What are the challenges of building utility-grade quantum computers?

As a cofounder of the Quantum Energy Initiative with an engineering background and perspective, I am looking at the challenges ahead to build utility-grade quantum computers, particularly in the FTQC regime. Understanding, evaluating and optimizing the energetics of these systems must be based on a holistic scientific and engineering perspective, up to looking at the economics of quantum computers. This complements a bottom-up approach that is focused on determining the fundamental bounds of qubits operations, belonging to the fields of quantum energetics and quantum thermodynamics.

In the second QEI Workshop held in Grenoble in January 2025, I had a chance to provide a view on current quantum computing roadmaps and their energetics aspects (video, slides).

In this talk, I inventoried some challenges and connected them to dozens of scientific and engineering questions:

• Finding quantum algorithms bringing actual practical speedups over their classical counterparts, and not just theoretical ones. And preferably in a generic way, meaning, for large classes of problems. This is a particularly difficult task for solving combinatorial and decision problems. You can forget about Grover’s algorithm which supports only some quadratic speedup, which is not sufficient in practice. See why here and there.
• Considering all classical software computing costs of any quantum solution, like all the circuit preparation (for a QPE-based chemical simulation in a FTQC regime) or classical optimizer preparation (for a variational VQE-based quantum simulation in an existing NISQ system), but also classical error syndrome detection (for FTQC) and large code compilation (for FTQC). By the way, all quantum algorithms are “hybrid”. Whether they are variational in NISQ or not in FTQC, they always have a significant classical part. I highlighted here the fact that quantum code compilation is a variable cost against the use case data. Given large circuits compilation and optimization may be costly, it may prevent quantum computing from being used on business applications with fast duty cycles. These costs, for a starter, can be expressed either in computing time or energy spent.
• Assessing the cost of QPU interconnect which will be necessary for all qubit modalities, even the ones which will squeeze the largest number of physical qubits in monolithic QPUs. In the case of Google, their plan is to create a 10K physical qubit chip. In order to support a “teraqops” chemical simulation (tera-quantum-operations), a single chip may support only a single logical qubit!
• Creating higher fidelity qubits at larger scale, which remains an overarching huge challenge. This has its load of engineering questions related to the cost of electronics, cabling, cryogenics, and lasers, which depend on the qubit type.
• Optimizing large algorithms computing times, with either speeding it up at the software/compiler level, or with having faster gates/readout/error correction, or at last, with parallelizing several circuit shots on multiple similar QPUs, provided they are affordable enough. I nicknamed this the EFP framework, for “efficient-faster-parallel”. Within this framework, I equated software to the total energy cost within the perspective of what to do to reduce computing time for solving a given problem. The total solution cost is proportional to energy costs, which can be optimized at the hardware level. You can then optimize computing time with three complementary means: reducing the total classical and quantum software cost, accelerating quantum gates, readout and error correction, and/or parallelize circuit shots. You can also reduce the number of circuit shots by reducing your outcome precision needs and by improving software.
• Defining some energy footprint acceptability threshold for large FTQC systems, with a reasoning that could be applicable to quantum computer overall economics and pricing. Based on current estimates, utility-grade QPUs supporting thousands of logical qubits may need a power ranging from less than a MW to several hundred MW. This power drain and economic question should soon drive many rationales, discussions and comparisons in the development of FTQC quantum computers.

I should add here various other scalability challenges like handling data preparation efficiently, particularly for solving linear equations and quantum machine learning tasks. Then, we are still in search of some potential qRAM, a memory type that is nearly mandatory to execute oracle-based algorithms.

Then, each qubit type (or “modality”) has its own scalability challenges. They can be very different. For example, solid-state qubits like superconducting and silicon-spin qubits are fighting against decoherence, noise and crosstalk whereas photonic qubits are hampered by low photonic sources and detector efficiencies and losses along the paths in photonic integrated circuits and optical fibers. These are very different scientific and engineering challenges.

What is the current TRL of quantum computers?

A TRL, aka, technology readiness level, is a metric from 1 to 9 defined by NASA in the mid-1970s. It characterizes the level of maturity of a given technology or product. 1 stands for an idea in the head of a researcher and 9 for a widely deployed technology like your smartphone or laptop computer. In between are physical experiments, prototype products and the first products being deployed. The TRL of a product can also be characterized by the readiness of its related ecosystem (training, skills, applications, documented deployed case studies).

