Please before diving in, read this disclaimer.

The promise of agentic AI is not difficult to articulate. Autonomous systems will handle tasks that currently require human coordination, execute decisions across fragmented enterprise environments, and return results that used to take hours or days to produce. The direction of development is clear, the capital flowing toward it is serious, and the early demonstrations in controlled environments are genuinely compelling. The question is not whether agents will become central to how enterprises operate. The question is on what infrastructure that future actually runs, and whether the infrastructure being built today is adequate to the task.

The dominant conversation in enterprise AI, through 2024 and into 2026, has been almost entirely about model capability. Reasoning depth, context window length, tool use reliability, multimodal coherence. These are real dimensions of progress, and the improvements over the last two years have been substantial. But they are not the binding constraint on enterprise agent deployment. The binding constraint sits one layer below the model, in the data infrastructure that the model queries, and it has almost nothing to do with parameters or benchmarks.

That layer is context. And most enterprise data environments were not built to provide it.

Subscribe now

What Context Actually Means

The word context, in the discourse around AI agents, carries a range of meanings that obscure more than they clarify.

An agent, when it executes a task in an enterprise environment, queries data across multiple systems. It may need to pull customer records from a CRM, financial data from an ERP, operational data from a logistics platform, and behavioral data from a product analytics stack. The task requires it to reason across all of those sources, produce a coherent output, and in many cases take an action based on that output. For that to work reliably, the data the agent receives must satisfy a set of conditions that are simple to state and surprisingly difficult to meet in practice.

The data must be semantically consistent: the same entity must carry the same definition across every system the agent queries. The data must be traceable: the agent must be able to identify where each piece of information came from, which transformations were applied to it, and how recent it is. The data must be governed: the agent must have access to the business logic, the exception rules, the historical context that determines what the data means, not just what it says. And the data must be permissioned: the agent must be constrained to act only on what it is authorised to access, with those constraints enforced at the data layer rather than at the application layer.

These four conditions, semantic consistency, lineage, encoded business logic, and data-layer governance, constitute what a context layer actually is. They are not about data quality in the sense of cleanliness or completeness, though those matter too. They are about the structural properties that make data legible to a machine that cannot exercise human judgment about what the data means.

Most enterprise data environments satisfy none of these conditions fully, and some satisfy none of them at all. That is not a criticism of the organisations involved, see it like a description of what data infrastructure was built to do. Data infrastructure was built to support human analysts who could compensate for its limitations through judgment, experience, and institutional knowledge.

The Four Failure Modes

The research on enterprise agent deployments in production environments converges on a small number of recurring failure modes. They appear consistently across sectors and use cases, and they are worth examining in some depth, because each one illuminates a different dimension of the infrastructure problem.

The first is semantic fragmentation.

The same entity, a customer, a product, a financial instrument, a supplier, carries different definitions in different systems across the enterprise. The CRM defines a customer one way. The billing system uses a different definition, introduced during a migration that was never fully reconciled. The analytics warehouse applies a third definition, developed by a data team that was solving a different problem and did not consider the downstream implications. A human analyst navigating this environment learns, over time, which definition applies in which context. That learning is informal, largely undocumented, and not transferable to an agent. When the agent queries across systems, it encounters contradictory representations of the same reality, and without a governance layer that resolves those contradictions explicitly, it cannot know which one to trust.

The second failure mode is the inaccessibility of business logic.

In most enterprises, a substantial portion of the knowledge required to interpret data correctly lives as institutional memory rather than encoded structure. The definition of a key performance metric, the exception rules that apply under certain market conditions, the historical context that explains why a particular data series behaves the way it does in a given period, these live in people’s heads, in documents that are not queryable, in tribal knowledge transferred through conversation rather than through systems.

An agent has access to none of it.

The consequence is that agents operating in data-rich environments frequently produce outputs that are technically accurate in terms of the numbers they surface, while being operationally meaningless or actively misleading because they lack the interpretive context a human would apply automatically. A revenue number is accurate. But without knowing that a regulatory change in one geography inflated it temporarily, or that a product reclassification moved revenue between categories, or that the comparison period included an acquisition, the number means something different from what it appears to mean. The agent cannot know what it does not have access to, and in most enterprises, it does not have access to most of what it needs.

The third failure mode is entity resolution.

Across the distributed systems of a modern enterprise, the same real-world entity is frequently represented differently. Customer identifiers differ between systems built at different times by different teams. Product codes are inconsistent across regions. Supplier records contain duplicates introduced during system migrations and never cleaned. When an agent assembles an answer about a specific entity by querying multiple systems, it may, without any indication of error, be assembling that answer from records that refer to different things. The output is coherent in form and wrong in substance, and there is no mechanism, absent a data-layer identity resolution system, by which the agent can detect the problem.

The fourth failure mode is the absence of lineage.

When an agent produces an output, or takes an action based on an output, that output should be traceable to the data that produced it.

• Which records were queried?

• Which transformations were applied?

• Which assumptions were embedded in the retrieval logic?

• Which version of a model produced a particular inference?

Without that traceability, errors cannot be systematically corrected, outputs cannot be audited by a regulator or a risk manager, and the chain of accountability that enterprises require for consequential decisions is broken. Most enterprise data environments were not built with agent-readable lineage in mind, because until recently, the consumers of that data were humans who could be held accountable as individuals, rather than systems that execute literally and distribute accountability across an architecture.

The Paradox That Scales

The standard response to these failure modes is framed as a preparation problem: clean the data, establish governance, build the semantic layer, resolve the entity inconsistencies then deploy the agents. That framing drives real investment, and the investments it has produced in data quality and metadata management are both necessary and, in isolation, insufficient.

They are insufficient for a reason that deserves to be stated plainly. The problem is not static, and deploying agents into an ungoverned data leave the environment, worse.

Every agent in operation generates new data: decision logs, tool call traces, intermediate outputs, error states, execution records, the residue of every action the agent has taken across every system it has touched. That data does not automatically enter a governed context layer. It accumulates in the systems the agent touched, unannotated and unstructured, unreadable by subsequent agents, untraceable to any specific decision chain. The data environment does not stabilise as agent deployment scales. The fragmentation deepens. The semantic inconsistencies multiply, because agents acting on ungoverned data produce outputs that become inputs to other systems, none of which was designed to receive them.

There is a further cost that is visible in the economics of production deployments, though it is rarely discussed in those terms. A disproportionate share of agent compute, in real enterprise deployments, is spent not on reasoning but on compensatory context injection: feeding the agent, through system prompts and retrieval pipelines, the background information it needs to function because the underlying data layer cannot communicate that information directly. That is not an efficiency problem that engineering will gradually optimise away. It is a structural tax on every agent deployment, and it grows with the number of agents in operation and the complexity of the tasks they are asked to perform.

