by Christian Dickel
Quantum computing and nuclear fusion are potential 21st century technologies based on 20th century physics and neither of them is currently market ready. But while they are sometimes bunched together as fascinating concepts that will at any time be twenty years away from being realized, some estimate the timescale for the commercialization of the quantum computer to be much shorter now. Quantum computing is currently in a hype phase: The company D-wave has already sold a few quantum annealers based on flux qubits for millions of euros. They can solve certain optimization problems, but their computational advantages are a topic of debate. Google, IBM, Intel, and Microsoft are major commercial companies investing in quantum technologies right now. Several startups such as Rigetti Computing and IonQ have been founded recently with the goal of commercializing quantum computing. A list of such companies can be found here.
First, let me stress that none of these near-term quantum computers are aimed at private customers. They either require expensive cryogenic technology or ultra-high vacuum chambers or both. This means they are about as affordable and mobile as the first classical computers. But one important difference between the time of the first classical computers and now is that the internet is fully developed to allow remote access and data transfer. In fact, it was partly build to distribute computing power in the early days. Thus, a cloud quantum computer seems like a natural business model for anyone trying to commercialize these technologies.
Cloud quantum computers are here
Naturally, as a researcher in this field, I would like to understand the cloud quantum computer business model. The first such systems are running and users on IBM’s Quantum Experience can currently use the quantum computer for free for teaching and research. Rigetti Computing also offers access to its quantum computer on the cloud as well as quantum simulations on classical computers. This means that they not only have the quantum and classical hardware ready but also programmed interfaces to allow users to run their own algorithms on their machine.
Right now, these machines are not powerful enough to beat classical simulations of qubits. In fact, a few qubits can even be simulated on a laptop. But the exponential scaling of simulating quantum systems means that the best classical supercomputers can currently only simulate around 50 qubits. Simulating 100 qubits on classical hardware will at best be possible in ~100 years if Moore’s law would hold. Soon, quantum processors might solve problems that cannot be simulated classically. However, that doesn’t mean that those problems will be useful. It is not clear when quantum computers will be able to perform the first computations that would generate commercial value. Currently, the market for the quantum computer would be research groups who want to try out ideas on real hardware or they could be used for teaching purposes in universities. Thus, the market would first be limited to the quantum information community, with early adopters like Volkswagen or Airbus mostly driven by the fear to miss the moment when quantum computation starts to be useful.
How much does a medium-scale superconducting quantum computer cost?
I’m going to focus on a superconducting system because I understand this platform better than ion traps or a spin qubit system. A budget for such a system would look quite different. It is not yet clear which quantum computing platform will scale to a commercially viable quantum computer or if any of them will.
How expensive would access to a superconducting quantum computer on the cloud have to be to cover the expenses of the provider? I will try to estimate the cost excluding the years of research and development into quantum and classical hardware and the upkeep of the fabrication facilities involved. These costs can only be recovered if the quantum computer can solve commercially interesting problems. However, then, it would likely be easy to make a profit. Until then, if you have the goal to make a quantum computer anyway and could just produce some extra chips with qubits and want to offer cloud access on a system that is an exact copy of a research quantum computer, how much do you have to charge to finance the service?
To be clear, I am not claiming that there is currently a big demand for quantum computer hours. In addition, as long as quantum computer access is offered for free by some competitors, it will be difficult to sell quantum computer hours. The tech-giants with other sources of income can keep the start-up competition at bay, because for them the advertising value might be more important than small profits on the side. As long as quantum computations do not solve commercially interesting problems, the market will be narrow. Once they approach computationally interesting problems, companies might have an incentive to try ideas on quantum machines to gauge their power. Charting the near-term developments in quantum computing by tracking the performance of different cloud services could be a strategy for investors to predict when quantum computers will have a true breakthrough and on which quantum computing platform it is going to happen. This knowledge would certainly be valuable.
The price for cold temperatures
Superconducting qubits require mili-Kelvin temperatures. The cryogenic hardware to achieve ~10mK is a dilution refrigerator. Probably, it makes sense to pick a bigger model that could potentially host many-qubit chips. Worldwide there are only a handful of companies building dilution refrigerators: Oxford Instruments, Bluefors, Janis, Leiden Cryogenics, CryoConcept. A big dilution refrigerator will cost about ~1 M€. A dilution refrigerator consumes about 10 kW of electricity and requires cooling water and access to liquid nitrogen for traps that filter the He3-He4 mixture that is at the heart of the cooling process that can reach ~10 mK. Thus a dilution refrigerator requires at least~10-20 k€ per year for upkeep.