According to this scale, quantum annealer’s TRL sits around 7 while gate-based systems are at around 5 given they are not yet commercially implementing quantum error correction. You can still buy a gate-based QPU with over a hundred physical qubit, like Cleveland Clinic who acquired a System One from IBM and installed it in Spring 2023, You can use them in the cloud through AWS, Microsoft, IBM or D-Wave online services. You can learn quantum programming and test your code. Even D-Wave has an (undocumented) SLA (service level agreement) should you use their online QPUs for production. There are already tens of thousands of quantum developers worldwide thanks to these systems being available online. You can test a 127-qubit IBM Eagle-generation QPU for free.

But the TRL is also currently quite low when considering the usefulness of quantum computers with regards to their added value and cost. Let’s say you test an online quantum computer with 20 operational noisy qubits. The system probably costs several $Ms and its hourly usage is charged at about $2K. But $2K is the price of a laptop where you could run your quantum code in a free open source emulator. And it would even run faster than on the quantum computer. For a higher number of qubits on a current noisy quantum computer, the situation is blurred depending on your use case as explained when responding to the previous question.

So, technically, quantum computers (QPU) TRL sits at around 5-8 depending on the case, but their actual value is still low compared to legacy classical systems.

Are NISQ computers useful for the industry?

NISQ computers are the current noisy qubit quantum computers. Those you can buy, rent or test today. The current wisdom is that, while they have some scientific value, their industry value is questionable. This is John Preskill’s view. I described NISQ challenges in a 2023 paper, and things have not significantly changed since then.

The main NISQ recent advances come from some improvements in qubit numbers and fidelities, with both superconducting and ions qubits, and from the use of quantum error mitigation, a technique promoted among others by IBM. Some use cases are presented, like with the interesting work from Algorithmiq, a software startup based in Finland.

A broad industry and academic team led by Sabrina Maniscalco, who is the CEO of Algorithmiq, published a paper willing to fight some myths on NISQ: Myths around quantum computation before full fault tolerance: What no-go theorems rule out and what they don’t by Zoltán Zimborás, Fernando G. S. L. Brandão, Elica Kyoseva, Ivano Tavernelli, Sabrina Maniscalco et al, arXiv, January 2025 (11 pages). It tried to show that the power and use cases of NISQ with quantum error mitigation could be expanded thanks to the upcoming improvements of qubit fidelities. However, the paper leaves the question open of whether NISQ systems will deliver some exponential or practical speedups. Another open question is lingering about the potential known and unknown limits of analog quantum computing.

What is it to be pessimistic or optimistic?

First, you need to understand where we are and to identify the scientific and engineering challenges ahead. Some are workable, others are really hard. It never necessarily means that “it’s impossible”. It is up to scientists and engineers to find solutions. You then need to appreciate how long it could take, understanding for example the lengths of the manufacturing-testing-experimenting cycles that dominate in this industry. These cycles are long!

Usually, the more you know a topic, the less optimistic you may become because you understand the sheer weight of the challenges ahead. It can lead to sheer pessimism. Listening to educated pessimists should be used to challenge the optimists. But you can also be excited by these challenges and by human creativity. That’s my situation.

I try to avoid flirting with optimism or pessimism, which are just two opposite moods. I stick to the scientific state of the art and remaining challenges. My own belief is that by doing so, you are neither pessimistic nor optimistic, you are just neutral with some optimism bias even though the pure optimists will perceive you as being a pessimist.

So, pessimism or skepticism? Source.

When you identify challenges in the quantum computing space, and there are many, most of them are currently being addressed in multiple ways by scientists and industry vendors. What can make you optimistic is that multiple technologies are being developed, like with the types of qubits or even with all the varieties of quantum algorithms around.

How to grade the related scientific vs technology uncertainty?

A common wisdom with many physicists, including the Nobel in physics laureates I met in Lindau in July 2024 is that the challenges ahead are mainly about engineering and technology developments. According to them, the theory is accepted and proven. But some scientists, like Alain Aspect, also think that creating large scale quantum computers may require significant time, a bit like the LIGO gravitational waves detection experiment which took about 25 years from design to completion (from 1989 to 2015).