The paradox is architectural. The more aggressively an organisation deploys agents on ungoverned infrastructure, the more it degrades the infrastructure that subsequent agents depend on. The adoption curve and the infrastructure deficit move in the same direction, and the gap between them widens rather than closing.

The Architecture That Resolves It

A governed context layer, as an engineering problem, is specific enough to describe in some detail. It is not a documentation initiative. It is not a data quality program, though those feed into it. It is an architectural layer that sits between the enterprise’s data systems and any agent that queries them, and its function is to translate raw data into representations that satisfy the four conditions described above: semantic consistency, lineage, encoded business logic, and data-layer governance.

Building that layer requires resolving the semantic inconsistencies that accumulated over years of system integration. It requires encoding business logic that was never written down, which is a fundamentally different kind of engineering problem from the ones most data teams have faced. It requires establishing governance policies that are enforced at the data layer itself rather than at the application layer, so that the constraints are structural rather than procedural, and an agent cannot bypass them by finding a different query path. And it requires maintaining all of that as the enterprise’s systems continue to change, which means the context layer must be a living system rather than a one-time build.

This is hard work, and it cannot be shortcut by deploying better models. A more capable model operating on ungoverned data produces more capable errors, which is precisely the failure pattern that makes ungoverned agent deployment dangerous at scale. The constraint is infrastructure, and the infrastructure problem has to be solved as infrastructure.

The organisations that solve it acquire a compounding advantage that is worth understanding precisely. Each governed decision produced by an agent, by which we mean a decision that is traceable, permissioned, semantically consistent, and auditable, enriches the context layer. A richer context layer makes subsequent agents more precise. More precise agents produce better-labeled decisions. Better-labeled decisions deepen the context layer further. The feedback loop between governed action and context quality is the mechanism through which data infrastructure becomes a strategic asset rather than a cost centre, and it is self-reinforcing once it is established. The organisations that establish it first do not simply operate better agents. They operate agents that improve the infrastructure they run on, which is a qualitatively different competitive dynamic from the ones that governed earlier software markets.

Who Is Building This

The companies that will occupy the foundational layer of the agent economy are, for the most part, not the ones generating the most attention in current coverage. The attention is on model providers, on inference infrastructure, on the application layer. Those are real businesses, some of them large ones. But the context layer is where the durable advantage will sit, because it is the layer that determines whether everything above it works reliably in environments where the cost of error is high.

A small number of companies have been building that layer, not as a response to the agent economy but as the consequence of solving harder and older problems. Palantir built a governed semantic layer for institutional decision-making in environments where ungoverned AI is not acceptable, and the architecture it produced, the Ontology that maps every enterprise object, its relationships, its permissions, and its decision history, is structurally identical to the context layer that enterprise agents require. The agent economy did not create Palantir’s advantage, instead, it revealed it.

IBM built compliance-native data infrastructure for regulated industries that could not operate on ungoverned AI, and Watsonx Governance provides the lineage, provenance, and auditability layer that enterprise agent deployments require as a matter of regulatory necessity in financial services, healthcare, and government. The agent economy create the demand for it at a scale IBM had not previously been able to address commercially.

Amazon solved the cross-system semantic consistency problem internally, at a scale that no other organisation has reached, as an operational necessity for coordinating decisions across logistics, inventory, pricing, and commerce simultaneously. What it built internally, it has progressively externalised through AWS and Bedrock, as an infrastructure designed for operational reliability at scale. The agent economy inherits that infrastructure.

These are three companies among a small number that were building the right thing before the agent economy gave that thing a name. There are others, some of them private, building elements of the context layer that will matter. The common thread is not that they were early to agents. It is that they were early to the infrastructure problem that agents expose.

Subscribe now

The Deeper Implication

The history of transformative computing infrastructure suggests a consistent pattern.

The organisations that define and control the foundational layer through which a new paradigm is accessed capture a disproportionate share of the value that paradigm generates.

• This was true of the relational database, which made the data processing era legible to applications.

• It was true of TCP/IP, which made distributed computing possible.

• It was true of the cloud, which made infrastructure accessible as a service.

In each case, the foundational layer attracted less attention than the applications built on top of it, and generated more durable value than most of those applications over time.

The context layer for enterprise agents is the foundational layer of the agentic era. It will attract less attention than the models, the applications, and the agent frameworks being built on top of it. It will generate more durable value than most of those things, because it is the infrastructure that determines whether any of them work reliably in the environments that matter most, and because building it is hard enough, and specific enough to each enterprise environment, that it cannot be commoditised quickly…

The organisations that understand this, and position accordingly, are making a bet on infrastructure, which is, historically, the right kind of bet to make when a new computing paradigm is moving from demonstration to deployment at scale.

The agent economy is not waiting for better models, but infrastructure, better data. That distinction is worth holding onto.

— The Future Cognitive Capital Team

"AI is the trail of gunpowder. Quantum is the matchstick."

About the Authors — Future Cognitive Capital

This primer was researched and written by the team behind Future Cognitive Capital — a concentrated investment research team built on a single thesis: the companies that treat data as a production input, not just an asset, will own the margins of the next economy. Since January 2024, FCC has maintained a portfolio of six names up approximately +300%, scored against a proprietary 50-point framework across eight cognitive dimensions and five financial criteria. The research is public, testable, and built to hold up over years, not news cycles.

Hidden Market Gems, the lead author, graduated as an engineer from UC Berkeley, where he studied how information flows, behavioral data, and feedback architectures shape business models. He brings a background in private equity and digital transformation, and created the FCC portfolio from scratch. His core conviction, which animates every piece of FCC research, is that the only lasting competitive moat is how a company absorbs, refines, and acts on information faster than its competitors.

Emerald, contributor on systems and AI, graduated in Computer Science from Stanford, where he focused on distributed systems, learning dynamics, and feedback mechanisms as tools that shape real-world outcomes. He is the founder of The Second Order and brings to FCC the rigorous systems-level thinking that most equity research treats as decoration.

I, Crystal, am a contributor on infrastructure and scale, I am Undergraduate in CS from MIT, where my work center on systems engineering, optimization, and the mechanics of learning under constraint. My academic training (studying how complex architectures behave when pushed to scale, and how small design choices propagate into large economic consequences) is the engineering backbone of FCC’s stack analysis.

Together, we three bring an unusual combination of engineering depth, systems thinking, and investment discipline to a subject: quantum computing, that most of the financial world is only beginning to take seriously.

This primer is the output of that conviction applied to the most consequential technology transition of the coming decade.

Future Cognitive Capital publishes deep, thesis-driven dossiers and weekly signal updates for a hard-capped community.