The dilution refrigerator would have to be wired for 50-100 qubits, which requires cables, attenuators, filters, isolators and amplifiers. A recent publication on this subject illustrates the design complexity. Let’s assume the design for the wiring is already done: we just have to pay for the components and for putting it all together. Wiring of dilution refrigerators is often done in-house in research groups and takes ~2 Months of work for two people. In addition, the quantum chip needs radiation shields and magnetic shields, which are often custom made. We also make infrared filters in house, which are used on all the microwave cables going into the dilution refrigerator. The quantum limited amplifiers used are currently not yet commercially available. Thus, giving an exact price for all components is difficult, but the wiring will cost will likely fall in the range of 200-500 k€.
Control electronics cost
Now consider the room-temperature electronics required for qubit control and readout. In superconducting qubits, we mostly use microwave equipment that has been perfected for radar and telecommunication applications. One would need arbitrary waveform generators and microwave sources for qubit control as well as fast analog to digital converters for readout. Often, there are also amplifiers at room temperature. Each readout building block costs about 60 k€ and can currently be used for ~10 qubits. The control hardware for each qubit costs about 15 k€. Thus room-temperature electronics will likely cost about 20-30 k€ per qubit at current rates. A lot of in-house developments of electronics are likely necessary to reach state-of-the art performance when feedback on qubit measurements is necessary. Cables and components at room temperature probably come in at 20-40 k€. The price of electricity for the room-temperature electronics is likely on the order of 10-20 k€ a year. The computer that controls all the electronics would be a high-end desktop computer whose price is negligible compared to everything else. Bringing the total cost of room temperature electronics to about 1.5M€.
Software development costs
How much it would cost to develop an interface that allows people to use the quantum computer via the internet? From an input output perspective this is not too difficult: a quantum computer could generate up to a few gigabytes of data if it runs for an hour, while the algorithms that clients would like to run can likely be represented in a few kilobytes. But, converting an abstract quantum algorithm into the specific hardware instructions that execute it in some close-to optimal way on the provided hardware is non-trivial. Either the clients would have to think about optimizations themselves or the provider could charge for consulting on this process. An intuitive interface could likely be written by a few good programmers in a year, if the drivers for the quantum computer hardware and the calibration routines for the qubits are already largely written. Thus we can maybe estimate the price for the interface at ~0.5 M€, including some development during a beta testing phase.
Despite having a nicely automated software to run our quantum computer experiments, we would likely still need to set aside someone with quantum computer experience to keep the cloud quantum computer operational. Whether one full-time position would be enough is an interesting question. It depends on how much support clients may need and how hands-off the calibrations are and if human intervention is required for scheduling and making judgement calls.
The bottom line
To conclude this very rough budget, one would likely need to invest at least ~3 M€ on hardware and another ~0.5 M€ on software to provide a cloud quantum computing service. A minimum for operational costs would likely come in at ~0.2 M€ a year just to be able to provide the service. The software investment is a one-time investment, while the several million euros for hardware will likely have to be invested for every additional quantum computer.
Now how much time on the quantum computer could we sell and how much time has to be used for calibration procedures? The amount of time that can be sold will have a large impact on the price one would have to charge clients per hours. If one could sell half the hours in five years, one would have to charge ~200 € per hour to earn back a 4-5 M€ investment. The price could be pushed down with economies of scale on the software and the infrastructure side, but currently the hardware investment dominates the price. The hardware price will depend on architecture decisions leading to compromises between performance and price. Current prototype quantum computers largely focus on performance, the commercialization would likely lead to cost becoming a more important figure of merit. In order to make quantum computing cheaper, a big opportunity to cut costs lies in developing cheaper solutions for the control electronics and for the wiring. This might lead to interesting new ideas in the field.
The price per hour on a superconducting cloud quantum computer is currently hypothetical and the budget is crudely estimated. A real world implementation will likely exceed this idealized budget. That being said, I was surprised by how cheap an hour of quantum computation on a medium-scale quantum computer would be. Compared to the price for supercomputer time to perform 50-qubit simulations, the quantum computer on the cloud could be competitive. Within a few years such cloud computers could venture into territory where classical computers cannot follow. Usually the quantum advantage is framed in terms of problems quantum computers can solve that classical computers cannot, but they might first beat classical supercomputers in terms of the cost to run quantum simulations. What fidelity simulations on near-term quantum computers could reach will naturally also be a factor. Of course, I neglect the dominant R&D costs, but the take-away message is that the quantum computing “space race” does not require investments comparable to the real space race, CERN or nuclear fusion reactors.
About Christian Dickel
Chris came to the Netherlands for the food and the weather but stayed for the quantum computer work. Apart from work he enjoys playing music with friends, ranting and soap-boxing.