I do not entirely agree with that focus on engineering. While for sure, engineering and technology developments are inescapable routes to building fault-tolerant quantum computers, there are still some scientific uncertainties to unfold. A bit like the whole history of physics, experimenting new artificial settings will uncover new mysteries which may require building new theories, a bit like Zeeman’s effect discovered in 1896 could be explained only about 30 years later with electron spins.

Here, we don’t know yet what we will discover when creating large-scale physical qubit systems. Will they showcase high-fidelity entanglement states at scale? How will we characterize it? Will we reach the famous Heisenberg cut, when macroscopic systems stop behaving quantumly with coherence and behave classically without it? Will we face the effects of the Lieb-Robinson bound that conditions the speed of correlation propagation in large scale entangled systems? It may affect large QPUs, particularly those which are implemented using quantum interconnect techniques. The confusion may come from the difference between theory and science in general. Quantum physics theory may be somewhat settled, but fundamental research and experimental research is still an open field.

Another huge quantum computing scientific uncertainty lies with algorithms. We don’t have that many quantum algorithms and even fewer quantum algorithms with provable exponential speedups, or practical speedups with reasonable times. There is still a lot of fundamental research to do there. It is likewise with quantum error correction. How about new sources of correlated errors that we’ll discover with scalable systems and for which current error correction codes are not well adapted? In the technology realm, we don’t know yet if the improvement of chips manufacturing processes will yield better quality qubits. We need to test it. At last, how about various “unknown unknowns”?

I’d say quantum computing is becoming a broad scientific experimental endeavor, mixing theoretical and experimental advances, one helping the other. We need these experiments to validate the theory and/or to extend it. We also need to accept failures and try various technology avenues. That’s why it doesn’t yet make much sense to launch a Manhattan like project on a specific qubit modality. It explains for example why DARPA is funding about 11 different such avenues as part of its Underexplored Systems for Utility-Scale Quantum Computing (US2QC) and Quantum Benchmark Initiative programs.

What is the cost of trying these various avenues? Is it acceptable? It is currently mainly supported by a mix of public and private investments with a high proportion of the former. It will go on like this if the strategic stakes are perceived as being as high as they are today.

A bird’s eye view on the realizations (green), challenges (orange/red), variations (grey) and paths to scalability (blue) for seven qubit types: superconducting, silicon, topological, cavities, cold atoms and trapped ions. They are all different! Source: Understanding Quantum Technologies 2024.

Is the pace of progress of quantum computing accelerating?

The perception of acceleration comes from industry vendor hyperbolic communication. Recent news pushed by large companies like Google and Microsoft are driving this impression. Microsoft oversold its topological qubit, that are supposedly around the corner. They are not! Too many scientific or industry communications are about “getting closer to scalable quantum computing” (example 1, example 2, example 3).

I don’t believe there is some acceleration. There is progress for sure, but no acceleration. First and foremost, how do you measure acceleration and speed of progress? What is the comparable? What is the meter and unit?

One way to proceed is to compare what scientific and technological advances have been achieved vs past plans. When I started working in the quantum computing space in 2018, I tracked the following announcements: Rigetti was planning to release a 128 qubit QPU in 2019, IonQ was mentioning that it had controlled 160 qubits on which it could handle operations on 79 qubits, and Intel was showcasing a silicon chip with the capacity to host 1,500 qubits in a chip thinner than a human hair. In 2019, Google had a 53 qubit chip (Sycamore) and IBM, a 65 qubit chip (Hummingbird r1).

Where are we now, 6 to 7 years later? Rigetti’s best chip is Ankaa-3 with 84 qubits. IonQ is at 36 qubits with its Forte system while Intel has a vague 12-qubit chip named Tunnel Falls released in 2023, that is not integrated in a full QPU system. Google reached 105 qubits with its Willow chip announced between August and December 2024 and IBM has a 156-qubit chip. Meanwhile, these qubits fidelities have somewhat improved but not in a stellar way. IBM’s 433 Osprey and 1,121 Condor qubit chips which were released in late 2022 and 2023 were even discontinued in favor of smaller chips (Heron 156 qubits) due to their lack of tunable couplers. With these couplers, IBM’s Heron is faring well, at about 99.7% two-qubit gate fidelities, not far from what is currently achieved with trapped-ions (99.84% at Quantinuum with its H2-1 system).

