Can quantum computing save generative AI?
The tools being built to automate the future are already unsustainable, but none of the Silicon Valley broskis seems to be talking about it.
Generative AI, for all its wondrous features (and it is, in many ways, wondrous), is a computationally expensive endeavour. Each interaction with a large language model is underpinned by immense backend activity, billions of parameters firing across dense neural networks, each weighted calculation demanding energy, time, and silicon.
The financial burden of training LLMs has escalated dramatically (training OpenAI’s GPT-4 was estimated to cost approximately $78.4 million). Projections suggest that the cost of training AI models could exceed $1 billion in the near future. The operational costs of running generative AI models are also substantial: estimates indicate that operating ChatGPT could cost over $700,000 per day, translating to approximately $21 million per month.
Where is the profitability? The monetisation paths that seemed obvious—subscription tiers, API licensing, enterprise integration—are already straining under the weight of their own limitations. Sure, interest has been considerable, but most sources suggest that paid subscription uptake is tepid (admittedly, it’s difficult to verify such numbers), and advertising integration (which has driven the Silicon Valley gold rush in recent decades) is technically fraught. In many instances, the value produced by the text, code, and images output by generative AI systems does not justify the cost of production. The hype curve is starting to plateau, competition is becoming increasingly fierce, and regulatory bodies are starting to flex their muscles. Already narrow margins will increasingly narrow.
It is particularly telling that frontier AI firms are falling back on familiar paradigms, most notably, productivity-suite tie-ins. OpenAI’s alignment with Microsoft, through Azure and Copilot, is emblematic of this trend: rather than generative AI disrupting existing software platforms, it is being retrofitted into familiar offerings. Google’s strategy is similar, embedding its Gemini model into Docs, Gmail, and Search. These are not signs of a revolutionary economic model, they are signs of absorption. Will OpenAI’s eventual legacy be the creation of a new feature in PowerPoint? Generative AI, for all its innovation, has yet to stand on its own feet: its monetisation is parasitic on legacy software architectures, which themselves are facing saturation.
Then there are the cultural and aesthetic applications of generative AI. Here, the irony is palpable: companies are, for profit-making purposes, building unprecedented compute apparatuses capable of producing compelling, often profound and sometimes beautiful, artefacts, yet they do not know how to extract capital from this process. Worse, these aesthetic outputs are now so cheap and abundant that their cultural value may be depreciating in real time. The infrastructure is operating, but it is producing excess without demand.
While venture capital continues to fuel the illusion of imminent return on investment, the actual cost of inference (the process of using these models to generate responses) at scale—particularly in real-time, high-availability contexts—is staggering. Each user query activates thousands of tensor operations, distributed across vast GPU arrays. These are not one-off costs, but persistent overheads that scale with user engagement. The more people use generative AI, the more expensive it becomes to maintain.
Unlike the general-purpose computing infrastructure of previous technological epochs, generative AI relies on single-purpose compute, optimised for matrix multiplication at scale. It relies on deeply specialised multi-billion dollar datacentres stocked with high-performance graphics processors. If the economic returns on this investment fail to materialise, such infrastructure might become this generation’s ghost town, built in anticipation of booms that never quite came—an edifice of compute power without alternative use cases.
But with business leaders and policymakers (and many cultural commentators) seemingly doubling down on generative AI and the large-scale infrastructure it requires, one can only assume they see something that many of us do not—is that something, quantum computing?
Its advocates argue that quantum systems could offer exponential speedups in certain computations, potentially reducing the energy and time costs associated with training and deploying large models. From this perspective, quantum computing appears as a deus ex machina for the economic woes of generative AI. Present computers are buckling under the weight of inference demands, but quantum machines will soon appear, ready to take the strain and save the visionaries.
The reality is more complicated. Quantum computers excel at highly specific tasks, like factorising large numbers, simulating quantum systems, and solving optimisation problems. They are not naturally suited to the kinds of probabilistic, high-dimensional computations that underpin generative AI. While there is promising research into quantum machine learning, we are far from seeing demonstrations of practical quantum advantage in this domain. It is possible that quantum hardware will eventually assist in training models more efficiently, or in optimising specific sub-processes, but these are incremental gains, not transformations of the kind that would make generative AI immediately sustainable.
But in this speculative economy of ours, future potential functions as present capital. Just as blockchain buoyed an entire ecosystem of dubious innovation with the allure of decentralisation, so too does quantum computing serve as a narrative scaffold, a way for firms to rationalise continued expenditure, to reassure investors, and to maintain momentum. Quantum is, in many ways, the perfect speculative asset: its complexity obscures its limitations, and its futurity immunises it against present critique.
What is most interesting about the quantum dream is how it speaks to deeper structural anxieties. The current model of generative AI is built on resource intensity—energy, data, compute—all of which are subject to ecological and political limits. Quantum computing, by contrast, promises efficiency. A fully operational quantum processor could perform certain calculations with a fraction of the energy required by classical supercomputers. In a world increasingly attuned to sustainability, this matters. It matters a lot more than many care to admit (even the billionaire tech bros have been preparing for societal and ecological collapse for a while now, buying islands and mega yachts; even building spaceships).
Training a single large language model can consume hundreds of megawatt-hours of electricity, with corresponding carbon emissions that rival small nations. This cost is not incidental, it is structural: the energy intensity of training and inference is embedded in the very architecture of these models. Quantum computing, with its promise of computational efficiency, is sometimes held up as a potential environmental corrective, but this too is speculative. In truth, the ecological cost of generative AI is already being incurred, and it is disproportionately borne by regions where raw materials are extracted and where data centres are sited for reasons of cost or regulatory laxity. We are outsourcing not only the labour of cognition but the labour of degradation. Maybe quantum can fix this, but even if it can, it may come too late.
Those of us who have been playing with computers for longer than we’d care to admit know that the promise is not the product. Quantum computing is still largely a laboratory endeavour. Scaling quantum systems is extraordinarily difficult, and the error correction required for useful computation remains an unresolved problem. In this light, to hinge the economic future of generative AI on the success of quantum computing is to swap one uncertain bet for another.
Geopolitics adds a further layer of complexity. Sovereign actors are investing heavily in both generative AI and quantum computing, not simply for commercial returns but as instruments of strategic advantage. The economic inefficiencies of these technologies may be tolerated, perhaps even encouraged, if they serve national interests. Infrastructure, in this view, is not evaluated by market logic alone but by its potential to shape global order. The quantum dream becomes not just a technological wager, but a geopolitical one.
There is a historical irony at play: generative AI was supposed to automate knowledge work, streamline productivity, and reduce the cost of creative and cognitive labour. Instead, it has triggered a new wave of infrastructural excess, one that mirrors the worst tendencies of extractive capitalism. We are told these systems will save us time, but we are spending unprecedented amounts of energy and capital just to keep them online. And so, in a strange inversion, quantum computing becomes less a step forward than a mechanism of temporal deferral, a way of delaying the moment when the balance sheets must be reconciled. The costs won’t ever disappear, they are simply being relocated to a more speculative future, where another miracle technology will, finally, change the maths. Until then, we build and burn.
Quantum computing, then, functions less as a saviour than as a symbol of faith, of futurity, of the enduring hope that the costs of innovation will eventually be justified by its rewards. In this sense, it is not the solution to the problem of generative AI, but a continuation of the present narrative.
Whether that faith is warranted remains to be seen. But we should at least ask: when the miracle fails to arrive, what will we have built in its name?