The tech industry was not ready for what came after ChatGPT. Power grids, cooling systems and data centers are being rebuilt right now because none of them were designed for the demand that arrived.
UN researchers expect AI to double the power and water that data centers use by 2030.
Had the industry known that demand was coming, it would have designed for it.
Quantum computing is early enough to do exactly that.
CEO and co-founder of QuantWare.
AI hyperscalers care about one thing: how much useful computation they get for each dollar invested.
This total cost of ownership is dominated by depreciation and power usage. In the current infrastructure buildout, they have sufficient access to capital to fund the capital expenditure but insufficient access to power.
This constraint is so acute that data centers in space are becoming an economic possibility if growth continues at this pace. But even if the AI industry matures and growth slows, total cost of ownership will dominate.
This directly translates into capital requirements and compute-per-watt being the most important metrics.
The quantum industry should care less about qubit counts and more about compute-per-watt
However, quantum computers are still sold on qubit count, and qubit count alone says nothing about economic returns on a system. What will matter at scale is how much compute-per-watt the end user gets — and this is not driven by qubit counts alone.
Take superconducting qubit processors, for instance, which are one of the quantum computing platforms most suitable for computation. They have been stuck on around 100 qubits for almost a decade. This is because on today’s chips, more than 90% of the surface is taken up by the wiring that controls the qubits and reads their answers, rather than by the qubits themselves.
The usual way around the problem is to network many small processors together. However, networking is lossy and gives sparse connections. That’s bad in classical chips, and exponentially bad in quantum chips. The work of holding the system together grows faster than the qubits it adds, and more of the power goes into running the machine than into computing with it.
This gives you higher qubit counts, but those are not equivalent to systems built with less networking: it costs you compute-per-watt.
The quantum industry should care more about experience curves
An experience curve is a simple idea: every time the total number made doubles, the cost of each one falls by a fixed amount. It is what drove down the price of solar panels and batteries, and quantum will not be an exception, but only if we build at the volumes that let the curve work.
Optimistic quantum roadmaps never talk about price reductions, when it is one of the most important problems our industry needs to solve. On today’s price per qubit, a million-qubit machine would cost somewhere between 100 billion and a trillion dollars. The cost per qubit has to fall at least a hundredfold in the coming years for quantum computers to be economically viable, which is doable, but not solved by more laboratory proof of concepts.
The transistor is the clearest case in computing itself. A single transistor once cost around a dollar. Today a chip carries billions of them, and each one costs a fraction of a cent. That fall came from decades of making more of them, driving down the cost year after year.
Costs fall the way they fell in classical computing: volume manufacturing of standardized parts combined with compounding (but not stepwise) technical progress. This requires an open architecture in which specialist companies each build one layer of the machine.
This Quantum Open Architecture — processors from one company, cryogenics from another, control electronics from a third — is what will drive specialization and in turn, drive the required economies of scale.
With the field transitioning from science to engineering, there are plenty of cool one-off demos in startups today; there are very few companies actually building the supply chain.
It matters to do this now
Quantum computers will change the world. They have the promise to change the way we do drug discovery, material development, and machine learning.
This also ties back to AI. The same workloads straining the grid today are the kind of heavy computation a quantum machine could one day take on, at a fraction of the power, once the economics are there. Getting there as fast as possible will create a massive amount of value and make all of our lives better.
The technology is progressing. But we have not focused on the economic metrics that matter, and if we don’t do so soon, we risk building an industry that makes expensive demos rather than actually useful computers. Qubit counts alone don’t matter. Building supply chains takes years, if not decades.
Whether the industrial fabrication, standardized parts and processors with exponentially more compute-per-watt will be there to achieve the economies of scale will be determined by choices made in the next few years.
Best business computers: leading desktop PCs organizations.
This article was produced as part of TechRadar Pro Perspectives, our channel to feature the best and brightest minds in the technology industry today.
The views expressed here are those of the author and are not necessarily those of TechRadarPro or Future plc. If you are interested in contributing find out more here: https://www.techradar.com/pro/perspectives-how-to-submit