Still, all this doesn’t look like being exponential. It proved more challenging to increase the number of qubits of decent quality at intermediate scales of about a hundred physical qubits. You can see the trend in the updated scatter plot below. The blue dots are about superconducting qubits and the green ones with trapped ions. Cold atoms in orange seem to scale well but have a hard time exceeding two-qubit gate fidelities of 99.5%. Ions achieve that, faring at >99.8% but are still struggling with scaling. Superconducting qubits are in-between, with 99.7% two-qubit gate fidelities at between 105 and 156 physical qubits.

In other qubit modalities, progress looked faster, like with cold atoms, first on large settings like Pasqal with the control of 828 atoms in 2024, and Caltech with 6,100 atoms, and then in gate-based mode (QuEra, starting in December 2023 with 48 basic logical qubits, Atom Computing with 24 logical qubits in November 2024).

Photonic qubits are either locked at low qubit count in the non-scalable KLM paradigm (Quandela, Orca Computing, …) or are bound to implement FTQC scalable architecture in the more or less distant future (again with Quandela, with PsiQuantum, and Xanadu).

Since 2018, we’ve also seen the emergence of new breeds of qubits implementing some autonomous error correction, like the cat-qubits from Alice&Bob and AWS, or the GKP qubits from Nord Quantique. These are encouraging new technology developments that may reduce the physical per logical qubits number overhead.

There are several technical limitations to the # of qubits per monolithic quantum processors. It leads their designers to envision using various quantum interconnect solutions (microwaves, optical photons, transduction), which will bring a wealth of new challenges, from physics to quantum code compiling and optimization. We thus have to be very careful when extrapolating physical qubit numbers, particularly toward million qubit ranges.

As I described in my Moore’s paper in 2023, scaling QPUs won’t probably follow any Moore’s law like with CMOS semiconductors because the scalability challenges are highly multidimensional. It’s not just a matter of shrinking transistors. I simplify things here since advances in CMOS were more complicated than that. Also, the test & learn cycles with various quantum computing technologies are relatively slow and don’t follow any Moore’s law.

However, we may underappreciate some interesting technology developments happening in the background that will pay off later: on cryogenic control electronics, on high-density cabling, control signals multiplexing, high-power and low phase noise lasers, and with high-power cryogenics.

What is special about quantum computing with regard to predictions?

In the not-so-distant past, predictions abounded about how fast a technology revolution would be adopted. The cloud, mobile computing, Internet of things, and more recently the metaverse. Artificial intelligence made its comeback in two periods, one between 2012 and 2022 with machine learning and deep learning around data science concepts, and since 2022, with the advent of LLM-based chatbots.

In these domains, the technology was more or less mature, but it had proven use cases that were beyond the state of the art. For example, mobile computing and cloud computing brought usability and easy access to multi-format communication on the go. In other domains like with IoT and metaverses, the generic user value was neither obvious nor universal.

The uncertainties in these various domains were however not so much about the underlying science and technology, than on other factors: costs, ease of use, user value, physiological acceptability in the case of the metaverse, social impact and so on.

With quantum computing, we are exposed to a flurry of market predictions including how much value it could generate, around $2T as soon as 2035 according to McKinsey before even the technology is ready. Even if it was ready in 2025, it would take more than 10 years to yield such value. You’d have to deploy it, to create algorithms, to run it, to try different options to create a digital twin of a new drug or chemical compound candidate. Then, you’d have to test it in the real world, either undertaking multi-year clinical tests or launching the buildout of new manufacturing plants. And so on. This takes time!

McKinsey quantum monitor as of April 2024. Notice the $42B total government investment which is wrong due to bad estimated of China’s investments as well explained in The Quantum Panic – The U.S. wants to be prepared for whatever quantum technologies bring, but is it time to rightsize the threat from China? by Rachel Cheung, The Wire China, February 2025.

How can we compare today’s situation and uncertainties with quantum computing to the ones we had 30 years ago or so in other domains?

There were not many doubts back then about the progress ahead with raw computing power, memory, storage and networking capabilities growth. Moore’s law was kind of predictable, at least until the mid-2000s.