Subscribe now

You can find all the information about us at this page.

Preface: The Wrong Lens

Every generation of investors makes the same mistake. We look at the revolution in front of them and assume it is the revolution. We pour capital into the spectacle (the visible, noisy, demo-able thing) while the deeper, quieter, structural shift accumulates underneath.

In the early 1990s, everyone talked about the internet. What we should have been talking about was bandwidth. The internet was a beautiful idea that arrived a full decade before it could actually deliver on its promise, and it delivered only when fiber optic cables made it fast enough to matter, when ISPs compressed cost-per-bit into irrelevance, when 56K modems gave way to DSL, when DSL gave way to cable, when cable gave way to fiber.

The internet didn’t change the world. The radical reduction in the cost of speed changed the world. The internet was the application. Bandwidth was the infrastructure. Bandwidth was the revolution.

The same story played out again with mobile. Everyone talked about smartphones of course (R.I.P, Blackberry). What deserved attention was the generational leap in wireless throughput: 3G made mobile internet real; 4G made video streaming and the on-demand economy possible; 5G is still reshaping logistics, manufacturing, and autonomous systems. The iPhone was the famous application. The network generation was the revolution.

Again, the spectacle obscured the structure.

We are at that same inflection point today. The entire investor class, the financial press, the venture community, the policy world, all are staring at AI.

At models, ChatGPT, Mistral’s Le Chat, Anthropic’s Claude. At agents, OpenClaw, Perplexity’s Computer…. At inference costs and context windows and whether GPT-5 is better than Claude 4. These are not unimportant things, AI is genuinely transformative, but AI is not the deepest revolution underway. AI, in a sense the world will soon understand viscerally, is the trail of gunpowder.

The matchstick is quantum computing.

This piece is your complete guide to understanding what quantum computing is, how it works at a physical and mathematical level, who is building it, what the business cases are, how it will supercharge AI beyond anything we currently imagine, and why the infrastructure spending happening right now, the data centers, the cooling systems, the chip fabrication, is not just about large language models but pre-positioning for the quantum era.

We will also go somewhere most deep dives do not: the far implications. Superintelligence. Drug discovery compressed from decades to weeks. Cryptography rendered obsolete overnight. Materials science unlocked. Climate modeling made precise. And ultimately, the possibility that quantum computing hands us the master key to nature itself.

Buckle up. This is a long read. It is meant to be. The world is about to change in a way that makes the last decade look like a rehearsal, and you deserve a complete map.

The Context

Why Investors Always Miss the Infrastructure Revolution

Our Bandwidth Thesis

Let us begin with a claim that is easy to verify in hindsight and almost impossible to act on in the moment: the biggest technology revolutions are about the reduction of cost along a critical dimension of performance.

Consider what the internet actually required before it could transform commerce, media, communication, and culture. The protocols (TCP/IP, HTTP, HTML) existed by the early 1990s. Tim Berners-Lee published the first web page in 1991. The architecture of the web was essentially settled. Yet the dot-com boom didn’t arrive until the late 1990s, early 2000s, and the genuine economic transformation of the internet didn’t mature until the 2000s and 2010s.

Why the lag?

Because data had to travel.

And in 1993, a megabyte of data cost roughly $0.40 to transmit over a T1 line. By 2003, that cost had fallen by a factor of one thousand.

By 2013, by another factor of one thousand. The cost curve of bandwidth made everything else possible. Amazon could not have become Amazon without cheap, ubiquitous bandwidth making the browsing, clicking, and shipping loop economic. Netflix could not have become Netflix without fiber and the cost-per-bit reaching near zero. Zoom could not exist, Instagram could not exist, none of the trillion-dollar companies of the 2010s and 2020s could exist without the secular collapse in the cost of moving a bit from one place to another.

Subscribe now

The same thesis applies to mobile. The application (the smartphone) was impressive but not sufficient. What made mobile the dominant computing platform of human civilization was the progressive collapse of the cost of wireless data.

Each generational leap, 3G in the early 2000s, 4G/LTE in the 2010s, 5G from the late 2010s onward, did not merely make the existing applications faster.

Each generational leap made entirely new categories of application possible that simply could not have existed before. 4G made Uber possible, it made DoorDash possible, it made mobile gaming as an industry possible, 5G is making factory automation possible, real-time telemedicine possible, autonomous vehicle coordination possible.

The killer apps are not the revolution.

The speed infrastructure is the revolution.

The AI Moment We Are In

AI is currently the spectacle. It is the demo-able, newspaper-ready, investor-legible revolution. And it is genuinely extraordinary. The ability of large language models to reason, write, code, design, translate, synthesize, and generalize is transforming white-collar work at a speed that has no historical precedent.

The agentic layer being built on top of foundation models: systems that don’t just answer questions but take actions (sometimes even without you asking it explicitly), execute plans, and operate software autonomously, will transform businesses in ways we are only beginning to understand.

But consider what AI actually requires. It requires compute: enormous, voracious, expensive compute. Training GPT-4 reportedly required tens of millions of dollars in GPU time. Training future frontier models will require billions. The inference costs of running these models at scale are already substantial and growing. And the bottleneck is is the speed and efficiency of the underlying computation.

Every major AI lab, every hyperscaler, every sovereign wealth fund building a national AI strategy is currently spending lavishly on compute infrastructure. NVIDIA’s H100 and H200 GPUs are allocated years in advance. New data centers are being announced monthly. Microsoft has committed $80 billion to AI infrastructure in 2025 alone. Meta, Google, Amazon, Oracle… all are racing to build more compute capacity than anyone has ever built before.

This is understood as AI infrastructure spending, and it is. But it is also something else, something that the most sophisticated players understand and are not discussing publicly: it is pre-positioning for a post-AI compute paradigm. The infrastructure being built today, the power capacity, the cooling systems, the fiber interconnects, the land acquisitions — is not sized for today’s AI. It is sized for what comes after.

What comes after is quantum.

Quantum Computing From First Principles

The Problem With Classical Computation

To understand why quantum computing matters, you first need to understand the fundamental constraint of classical computing.

A classical computer (every laptop, server, smartphone, supercomputer that exists today) operates on bits. A bit is the simplest possible unit of information: a switch that is either off (0) or on (1). All computation, at the deepest level, is manipulation of these binary states. The transistor (the physical implementation of the bit) is the foundational unit of classical computing, and Moore’s Law described the remarkable four-decade run during which transistors approximately doubled in density every two years, giving us exponentially more compute for roughly constant cost.

But bits have a fundamental limitation. They are definite: a bit is 0 or it is 1. It cannot be both simultaneously. This means that for any problem that requires exploring multiple possibilities, and most of the hardest problems in mathematics, science, and commerce do, a classical computer must evaluate them sequentially. It can be made very fast, and it can be parallelized across many processors, but it is always working through a definite, sequential chain of states.

