AI and Quantum Computing: History, Science and the Road Ahead

A classical computer, whatever its speed, ultimately processes information as bits — each definitively a 0 or a 1 at any given moment. Every operation a classical computer performs, from running a spreadsheet to training a neural network with billions of parameters, is built from logic gates manipulating these definite bit values. This model has proven extraordinarily powerful: the transistor, the integrated circuit, and decades of relentless miniaturization following the pattern popularly known as Moore’s Law have made classical computing exponentially cheaper and faster for seven decades, and modern GPUs — built from billions of transistors performing massively parallel classical arithmetic — are what actually train and run every major AI system in production today, including the largest language models.

A quantum computer processes information differently, using qubits that can exist in superposition and become entangled with one another. This is not simply “more states at once” in a way that trivially multiplies computational power — the mathematics of quantum mechanics means a quantum computer’s advantage, where one exists, comes from interference effects that can amplify correct answers and cancel out wrong ones for specific, carefully designed algorithms. This is why quantum computers are not a general speedup for arbitrary computation: for the vast majority of tasks — including most of what AI systems do — no known quantum algorithm outperforms the best classical approach, and researchers do not expect that to change. Quantum advantage, where it exists or is expected, applies to specific problem classes: certain optimization problems, quantum chemistry and materials simulation, and specific number-theoretic problems including integer factorization.

This distinction matters enormously for how this guide — and any credible source — discusses AI and quantum computing together. They are not on a collision course to merge into one technology, and quantum computers are not “next-generation AI chips.” They are two different tools, developed by overlapping but distinct research communities, that may eventually be used together in hybrid systems for specific tasks — a classical AI system handling the bulk of a workload, calling out to a quantum coprocessor for a narrow subroutine well-suited to quantum hardware — without either replacing the other.

It is easy to credit AI’s recent progress entirely to clever algorithms, but the historical record shows hardware evolution was at least as important. The transistor (1947) enabled the integrated circuit, first demonstrated separately by Jack Kilby and Robert Noyce in 1958-1959, which let engineers place multiple transistors on a single chip. This began the decades-long pattern of exponentially increasing transistor density per chip that became known as Moore’s Law, after Intel co-founder Gordon Moore’s 1965 observation that transistor counts on integrated circuits were roughly doubling every year (later revised to about every two years). This relentless hardware scaling made possible not just faster classical computers generally, but eventually the specific hardware — graphics processing units — that would prove unexpectedly well-suited to AI.

GPUs were originally designed in the 1990s and 2000s for a narrower purpose: rendering 3D graphics for video games, a task requiring the same mathematical operation — matrix multiplication — performed millions of times in parallel across a screen’s pixels. Researchers including Andrew Ng and others noticed in the late 2000s that this same parallel matrix-multiplication capability was exactly what neural network training required, since a network’s forward and backward passes are, at their mathematical core, large sequences of matrix multiplications. NVIDIA’s 2006 release of CUDA, a programming platform letting developers use GPUs for general-purpose computation beyond graphics, made this connection practically exploitable. AlexNet’s 2012 ImageNet victory (covered in this guide’s timeline) was trained on consumer NVIDIA GPUs, a detail often glossed over in accounts that credit the result purely to algorithmic insight — the hardware match was just as decisive.

This hardware dependency has only deepened since. Training today’s largest language models requires thousands of specialized AI accelerator chips — NVIDIA’s data-center GPUs and increasingly custom chips like Google’s Tensor Processing Units (TPUs), first deployed internally in 2015 — running for weeks at a time, consuming enormous amounts of electricity and capital. This is the concrete, measurable reason AI infrastructure spending has become a major line item for cloud providers and technology companies, and it is entirely a classical computing phenomenon: every dollar of this spending buys classical silicon, not quantum hardware. Cloud AI — the ability to rent GPU compute by the hour from providers like AWS, Google Cloud and Microsoft Azure rather than building physical data centers — has been essential to democratizing access to this hardware, letting startups and researchers without capital for their own GPU clusters still train and deploy serious AI models.

Superconducting qubits, the approach Google and IBM have invested in most heavily, use tiny circuits made from superconducting materials, cooled to temperatures colder than deep space (near absolute zero, using specialized dilution refrigerators) so that electrical resistance vanishes and quantum effects can be precisely controlled and measured. This approach currently leads in raw qubit counts and gate speed — operations can be performed in nanoseconds — but the extreme cooling requirement makes the systems large, expensive, and difficult to scale into the eventual multi-million-qubit systems some cryptographic threat estimates assume would be needed. Google’s Sycamore and Willow chips, and IBM’s Osprey, Condor and subsequent processors, are all superconducting systems.