Telecom infrastructures were key for developing the Internet and for smartphones and the cloud. There were some infrastructure built out delays for fiber installation around the world. However, this was not due to uncertainty in scientific or technology. It was more an economic issue. Also, many Internet developments benefited from the technology but were accelerated for societal reasons, like with the advent of the web 2.0 around 2004. There are some parallels to draw here with FTQC QPUs. At some point, they will become heavyweight infrastructure programs requiring significant funding, beyond classical research funding in the academic and startup world.

In our analysis, we are often affected by a “survival bias”. We use the example of successes as proof of our reasoning, forgetting past failures. They don’t resonate in memories since they didn’t affect us much. In the mid-1980s, there was some buzz on GaAs electronics, that was supposed to supplant CMOS. It didn’t, mostly for engineering and economic reasons. Nanomaterials and graphene technologies had similar troubles, ranging from scientific and technology overpromises and societal backlash.

Classical analog computing was promoted in the 1960s to 1970s but failed, being overcome by digital computing. It was not generic enough and didn’t scale well due to noise. Quantum computers are also analog to some extent, of course, with analog quantum systems ala Pasqal and D-Wave, but also with NISQ systems implementing arbitrary rotation gates. One key difference is that FTQC mixes analog and discrete operation, which is to enable quantum error correction, a feat that was not accessible to classical analog computers. Also, analog computing didn’t benefit from the same attention and public investments as we have right now with quantum computing. There was no equivalent of Shor’s algorithm to drive government interest.

By the way, where are we with the promises of the Hyperloop, promoted by a certain Elon Musk in a 2013 white paper? Several ambitious infrastructure projects were supposed to be launched around the world. A couple startups were created. Twelve years later, it’s not really in the radar. The idea is plagued with  a mix of engineering problems, safety problems, capacity and economic limitations and huge infrastructure costs. How about drones flying around to transport passengers from airports to cities? They can indeed fly. But with limited range, air trafic control challenges, economic challenges again, and so on.

Now, let’s consider the Apollo program. It took only 8 years from JFK’s speech in May 1961 to Congress to having men landing on the Moon in July 1969. Yuri Gagarine had been sent to space a couple weeks earlier, in April. JFK confirmed the goal in another speech in September 1962, listing some technical challenges. Weeks from the start of the Cuban missiles crisis! Of course, the Moon quest project was fueled by huge investments but it was mostly about solving technology challenges. In 1961, engineers already knew how to launch a rocket and send a human into space. Newton’s laws were already vindicated. But there was the Cold War between the USA-USSR rivalry.

Quantum computing seems much different and more challenging than all the above examples from the scientific and engineering perspective. There are superposed stacks of uncertainties from securing large scale entanglement to designing useful algorithm, and with relatively long test and learn cycles. And we have not yet started the real economical discussion about the price of these systems and their return on investment. From the scientific and engineering standpoint, quantum computing seems closer to nuclear fusion. Both are linked to the limits of physics and experimentation. The advantage of quantum computing is that it is way less capex intensive than nuclear fusion. At least at this point.

Will quantum computers run LLMs more efficiently than classical computers?

Some quantum computing companies like IonQ and Quantinuum are flirting with the idea that quantum computers may someday run LLMs more efficiently than classical computers. They could even be more energy efficient.

What investors got from some IonQ presentation in 2024. Source.

Unfortunately, quantum computers will probably not help mitigate the growth of energy consumption linked to consumer LLM based-generative AI. Data loading is quantum computers’ Achilles’ heel and will remain so for a long time given they will unlikely benefit from some hardware Moore’s law in relation to their clock speed. Given the size of both the training data (in petabytes and tera-tokens encoded in highly multidimensional vectors) and models (in tera-parameters), they won’t be loaded at a reasonable timescale on any quantum computer for training, even with some classical data ingestion and preparation. Variable computing and energetic costs are mostly in inferences which are run nearly real-time. Any quantum algorithms running in a quantum advantage regime will probably not be real-time. We can also use some simple logic here. An inference would require one quantum computer and there are hundreds of millions of users of these LLM chatbots. Unless some clever algorithm enables multiple inferences to run on single QPUs, this can’t scale well.