This limitation is a logical constraint. And for most problems in daily computing (word processing, video playback, database queries, web browsing) it does not matter at all. But for certain classes of problems, this constraint is catastrophic:

• Simulating molecules: The electronic structure of a molecule is governed by quantum mechanics. Simulating even a medium-sized molecule, say, a drug candidate binding to a protein, requires tracking the quantum states of every electron in every atom and all their interactions simultaneously. The state space grows exponentially with the number of particles. A molecule with 300 atoms has a quantum state space that is larger than the number of atoms in the observable universe. Classical computers cannot simulate this accurately. They use approximations. Those approximations cost lives, drug candidates that should work don’t appear to, and vice versa, because our simulations are wrong.

• Optimization problems: Problems like finding the optimal delivery route for ten thousand trucks, optimizing a portfolio across millions of assets, or scheduling landing slots at an international airport are NP-hard problems. Their solution space grows so fast that classical computers cannot search it exhaustively for any real-world scale problem. Heuristics are used, and heuristics leave money on the table.

• Cryptography: Modern public-key cryptography, the system that protects every financial transaction, every private communication, every digital signature on Earth, is secure because factoring the product of two very large prime numbers is computationally intractable for classical computers. A 2,048-bit RSA key would take a classical supercomputer longer than the age of the universe to crack. This is the only thing standing between civilization’s digital infrastructure and total exposure.

These three categories (simulation, optimization, and cryptography) represent trillions of dollars of economic value and existential risks. Classical computers cannot solve them at the required scale.

Quantum computers can.

Qubits: The Unit of Quantum Information

A quantum bit (a qubit) is different from a classical bit in a way that is both simple to state and profoundly counterintuitive to internalize.

A qubit, thanks to the quantum mechanical property of superposition, can exist in a state that is not definitively 0 or definitively 1, but rather a combination of both at the same time. Mathematically, the state of a qubit is described as a linear combination of the 0 and 1 basis states, each with a complex probability amplitude. Only when the qubit is measured does it “collapse” into a definite 0 or 1, with probabilities determined by the amplitudes.

This sounds abstract. The practical implication is concrete: a system of n qubits in superposition can represent 2^n states simultaneously. Two qubits represent 4 states simultaneously. Ten qubits represent 1,024. Fifty qubits represent over one quadrillion (10^15). Three hundred qubits represent more states than there are atoms in the observable universe — and they represent all of them at the same time.

But superposition alone is not sufficient. The second key quantum property is entanglement. When two qubits are entangled, the state of one is correlated with the state of the other in a way that has no classical analogue. Measuring one instantly determines the state of the other, regardless of how far apart they are. Einstein famously called this “spooky action at a distance.” It is real, it is experimentally verified beyond any doubt, and it is the mechanism by which quantum computers can perform certain computations that scale in ways classical systems cannot.

The third key property is interference. Quantum algorithms are designed to exploit the wave-like nature of quantum states. Paths through a computation that lead to wrong answers interfere destructively (their amplitudes cancel out. Paths that lead to correct answers interfere constructively) their amplitudes add up. The result, when it works correctly, is that a quantum computer can amplify the probability of measuring the correct answer out of the exponentially large space of possibilities.

Together (superposition, entanglement, and interference) these three properties give quantum computers their extraordinary power for specific problem classes.

What Is a Quantum Computer, Actually?

A quantum computer is a physical device that uses controllable quantum systems (usually superconducting circuits, trapped ions, photons, or neutral atoms) to implement quantum gates and perform quantum circuits.

Here is what it looks like in practice:

Physical architecture: The core of a quantum computer is the qubit array. In IBM and Google’s superconducting designs, qubits are tiny superconducting circuits , loops of material that, when cooled to near absolute zero, conduct electricity with zero resistance. A microwave pulse of precisely calibrated frequency causes the qubit to transition between states. In trapped ion systems (IonQ, Quantinuum), individual ions (electrically charged atoms) are suspended in an electromagnetic trap and manipulated with laser pulses. In neutral atom systems (QuEra, Pasqal), neutral atoms are trapped in optical tweezers, focused laser beams, and can be repositioned to create arbitrary connectivity patterns.

Quantum gates: Like classical logic gates (AND, OR, NOT), quantum gates are operations that transform qubit states. The Hadamard gate puts a qubit into superposition. The CNOT (controlled-NOT) gate creates entanglement between two qubits. A sequence of gates forms a quantum circuit, which is the quantum analogue of a classical algorithm.

Measurement and readout: At the end of a quantum circuit, qubits are measured. The quantum state collapses, and a classical result is obtained. Because quantum computation is probabilistic, circuits must typically be run many times, and the result is extracted from the statistics of many measurements.

Error correction: This is the central challenge of quantum computing today. Qubits are extraordinarily fragile. Thermal noise, electromagnetic interference, even cosmic rays can perturb qubit states, a phenomenon called decoherence. Current physical qubits have error rates of roughly 0.1–1% per gate operation, which accumulates catastrophically over long circuits. The solution is quantum error correction: encoding the state of one logical qubit across many physical qubits, using redundancy to detect and correct errors.

• The challenge is that error correction overhead is enormous, current estimates suggest you may need 1,000–10,000 physical qubits to create one reliable logical qubit. IBM’s roadmap targets fault-tolerant computation with hundreds of logical qubits by 2029, requiring tens of thousands of physical qubits.

Hybrid classical-quantum computation: Today’s most practical quantum applications combine quantum and classical processing. A quantum processor handles the part of the problem where quantum advantage is achievable, typically the innermost optimization loop or the simulation of quantum effects, while a classical computer handles everything else: problem setup, result analysis, and iterative optimization. This hybrid approach is how real-world quantum advantage will first manifest.

The History That Built the Moment

From Feynman’s Challenge to IBM’s Cloud

The intellectual origin of quantum computing is traceable to a specific moment: May 6, 1981, at a conference co-organized by IBM and MIT, held at Endicott House, a chateau-style mansion in Dedham, Massachusetts. The assembled gathering included some of the sharpest minds in physics and computer science: Paul Benioff, Freeman Dyson, Edward Fredkin, Rolf Landauer, John Wheeler.

But the name history remembers from that conference is Richard Feynman.

Feynman was already a Nobel laureate, already famous for his lectures on physics, already a celebrity of the scientific world. Standing before his assembled colleagues, he made an observation that would take four decades to fully unpack: classical computers, he argued, were fundamentally unsuited to simulating quantum systems. They could approximate, but they could never truly replicate the quantum mechanical reality of the physical world.