Trapped-ion qubits, pursued by companies including IonQ and Quantinuum, use individual charged atoms (ions) suspended in electromagnetic fields and manipulated with precisely tuned lasers. This approach generally achieves higher gate fidelity — fewer errors per operation — than superconducting systems, because individual atoms are naturally identical to one another in a way manufactured superconducting circuits are not. The tradeoff is speed: laser-based operations are typically slower than electrical ones, and connecting large numbers of trapped ions into a single scalable system remains an active engineering challenge, though companies in this space have published roadmaps targeting continued scaling.

Neutral atom qubits, pursued by companies including QuEra and Pasqal, use individual uncharged atoms held in place by tightly focused laser beams (optical tweezers) arranged in programmable grid patterns. This is a comparatively newer approach that researchers consider promising specifically for scaling to larger qubit counts, since optical tweezer arrays can in principle be reconfigured and expanded more flexibly than the fixed physical wiring superconducting chips require. It remains earlier-stage than superconducting approaches in terms of published error-correction demonstrations at comparable scale.

Photonic quantum computing, pursued by companies including PsiQuantum and Xanadu, uses individual particles of light (photons) as qubits, manipulated using specialized optical components. Its major advantage is that it does not require the extreme cryogenic cooling superconducting systems need, and photons naturally travel well through optical fiber, potentially easing the networking of multiple quantum processors together. It faces its own distinct engineering challenges around reliably generating and detecting single photons with high efficiency.

No single one of these four approaches has been declared the industry-standard winner, and researchers across the field generally expect this diversity of hardware approaches to continue for years — a useful reminder, when evaluating any single company’s quantum computing claims, that the underlying hardware question remains genuinely open.

Training a modern large language model begins with assembling a training dataset — typically hundreds of billions to trillions of words drawn from books, websites, code repositories and other text sources, filtered and processed to improve quality. The model itself starts as a neural network with randomly initialized parameters — for the largest current models, hundreds of billions of individual numerical weights. Training proceeds in small steps: the model is shown a chunk of text, asked to predict the next word, compared against the actual next word, and its weights are adjusted slightly via backpropagation to reduce that specific error. This step is repeated an enormous number of times — typically trillions of individual weight updates — across a cluster of thousands of GPUs working in parallel, a process that can take weeks and cost tens of millions of dollars in compute for the largest models, according to published estimates from AI labs and independent analysts.

This is followed, for most deployed models, by additional training stages: instruction tuning, where the model is further trained on examples of following instructions well, and reinforcement learning from human feedback (RLHF) or related techniques, where human reviewers rate model outputs and the model is adjusted to produce more highly rated responses. This multi-stage process is why a “raw” trained language model and the assistant-style chatbot product built from it can behave quite differently — the additional stages shape tone, helpfulness and safety behavior on top of the base model’s learned language capability.

Inference — actually using a trained model to answer a question or generate text — is computationally far cheaper per request than training, but at the scale of millions of daily users, inference compute costs can exceed training costs over a model’s lifetime. This is why AI companies invest heavily not just in training infrastructure but in inference-optimized hardware and software, including techniques like model quantization (reducing the numerical precision of a model’s weights to speed up computation with minimal accuracy loss) and specialized inference chips. None of this — training or inference, for any commercially deployed model as of this guide’s writing — involves quantum hardware at any stage.

AI ethics research addresses a specific, documented set of concerns: bias in training data producing unfair or discriminatory model outputs, misinformation risks from generative models producing convincing but false content, privacy implications of training on large-scale scraped data, environmental costs from the electricity large training runs consume, and labor-market effects as AI tools automate specific tasks. Major AI labs — Google DeepMind, OpenAI, Anthropic and others — publish research and policy documents addressing these concerns, along with technical safety research on topics including model alignment (ensuring a model’s behavior matches its developers’ and users’ intentions) and robustness against misuse. This guide treats these as genuine, actively researched, imperfectly solved problems, not as either dismissible non-issues or as evidence of an unmanageable crisis — the accurate picture, per the published research itself, sits between those two extremes.

Responsible quantum innovation is a narrower, more concretely defined conversation at this stage of the field’s development, centered substantially on the cryptographic transition this guide has returned to repeatedly. Because Shor’s algorithm’s threat to public-key cryptography is mathematically proven even though the hardware to exploit it does not yet exist, the field’s most active “responsible innovation” work is coordinated, proactive: NIST’s standards process, vendor migration roadmaps, and international standards coordination aimed at completing the transition to post-quantum cryptography before it becomes urgent rather than after. Quantum computing researchers and companies have generally avoided the more speculative ethical debates surrounding AI — there is no equivalent “quantum misinformation” or “quantum bias” concern with comparable evidence behind it — precisely because the technology remains further from broad deployment.