There are no available resource estimates on how some FTQC quantum computers could handle any LLM task. The path to quantum transformers by David Wakeham, Xanadu, April 2024, provides sketchy classical and quantum computing requirements to train LLMs. It shows that even with very optimistic assumptions on the clock speed of quantum computers (in GHz), and with several tens of thousands of logical qubits, quantum computers would be faster than classical computers for the embedding part of LLM only after a couple decades, but this task could be run classically and only kernel operations handled quantumly (linear operations using block-encode matrices). Quantum linear algebra is all you need for Transformer architectures by Naixu Guo, Zhan Yu, Matthew Choi, Aman Agrawal, Kouhei Nakaji, Alán Aspuru-Guzik, and Patrick Rebentrost, arXiv, February-May 2024 (31 pages) states that “embedding large data into quantum computers is difficult in the absence of the availability of functioning quantum RAMs”.

So maybe, we should invent the notion of qSLM for quantum small language models!

However, the broader field of quantum artificial intelligence, not including LLMs for consumer chatbots, is a good exploration path, particularly when there is a need to use smaller data for training and to obtain better accuracy in inferences. Quantum machine learning can potentially be used for both training and inferences. There are still many technology interdependencies here: on data preparation and on the development of qRAM and on the connection between classical and quantum components of an AI solution.

One thing is sure: classical AI already helps develop quantum technologies, in quantum error mitigation, quantum error correction at the error syndrome detection level, qubit calibration, and even circuit design (using custom Transformers).

How about the if/when mantra?

Analysts and others are frequently parroting the “if not when” mantra, saying that the question is not if scalable quantum computers will appear but when. As shown above, there are still many “ifs” which explain a lot of the uncertainty on the “when”. Otherwise, we would have an infinite certainty on the “when”. It is a bit like if “if” and “when” were complementary observables in Niels Bohr’s parlance. And due to Heisenberg indeterminacy principle, you can’t be precise on both observables. So if you dare want to be a bit more precise on the “when”, you need to be less accurate on the “if”.

Should I wait, or should I go?

That’s the typical question for end-user companies. Some analysts are playing with FUD (fear, uncertainty and doubt) to drive customers on the quantum computing bandwagon: human skills and quantum computing resources will be scarce, there will be a significant first-mover advantage, and your competitiveness will highly depend on your adoption of quantum technologies.

I don’t believe any of these assertions to be true or verifiable. If quantum computers work at scale and benefit customers, hardware vendors will manufacture them and sell them. In the history of technology, even for large systems like supercomputers, there’s never been a real scarcity problem. And if there was one, it would mean that scalable quantum computers are way too large and costly. Then, human skills can follow-on. We already have a good stream of young people being trained. Finally, your competitiveness may only marginally depend on quantum computing.

Some early adopters of quantum computing in Europe across multiple verticals: energy utilities, chemicals, telecoms, financial services, air/defense, automotive and transportation. Source: OE.

Still, I would advise customers to invest in quantum computing. Am I contradicting myself? No, I’m just basing my rationale on different arguments. One is that testing quantum computing is not expensive and is part of any techno-screening of any large company. You don’t need to acquire a $40M QPU. You can rent it for cheap in the cloud. You can even learn quantum coding with free open source emulators.

Second, learning quantum computing is an excellent way to revisit your existing applications portfolio. It helps you identify pain points in scalability or output quality, revisiting the algorithms you are using. At some point, it will help you modernize your IT, and not necessarily with quantum computing. You may land in using GPUs and/or tensor network “quantum inspired” classical applications.

Learning about quantum computing can help you upgrade your algorithms teams. It will push them to develop better mathematical skills, like in the financial services vertical. It will also help customers deal with the aggressive push from industry vendors and to decipher the continuous news stream. In just one week, you’ll have thousand of talks from quantum physicists at the APS March Meeting held in Anaheim, and the Nvidia Quantum Computing Day in San Francisco, hosting a dozen quantum computing vendors. Get ready for a flurry of announcements!

Of course, another minimal task to launch is the audit and upgrade of the company cybersecurity infrastructure. Whether you like it or not, you’ll have to fold in deploying post-quantum cryptography.

So, learn quantum computing at small scale, educate your own technical people, try small things, be ready to adopt the technology should quantum computing scale well on a reasonable time horizon.