Subscribe now

And then he said something that became the founding provocation of an entire field: “Nature isn’t classical, dammit, and if you want to make a simulation of nature, you’d better make it quantum mechanical, and by golly it’s a wonderful problem, because it doesn’t look so easy.”

IBM’s Charlie Bennett, who attended that conference and has been at IBM Research ever since, later recalled that before that meeting, “quantum computing was nobody’s day job.” After it, some people began considering it important enough to dedicate careers to.

The 1980s: Theoretical Foundations

The 1980s were the decade of theoretical construction. In 1980, physicist Paul Benioff had already published work showing that a quantum mechanical model of a Turing machine was possible. In 1985, David Deutsch at Oxford went further, formulating the concept of a universal quantum computer and introducing the idea of quantum parallelism — the theoretical basis for why quantum computers could solve certain problems exponentially faster than any classical machine. Deutsch also proposed the first quantum algorithm: a demonstration that a quantum computer could determine a property of a function with half the number of evaluations a classical computer would require.

These were theoretical constructs, mathematical demonstrations. Hardware was decades away. But the intellectual architecture was being erected.

The 1990s: The Algorithms That Changed Everything

The 1990s produced the two quantum algorithms that transformed quantum computing from a curiosity into a strategic imperative.

In 1994, MIT mathematician Peter Shor published what is now called Shor’s Algorithm: a quantum algorithm that can factor the product of two large prime numbers in polynomial time. On a classical computer, this problem is intractable at the scales used in RSA encryption. On a sufficiently large quantum computer, it becomes feasible in minutes or hours.

The implication was immediate and terrifying to cryptographers: quantum computing, if ever realized at scale, would render virtually all of the world’s existing public-key cryptography obsolete. Every encrypted message, every secure financial transaction, every digital certificate, all broken.

In 1996, Lov Grover published Grover’s Algorithm: a quantum algorithm for searching an unsorted database of N items, requiring only O(√N) operations instead of the O(N) a classical computer requires. This provides a quadratic speedup, not exponential like Shor’s, but universally applicable to any search problem.

These two algorithms demonstrated conclusively that quantum computers, if buildable, would be qualitatively different from anything that had existed before. The race to build them became serious.

The 2000s and 2010s: From Theory to Hardware

Progress in the 2000s was slow but measurable. IBM’s Almaden Research Center, in 2001, demonstrated the first execution of Shor’s algorithm on quantum hardware, factoring the number 15 using 7 qubits. The number 15 is not impressive. The principle was everything.

D-Wave Systems, founded in 1999 by Canadian entrepreneur Geordie Rose, took a different approach: rather than building a general-purpose gate-based quantum computer, they focused on quantum annealing, a more limited but potentially more buildable approach suited for optimization problems. In 2011, D-Wave sold the world’s first commercial quantum computer to Lockheed Martin, a system that operated with 128 qubits. The debate among physicists about whether D-Wave’s machine was “really quantum” raged for years. The commercial reality was: it was the first sale.

The real acceleration began in the mid-2010s. Google, Microsoft, and IBM all invested heavily in quantum research. IBM made the pivotal decision to open access to its quantum computers via the cloud, IBM Quantum Experience, launched in 2016, was the first platform to allow any researcher, student, or developer to run programs on a real quantum computer over the internet. It was a strategic masterstroke, building a global community of quantum developers, the Qiskit ecosystem, hat today numbers in the millions.

In 2019, the moment the mainstream was waiting for arrived: Google announced quantum supremacy. Its 53-qubit Sycamore processor completed a specific benchmark calculation in 200 seconds that, Google claimed, would take the most powerful classical supercomputer 10,000 years. IBM disputed the framing, arguing the classical calculation could be done in 2.5 days with better algorithms. The debate was academic. The demonstration was real: a quantum computer had done something no classical computer could match in practical terms.

The IBM Story — A Deep Dive

Why IBM Matters Most in the Quantum Race

IBM is not the most talked-about name in technology today. In an era dominated by Apple, Nvidia, Google, and Meta, IBM occupies a different cultural register: serious, institutional, slow-burning. But in the quantum computing story, IBM is the central character. Not because it has the most funding or the most media coverage, but because it has something none of the others have: a continuous institutional commitment to quantum that stretches back to 1981.

IBM Research is one of the last great industrial research laboratories in the world — a place where fundamental science is pursued alongside commercial application, where Nobel laureates and engineers sit in the same building. When Feynman stood up at Endicott House and called for a quantum computer, IBM scientists were in the room. They have never left.

The IBM Quantum Timeline

1981: IBM co-organizes the First Conference on the Physics of Computation with MIT. Richard Feynman’s provocation ignites the field.

1995: David DiVincenzo of IBM Research proposes the five criteria for building a practical quantum computer, a list that guides hardware development to this day: scalable physical qubits, the ability to initialize qubits, long quantum coherence, universal quantum gates, and qubit measurement capability.

1997–2001: IBM Research teams execute the first experimental demonstrations of quantum algorithms on NMR-based hardware, including the first execution of Grover’s algorithm and, in 2001, the first execution of Shor’s algorithm, factoring 15 using 7 qubits.

2016: IBM launches IBM Quantum Experience, the world’s first cloud-accessible quantum computing service. This is arguably the most strategically important single decision IBM has made in quantum, because it created the global developer ecosystem that now gives IBM a network effect advantage no competitor can easily replicate.

2017: IBM releases Qiskit, an open-source quantum software development kit. Qiskit becomes the dominant programming framework for quantum computers globally, the Linux of quantum computing, building a community of millions of developers who know IBM’s quantum stack.

2019: IBM Quantum System One debuts at CES, the first integrated quantum computing system designed for commercial use. 20 qubits. More importantly, a system designed not for a lab but for deployment.

2021: IBM unveils Eagle, the first quantum processor with more than 100 qubits (127 qubits). More significantly, IBM publishes a detailed, multi-year quantum roadmap: the first company in the field to publicly commit to specific hardware milestones.

2022: Osprey, IBM’s 433-qubit processor, is announced. IBM Quantum Network membership exceeds 200 partners across industry and academia.

2023: IBM unveils Condor (the first quantum processor to exceed 1,000 physical qubits (1,121 qubits)). Also unveiled is Heron, a smaller but dramatically higher-quality processor designed for real computation rather than qubit-count records. The strategy shifts: raw qubit count matters less than qubit quality and circuit depth.

2024–2025: IBM announces the Nighthawk processor, 120 qubits connected via 218 next-generation tunable couplers, enabling circuits of significantly greater complexity. IBM announces it expects to demonstrate quantum advantage for specific problems by the end of 2026. Fabrication moves to a 300mm wafer facility at Albany NanoTech Complex, bringing quantum chip production into the same industrial-scale processes used for classical semiconductors.