Both fields share one governance challenge in common: the gap between a technology’s proven scientific capability and the public’s understanding of that capability, a gap this guide has aimed to narrow by consistently distinguishing what has been proven, announced, deployed, or merely projected. Readers evaluating any specific claim about either field — in the news, in a company announcement, or in casual conversation — are well served by asking the same question this guide asks of every claim it makes: is this proven science, peer-reviewed research, an official company announcement, a commercially available product, or a future projection? The answer changes how much weight the claim deserves.

AI investment today is overwhelmingly near-term and revenue-focused: cloud providers building GPU data-center capacity to meet immediate customer demand, enterprises adopting AI tools for measurable productivity gains, and AI-focused companies raising capital against demonstrated or clearly near-term commercial traction. This is reflected in publicly reported capital expenditure figures from major cloud providers, which have devoted tens of billions of dollars annually to AI infrastructure buildout in recent years, an investment pattern grounded in already-proven commercial demand rather than speculative future capability.

Quantum computing investment follows a different, longer-horizon logic. Technology companies including Google, IBM and Microsoft fund quantum research programs as part of long-term strategic bets on a technology whose commercial payoff, per their own published roadmaps, is generally not expected for years. Venture capital investment in quantum computing startups has grown substantially over the past decade, but investors and company executives alike generally describe this as patient, research-stage capital rather than investment expecting near-term revenue at the scale AI companies are already generating. Government funding plays a proportionally larger role in quantum computing than in mainstream AI development today, reflecting both the technology’s earlier stage and its national-security relevance through the cryptographic transition this guide has discussed throughout.

This difference in investment horizon and risk profile is precisely why this guide’s Business Insight box, earlier in this article, distinguishes between organizations investing in AI for measurable near-term returns and those simultaneously funding quantum research as a longer-term strategic hedge — both are rational strategies, but conflating their timelines and expected returns leads to poor decision-making on either front.

What Is a Qubit?

A classical bit is always either 0 or 1. A qubit, the basic unit of quantum information, can exist in a superposition of both states at once, described mathematically by two probability amplitudes. This does not mean a qubit “is both 0 and 1 simultaneously” in an intuitively simple sense — it means that until measured, the qubit’s state is a combination the mathematics of quantum mechanics precisely describes, and measurement collapses it to a definite 0 or 1 with specific probabilities. Multiple qubits can also become entangled, meaning their combined state cannot be described independently of one another, a resource quantum algorithms exploit for computational advantage. Physically, qubits are built from superconducting circuits cooled near absolute zero, individual trapped ions controlled by lasers, neutral atoms held in laser traps, or photons — each approach with different tradeoffs in speed, stability and scalability, none yet a settled industry standard.

How AI Models Learn

A modern AI model, such as a large language model, learns through a process called training: it is shown enormous amounts of data (text, images, or other structured information) and adjusts millions or billions of internal parameters to get better at predicting or generating correct outputs. The core mechanism is backpropagation (1986, per this guide’s timeline): the model makes a prediction, compares it to the correct answer, and works backward through its layers using calculus to determine exactly how much each internal parameter contributed to the error, then nudges each parameter slightly to reduce that error. Repeated over enormous datasets and computing resources — almost always classical GPUs — this process is what produces a model capable of translating language, generating images, or predicting protein structures. No quantum hardware is involved in training any major commercial AI model available today.

Why Quantum Computers Need Error Correction

Physical qubits are extremely fragile: interactions with their environment — stray electromagnetic fields, tiny temperature fluctuations, imperfect control signals — cause them to lose their quantum state, a process called decoherence, typically within microseconds to milliseconds depending on the hardware. This makes any individual physical qubit far too unreliable to run a long, useful computation on its own. Quantum error correction addresses this by combining many physical qubits into a single, more reliable “logical qubit,” using redundancy and clever encoding to detect and correct errors without directly measuring (and thereby destroying) the fragile quantum information itself. Google’s Willow chip result (December 2024) demonstrated that, for the first time in a peer-reviewed, published result, error rates in a logical qubit actually decreased as more physical qubits were added — a “below threshold” result researchers had targeted for years, because without it, adding more qubits would only add more opportunities for error, not less.

What Is Quantum Advantage?

