The sobriety of a quantum computer depends on its architectural choices, not in the number of qubits displayed. It is the real efficiency of the calculation which must guide the judgment of future announcements.
Not all quantum computers are created equal on energy. The energy frugality of a quantum computer is not a promise that we discard after the fact. It lies in the physics and engineering choices made years earlier, when no one was still looking at the meter. At a time when the projections anticipate a doubling of the electricity consumption of data centers by 2030, the question of sobriety of calculation is no longer theoretical. Those who made these choices early will not have to make up for them.
We often hear that quantum computers are, by nature, energy-intensive. We also hear, sometimes in the same sentence, that it would be the promise of a finally sober calculation. Both statements are false for the same reason. The quantum computer does not have an energetic nature. Its consumption is the product of architectural choices and engineering decisions made well in advance. Two machines of comparable size can differ today by a factor of three on their power consumed. At the scale of fault-tolerant systems, the roadmaps suggest differences of up to two orders of magnitude between architectures.
For a CTO or an architect who one day wants to integrate these machines, the gap is not anecdotal. Understanding where it comes from means understanding what is really at stake when you design one of these machines. Four factors weigh in, and they are linked.
Operating temperature
This is the first position, by far. On most platforms, the energy spent to cool the system exceeds that used to compute. For example, a superconducting processor must be kept between 15 and 20 millikelvins, or near absolute zero. It takes approximately 1 kW of external power to generate 1 µW of cold at 20 mK : it’s like burning the electricity consumption of a washing machine to prevent a snowflake from melting. Operating a notch higher lightens the entire cryogenic chain.
Other approaches, for example a silicon quantum processor, operate at a few hundred millikelvins, more than an order of magnitude above superconducting systems. The cryogenic station is not a common inevitability. This is a consequence of the platform chosen.
The material of the qubit and its quality
Then comes the intrinsic quality of the qubit, that is to say its capacity to retain its information without losing it, which is measured by its coherence time. The more coherent and stable a qubit is, the fewer must be added to obtain a reliable calculation. This point goes beyond pure performance. It directly controls consumption, since each added physical qubit requires its control and reading electronics, which has a concrete energy cost.
For example, a cryo-CMOS controller (control electronics designed to operate at very low temperatures), whether spin qubits or superconductors, dissipates on the order of 1 to 10 milliwatts per qubit. Taken in isolation, this is negligible. Multiplied by the number of qubits in a scaled system, this is a thermal load that ends up sizing the entire cooling infrastructure.
To give a reference: a latest generation GPU server, like the NVIDIA Vera Rubin NVL72 rack, already consumes between 180 and 220 kW. A scaled superconducting quantum processor would run in parallel in the same infrastructure, adding several times that power, just for cryogenics. The material from which the qubit is made therefore determines part of the energy budget from the design stage.
Quantum error correction
This is the silent multiplier. Current qubits are noisy. To obtain a reliable logical qubit, its information is encoded redundantly over a large number of physical qubits, which makes it possible to detect and correct errors during the calculation. This correction only works if the physical quality of the qubits is below a certain error threshold, the fault tolerance threshold. The further we are below this threshold, the fewer physical qubits are required per logical qubit. This ratio, the overhead, can be considerable, and it is directly governed by the quality mentioned above.
The effect is cumulative. Less overhead means fewer physical qubits, therefore less control electronics, less wiring between the cryogenic stages, and a reduced thermal load to cool. Quality upstream is paid for, or saved, several times downstream. This is one of the rare examples in engineering where the same choice is beneficial on three or four different levels.
Integration density
What remains is the way in which it all fits into space. Housing a qubit is not housing a point. It means housing the volume necessary for its connections and its geometric constraints. A dense architecture reduces wiring lengths, limits heat input, and allows the same cryogenic system to serve more qubits. A dispersed architecture does the opposite, and the bill follows. On a large scale, when a system has tens of thousands of qubits, this wiring becomes one of the main points of tension in cryogenic platforms: each added link is one more route through which heat infiltrates. On the scale of a system with 100,000 physical qubits, the cumulative length of these links can represent several hundred meters.
Read upcoming announcements
These four factors are not independent. The quality of the qubit controls the overhead, which controls the number of components, which controls the thermal load, which controls the cooling infrastructure. A good decision upstream spreads favorably throughout the chain. A bad one increases the bill on each floor.
This is why an isolated consumption figure says nothing useful. The sector will increase its energy announcements in the coming months. To evaluate them, a few questions are enough. At what temperature does the machine operate, and what does this temperature mean for the infrastructure? Does the figure given include cryogenics and control electronics, or only the computing core? Is consumption independent of the number of qubits, or does it increase with it? And above all, is efficiency reduced to the calculation actually produced, and not just to the number of qubits displayed? The number of qubits is an indicator of communication, not a measure of power, and even less a measure of sobriety.
These questions guided our architectural choices from the start. If the answers are not always simple, the questions are essential.