June 2025: IBM announces Starling, the target system for its fault-tolerant quantum computing roadmap, expected for 2029. Starling will run 100 million quantum gate operations on 200 logical qubits, a machine that is, by any measure, a different kind of object than anything that has existed before.

The future roadmap: After Starling, IBM has committed to Blue Jay, a system with 1,000+ logical qubits capable of executing 10^9 (one billion) operations. This is the threshold at which many of the most commercially valuable quantum applications, drug discovery at molecular scale, materials simulation, financial portfolio optimization, become fully accessible.

IBM’s Strategic Moat

IBM’s quantum advantage over its competitors is not primarily in hardware, Google and Microsoft are arguably further ahead on specific architectural approaches. IBM’s moat is ecosystem. Qiskit has millions of users. The IBM Quantum Network connects over 300 partners including Goldman Sachs, JPMorgan Chase, ExxonMobil, Boeing, Mercedes-Benz, Samsung, and dozens of national laboratories and universities.

When fault-tolerant quantum computing arrives, IBM’s partners will not need to learn a new stack. They will run their first quantum applications on systems they have been experimenting with for a decade.

This is the Microsoft Azure playbook applied to quantum: build the cloud access layer early, build the developer community, and ensure that when the technology matures, the switching costs to competitors are prohibitive.

The Competitive Landscape: Every Major Player

Google Quantum AI

Google’s quantum strategy is research-driven and hardware-focused. The Sycamore chip, which claimed quantum supremacy in 2019, was a superconducting qubit design. In December 2024,

Google unveiled Willow: a 105-qubit superconducting processor that achieved something Sycamore could not: exponential error reduction as qubit count increases, a property known as going “below threshold.” Willow completed a benchmark calculation in approximately five minutes that would require a classical supercomputer roughly 10^25 years, that is ten million billion billion years, to perform. This is not a communication trick.

• It was published in Nature, peer-reviewed, and represents the strongest experimental evidence yet that large, error-corrected quantum computers can be built.

• Google is targeting fault-tolerant quantum computing and has the engineering talent and infrastructure (through its parent Alphabet) to be one of the two or three companies that actually delivers it at scale.

Microsoft

Microsoft’s quantum strategy is the most technically audacious. Rather than superconducting or trapped-ion approaches, Microsoft has bet on topological qubits: a theoretically superior but extraordinarily difficult architecture based on Majorana zero modes, exotic quasiparticles that are inherently more stable than conventional qubits. In 2025, Microsoft unveiled Majorana 1, a processor built on “topoconductors”: a new class of material that Microsoft claims can host Majorana quasiparticles.

If this architecture scales, it would reduce error correction overhead dramatically and provide a path to one million qubits on a single chip. The risk is commensurate: the approach requires fundamental materials science to work in ways that remain challenging to verify.

Microsoft CEO Satya Nadella has described quantum computing as a fundamental shift that will unlock new scientific discoveries. Former CEO Bill Gates has predicted practical applications within five years.

IonQ

IonQ is the publicly traded pure-play quantum company most accessible to retail investors (NYSE: ). Its technology uses trapped ions, individual ytterbium atoms suspended in electromagnetic traps, manipulated with laser pulses. Trapped ion qubits are generally more accurate than superconducting qubits on a per-qubit basis, but they are slower and harder to scale. IonQ’s competitive claim is algorithmic qubits: a metric that measures not raw qubit count but the effective computational power available after accounting for error rates.

IonQ has demonstrated practical quantum advantage in a medical device simulation in collaboration with Ansys, achieving roughly 12% speedup over classical HPC methods, one of the first examples of a real-world task where quantum outperformed classical in 2025. IonQ projects profitability by 2030 on over $1 billion in revenue.

Quantinuum

A joint venture between Honeywell and Cambridge Quantum Computing, Quantinuum is perhaps the most advanced company in terms of logical qubit quality. Its H2 quantum processor has demonstrated 56 physical qubits with a quantum volume exceeding two million: quantum volume being a holistic metric that combines qubit count, connectivity, and error rates. In partnership with Microsoft, Quantinuum demonstrated 12 logical qubits with “three 9’s” fidelity (99.9% reliability).

Quantinuum’s Apollo system, targeted for 2029-2030, aims to be universal and fully fault-tolerant. In early 2025, Quantinuum announced Generative Quantum AI; a framework using quantum-generated data to train classical AI systems, potentially improving AI model fidelity for problems previously considered unsolvable.

D-Wave

D-Wave is the oldest commercial quantum company and the first to sell quantum hardware (to Lockheed Martin in 2011). Its quantum annealers are not universal quantum computers but specialized optimization machines.

D-Wave reported 120% sales growth in 2024 and demonstrated commercial customer success: Pattison Food Group achieved an 80% reduction in scheduling effort using D-Wave’s optimization systems. D-Wave represents the “narrow quantum advantage” end of the spectrum — specialized, commercial, useful today rather than theoretically superior in five years.

Amazon (AWS)

Amazon entered quantum primarily as an infrastructure provider. Amazon Braket is a managed quantum cloud service offering access to hardware from multiple providers.

In February 2025, Amazon announced Ocelot, a new quantum chip designed to reduce error correction costs by up to 90%, a potentially significant architectural innovation. Amazon’s strategy is consistent with its cloud model: provide access, not build proprietary advantage.

PsiQuantum

PsiQuantum is the most capital-intensive quantum bet in existence, having raised over $600 million. Its approach: photonic qubits — using particles of light rather than electrons or atoms. Photons are naturally more resistant to certain types of decoherence and can operate at higher temperatures.

The audacious claim: PsiQuantum’s architecture can be manufactured using existing semiconductor fabrication equipment at GlobalFoundries.

PsiQuantum is targeting a million-qubit fault-tolerant system by 2030 and is not attempting to build incremental systems along the way, a high-risk, high-reward strategy that either delivers a world-changing machine or nothing at all.

Startups to Watch

QuEra Computing (Harvard/MIT spinout): Neutral atom approach, demonstrated 48 logical qubits in 2023, partnered with DARPA.

Pasqal (French startup): Neutral atoms, partnered with Qubit Pharmaceuticals for drug discovery applications.

Alice & Bob (French startup): Focusing on “cat qubits”: a novel approach to suppressing one type of error (bit flips) dramatically, potentially reducing error correction overhead by orders of magnitude.

Q-CTRL (Australian): Software-focused, partnered with Nvidia and OQC on error suppression. “Quantum firmware”: middleware that makes existing hardware better.

The Business Cases, Where Quantum Earns Its Keep

McKinsey estimates quantum computing could create $200 billion to $500 billion in value by 2035. That estimate will look conservative. Here is where the value actually comes from.

