that run efficiently on CPUs and GPUs. At the Quantum Applications Lab, we are finding cases where returning to the original, harder formulation can offer significant advantages. The opportunity is not to make today’s approximation faster. It may be to make the problem we originally wanted to solve affordable again.

Using a quantum device also introduces a handover problem. Somewhat paradoxically, this means that much of the engineering required to make quantum computing useful consists of making the classical parts faster. Better preprocessing, decomposition and orchestration allow the specialist machine to work only on the part of the problem where it creates disproportionate value.

This is also why application development eventually has to become hardware-specific.

Remaining hardware-agnostic is useful during an initial technology scan because it avoids premature commitment. But once the goal is to create value, abstraction has limits. Different quantum machines expose different connectivity, interactions, precision and control. A next-generation processor may not simply run today’s algorithm faster; the algorithm and workflow may need to change to exploit the new hardware.
Let’s return to our 400-kilometer journey.

What if the new transport technology does not take us from Amsterdam to London, but 400 kilometers vertically, into orbit? The question changes. What is the business case for going there?

Satellite communications provide one clear answer. Much of space exploration, however, remains curiosity-driven: we go because the environment allows experiments and discoveries that cannot easily be made on Earth. That is a legitimate and important mode of innovation, but it is difficult for businesses to prepare today for industries whose eventual applications we cannot yet describe.

Long-term, fault-tolerant quantum computing is somewhat like that journey into space.

Near-term quantum optimization is different. It is closer to aviation. We already understand the potential business case, but the journey has to be long enough: the optimization problem large and valuable enough, and the surrounding workflow efficient enough, for the specialized mode to beat the alternatives. Sometimes the train will still win.
For business leaders, regardless of how quantum hardware develops, a robust strategy might be relatively straight forward:

Identify where computational limitations have quietly removed value from your business. Quantify what recovering that value would mean. Then determine which hybrid of computational modes can economically recover it.