Quantum advantage refers to a demonstrated case where a quantum computer solves some specific computational problem faster than the best known classical method — a narrower, more careful term than “quantum supremacy,” which researchers increasingly avoid because it can be misread as claiming general superiority. Google’s 2019 Sycamore demonstration and its 2024 Willow follow-up are examples of quantum advantage on specific, often deliberately chosen benchmark tasks — genuine scientific achievements, but not evidence of practical, broad quantum superiority over classical computers on real-world problems businesses and researchers actually need solved. No quantum computer has yet demonstrated quantum advantage on a practically useful problem at a scale that outperforms the best classical approaches, which is why companies pursuing this goal describe it as a future roadmap target, not a current capability.

How AI and Quantum Computing Work Together

Today, the relationship runs mostly in one direction: AI helps quantum computing, more than quantum computing helps AI. Machine learning models are used to calibrate quantum hardware’s control pulses, to search large design spaces for better qubit layouts, and to help design more efficient quantum error-correction codes — all published, active research directions at Google Quantum AI and other labs. The reverse direction — quantum computers accelerating AI training or inference — remains a research field called quantum machine learning, with published papers exploring specific narrow tasks, but no demonstrated case of a quantum system outperforming classical machine learning on a practically useful problem at meaningful scale. The most credible near-term shared ground between the two fields is not AI acceleration at all, but quantum chemistry and materials simulation — a domain where AI models like AlphaFold already help today, and where future fault-tolerant quantum computers may eventually contribute complementary, more precise simulation capability for specific molecular systems classical methods handle poorly.

Hybrid classical-quantum computing — where a classical computer handles most of a workload and calls out to quantum hardware for a specific, well-suited subroutine — is the architecture most current research programs, including IBM’s and Google’s published roadmaps, actually target, rather than a vision of quantum computers replacing classical infrastructure outright. Variational algorithms such as the Variational Quantum Eigensolver (VQE), used for quantum chemistry problems, are explicitly hybrid: a classical optimizer iteratively adjusts parameters for a quantum circuit, combining the strengths of each. This is also the architecture most credible discussions of “AI plus quantum” describe: not quantum computers training language models, but classical AI systems and quantum coprocessors handling complementary parts of a larger workflow, each doing what it does best.

Responsible development is an explicit, published concern in both fields, not an afterthought this guide adds unprompted. Major AI labs including Google DeepMind, OpenAI and Anthropic publish safety and responsible-AI research and policies addressing model behavior, misuse risks and societal impact. Quantum computing’s responsible-innovation conversation centers substantially on the cryptographic transition described throughout this guide — ensuring the migration to post-quantum cryptography happens smoothly and in time, a coordination challenge spanning governments, standards bodies and every major technology vendor, and one where NIST’s 2024 standards represent significant, verifiable progress rather than a completed task.

Artificial intelligence and quantum computing represent two complementary frontiers of modern computing, not competing bets on the same outcome. AI, in its current deep-learning form, is already transforming industries through data-driven learning — running entirely on classical GPU and CPU hardware, from language models to protein-structure prediction. Quantum computing continues progressing toward practical applications in chemistry, optimization, materials science and cryptography, with genuine, peer-reviewed milestones — Google’s 2019 supremacy demonstration, its 2024 Willow error-correction result — marking real scientific progress without yet constituting broad, practical advantage over classical methods.

The next decade is likely to see increasingly powerful hybrid systems combining CPUs, GPUs and quantum processors, each handling the part of a computational workload it does best — not quantum computers replacing classical infrastructure, and not AI running on quantum hardware as a matter of course. That progress depends on continued advances in hardware, error correction, software and responsible scientific research, published incrementally through peer review and company roadmaps, rather than on hype or premature claims of arrival. The single most consequential real-world action already underway is the migration to post-quantum cryptography, prompted not by an existing threat but by prudent, proactive engineering ahead of one that current mathematics says could eventually exist.

Readers following ongoing developments are encouraged to consult official sources directly — Google Quantum AI, IBM Quantum, Microsoft Azure Quantum, NIST, and peer-reviewed journals including Nature and Science — rather than relying on any single account, including this one, as the final word on two fields that are, by design, still being actively written.

This guide draws on peer-reviewed publications including Google Quantum AI’s 2019 and 2024 Nature papers on Sycamore and Willow, Peter Shor’s 1994 and Lov Grover’s 1996 foundational algorithm papers, the 2017 “Attention Is All You Need” Transformer paper, NIST’s FIPS 203/204/205 post-quantum cryptography standards (August 2024), and official published roadmaps from IBM Quantum, Google Quantum AI and Microsoft Azure Quantum. Readers seeking primary documents should consult NIST’s Computer Security Resource Center, Nature’s physics and computer science sections, and each company’s official quantum-computing research pages for current updates.