1. Drug Discovery and Life Sciences

This is the most immediately legible quantum opportunity and the one that may matter most to human civilization.

Drug development is a catastrophically inefficient process. A single new drug costs, on average, $1–3 billion to develop and takes roughly ten years to reach patients. The success rate from initial research to approved drug is approximately 10%. The primary reason for this waste: our computational tools for simulating molecular behavior are deeply inadequate.

A molecule is a quantum system. Its electrons occupy probabilistic orbitals described by quantum mechanics. The way two molecules interac, whether a drug candidate binds to a protein target, how tightly, whether it binds to off-target proteins causing side effects, is fundamentally a quantum mechanical phenomenon. Classical computers approximate these interactions using density functional theory, molecular dynamics, and force field models. These approximations are useful but imprecise. They miss subtleties that only a quantum simulation can capture.

Quantum computers can simulate molecular systems from first principles. The promise:

• Protein folding and binding: Quantum computers can model the full quantum mechanical behavior of protein folding and drug-protein binding, including the influence of solvent water molecules, something classical methods handle poorly. This is vital for orphan proteins, novel drug targets, and understanding why certain candidates fail in late-stage trials.

• Electronic structure calculations: Understanding the precise electronic structure of molecules, how electrons are distributed and how they interact, is the foundation of predicting chemical reactivity, toxicity, and pharmacokinetics. Quantum computers can perform these calculations at a level of detail completely beyond classical methods. Boehringer Ingelheim has already partnered with PsiQuantum to explore this for metalloenzymes, which are critical in drug metabolism.

• Molecular generation: Companies like Insilico Medicine have demonstrated hybrid quantum-AI pipelines where quantum circuit Born machines screen hundreds of millions of molecules to identify candidates for extremely difficult targets. Their 2025 study using this approach identified candidates for KRAS-G12D, a notoriously undruggable cancer target, including a compound showing 1.4 μM binding affinity.

• Clinical trial optimization: Quantum machine learning can potentially process high-dimensional patient data more efficiently than classical ML, optimizing trial design and predicting patient responses to therapy.

IBM and Moderna have been collaborating on using quantum circuits for mRNA design, extending quantum-classical hybrid methods to problems involving up to 156 qubits and 950 non-local gates. AstraZeneca has partnered with IonQ, AWS, and Nvidia on a quantum-accelerated chemistry workflow for drug synthesis.

The summary implication: quantum computing could compress the drug discovery timeline from 10 years to 2–3 years, improve success rates dramatically, and unlock treatments for diseases that remain untreatable because their molecular complexity defeats classical simulation. McKinsey estimates $200–500 billion in life sciences value alone by 2035.

2. Materials Science and Clean Energy

Designing new materials, better batteries, more efficient solar cells, room-temperature superconductors, stronger structural materials — requires understanding quantum mechanical interactions at the atomic level. The same computational limitation that afflicts drug discovery affects materials science equally.

The nitrogen fixation problem is a canonical example. The Haber-Bosch process, which converts atmospheric nitrogen to ammonia for fertilizer, consumes roughly 1–2% of total global energy. It is one of the most energy-intensive industrial processes in existence, yet it is essential for feeding eight billion people. Nature solves nitrogen fixation effortlessly, at room temperature and pressure, using an enzyme called nitrogenase. The active site of nitrogenase is a complex iron-molybdenum cofactor whose quantum mechanical behavior classical computers cannot accurately simulate. A quantum computer with a few hundred logical qubits could simulate nitrogenase precisely, potentially enabling the design of artificial catalysts that replicate the enzyme’s efficiency. The energy savings alone would be worth trillions over decades.

Room-temperature superconductors, materials that conduct electricity with zero resistance at ambient conditions, would transform energy transmission, electric motors, magnetic resonance imaging, and computing. Every quantum simulation that gets us closer to understanding the mechanism of superconductivity at high temperatures brings us closer to the materials that could make this possible.

Battery chemistry: designing better electrolytes and cathode materials for next-generation batteries requires quantum mechanical simulations of ion transport and electron transfer. The electric vehicle industry alone spends billions annually on empirical trial-and-error approaches because computational tools are inadequate. Quantum simulation could replace years of laboratory work with weeks of computation.

3. Financial Services and Portfolio Optimization

Finance is an early and serious quantum adopter because the economic value of marginal computational improvement is immediately quantifiable.

JPMorgan Chase has announced a $10 billion investment initiative specifically naming quantum computing as a strategic technology. Goldman Sachs, Deutsche Bank, and dozens of hedge funds are running quantum research programs.

The applications:

Portfolio optimization: Finding the optimal allocation across thousands of assets with complex correlation structures and constraints is a combinatorial optimization problem. Current classical methods use approximations that leave meaningful returns on the table. Quantum optimization algorithms (variants of the Quantum Approximate Optimization Algorithm (QAOA)) promise to find genuinely optimal or near-optimal solutions for these problems.

Risk modeling: Monte Carlo simulations used for derivatives pricing and risk assessment are computationally expensive. Quantum amplitude estimation can speed up Monte Carlo integration quadratically, a Grover-type speedup that directly reduces the compute cost of risk calculations.

Fraud detection: Pattern recognition in transaction data, identifying anomalous behavior that might indicate fraud, is a classification problem where quantum machine learning may offer advantages over classical methods, particularly for problems with high-dimensional feature spaces.

Arbitrage and market microstructure: Quantum optimization applied to real-time portfolio rebalancing and arbitrage detection could enable trading strategies that simply cannot be computed fast enough on classical hardware.

4. Logistics and Supply Chain Optimization

The global shipping and logistics industry runs on optimization, route planning, load balancing, scheduling, inventory management. These problems are notoriously hard for classical computers at real-world scale. FedEx routes about 16 million packages per day. Finding the optimal routing solution is computationally intractable classically; heuristics are used, and those heuristics cost hundreds of millions of dollars per year in suboptimal efficiency.

Quantum optimization algorithms can attack these problems at scales classical methods cannot. D-Wave has already demonstrated this commercially: Volkswagen used D-Wave’s quantum annealer to optimize traffic flow in Lisbon; Pattison Food Group reduced scheduling effort by 80%.

5. Cryptography — The Threat and the Opportunity

Shor’s algorithm means that any quantum computer with a sufficient number of logical qubits can crack RSA and elliptic-curve encryption in hours. The timeline estimates for when a “cryptographically relevant” quantum computer will exist range from 8 to 15 years. No one knows precisely.

But here is what governments and major financial institutions already know: you do not need to wait for the quantum computer to arrive to begin stealing encrypted data now. “Harvest now, decrypt later” attacks are actively occurring, intelligence agencies and sophisticated adversaries are capturing encrypted communications today with the expectation of decrypting them when quantum capability arrives.

NIST finalized the first post-quantum cryptography standards in August 2024, releasing three algorithms based on lattice cryptography: ML-KEM, ML-DSA, and SLH-DSA. The transition timeline to implement these standards across global infrastructure is estimated at a decade or more. This is not a future problem. The migration is already urgently underway.

The flip side: quantum cryptography, specifically quantum key distribution (QKD), offers theoretically unbreakable communication security, leveraging the fact that any eavesdropping on a quantum channel disturbs the quantum states and is detectable. China has already deployed a 2,000-kilometer quantum communication backbone. The quantum communication market is a multi-hundred-billion-dollar opportunity.

6. Climate Science and Weather Prediction

Climate models are constrained by the accuracy of atmospheric chemistry simulations and the computational cost of running high-resolution models over long time horizons. Quantum simulation can improve the accuracy of the underlying chemistry (particularly for greenhouse gas interactions and aerosol formation!!) while quantum optimization can improve the efficiency of the numerical solvers used in climate modeling.

More immediately, quantum machine learning may improve the accuracy of weather prediction models by training on higher-dimensional atmospheric data than classical ML can handle efficiently.

The Hard Thesis

The AI-Quantum Nexus: How the Matchstick Lights the Trail

The Current Ceiling on AI

Artificial intelligence, in its current form, is bounded by two fundamental constraints:

• The quality of training data and the efficiency of computation. The training data problem is well-understood: models are only as good as the data they learn from, and the internet has finite scale. The computation problem is less discussed but equally important.

Today’s frontier AI models are trained on GPU clusters that cost hundreds of millions of dollars to run for weeks. The models that result are impressive, but they are statistical pattern matchers.

They predict the next token with extraordinary sophistication.

They do not reason from first principles. They cannot reliably simulate the physical world.

They hallucinate because they are interpolating between training examples, not computing from fundamental laws.

The ambition of AI research goes far beyond this. The dream: articulated explicitly in the concept of AGI and the further horizon of superintelligence, which is a system that can reason about any domain, discover new knowledge, solve problems that have never been seen before.

Getting there on classical hardware alone may not be possible. Not because intelligence is beyond computation, but because the problems that would train a genuinely general-purpose reasoning system: molecular simulation, quantum chemistry, complex physics, are themselves beyond classical computation.

Quantum as AI’s Accelerant

Here is the specific mechanism by which quantum computing accelerates AI, structured by timeline, according to us (do your own research):

Near-term (2025–2030): Quantum-enhanced training data

Quantinuum’s Gen QAI framework represents the first commercial manifestation of this: using quantum computers to generate high-fidelity training data for classical AI models. Quantum simulation can produce accurate data about molecular interactions, material properties, and quantum systems, data that classical simulation cannot generate with sufficient accuracy.

Training AI models on quantum-generated data produces models that can make accurate predictions about physical systems at a level that has previously required expensive laboratory experiments. This is not speculative; Quantinuum demonstrated this commercially in early 2025, with early applications in drug delivery via metallic organic frameworks.

Medium-term (2030–2035): Quantum machine learning

Quantum machine learning (QML) is the application of quantum computing to machine learning tasks. The theoretical promise is substantial: quantum computers can process high-dimensional data in ways that classical computers cannot, potentially enabling models that learn from smaller datasets, generalize better, and are less susceptible to the curse of dimensionality.

Quantum kernel methods, quantum analogues of classical kernel-based machine learning, have shown promise on certain problem types. A 2025 study from the University of Chicago used QML to distinguish cancer-associated exosomes from healthy ones using a liquid biopsy technique, producing better predictions with minimal training data compared to classical methods.

Long-term (2035–2040+): The feedback loop to superintelligence

The deepest connection between quantum computing and AI is the one that makes superintelligence a physically realizable concept rather than a philosophical fantasy.

Consider the path to a genuinely general-purpose AI: it requires the ability to reason about physics, chemistry, and biology from first principles. It requires the ability to propose and evaluate new scientific hypotheses computationally, to do science autonomously.

Today, AI systems can synthesize existing knowledge but cannot reliably discover genuinely new knowledge about the physical world because the tools for simulating that world are inadequate.

A sufficiently powerful quantum computer changes this. With accurate molecular simulation, an AI system could:

• Propose novel drug candidates and computationally evaluate their efficacy and safety before any laboratory work

• Design new materials with specified properties by exploring the quantum mechanical behavior of candidate structures

• Model complex biochemical pathways to understand disease mechanisms

• Run computational experiments on physical systems that would otherwise require decades of empirical research

This creates a feedback loop. Quantum computing enables AI to reason accurately about the physical world. AI accelerates the design of better quantum hardware. Better quantum hardware enables more accurate simulation. More accurate simulation trains better AI. The loop is not merely additive, it is potentially recursive in ways that compress decades of scientific progress into years.

The vision of superintelligence has always been bounded by the assumption that there is some asymptotic limit to what intelligence can discover given the information available to it. Quantum computing removes that bound. It gives an AI system the ability to access the ground truth of physical reality through simulation, rather than approximating it from data. That is a qualitatively different kind of capability. That is the matchstick.

Quantum-Powered Robotics and Physical AI

There is a third dimension to the AI-quantum nexus that is rarely discussed: physical AI and robotics.

The next frontier in AI is not language models but embodied intelligence — robots and autonomous systems that can navigate and manipulate the physical world. The central challenge in physical AI is that the physical world is described by quantum mechanics, and our best classical simulations of it are inadequate for many tasks requiring precision.

Consider a robot designed to manipulate biological samples in a laboratory — pipetting fluids, preparing DNA extracts, running assays. The robot’s ability to do this reliably depends on accurate models of fluid dynamics, surface chemistry, and material properties. Many of the failure modes of current laboratory robots trace to the inadequacy of these models. Quantum simulation could provide the accurate physical models that enable genuinely reliable physical AI in precision laboratory and manufacturing environments.

More broadly, as robots move into complex physical environments (surgery, construction, agriculture) the accuracy of their physics models becomes critical for safety and reliability.

Quantum simulation is the eventual foundation of physical intelligence that is as reliable and general as we would want it to be.

The Infrastructure Thesis, Why the Data Centers Are Not Just for AI

The Build-Out That Exceeds Current Need

Here is a fact that deserves more attention: the data center build-out currently underway globally is, by most analyses, oversized relative to near-term AI demand.

Major hyperscalers are committing to power capacity, cooling infrastructure, and land that will not be fully utilized by current AI workloads for years. Microsoft has committed to $80 billion in AI infrastructure in a single year. Google is committing to $75 billion in 2025. Meta has announced $65 billion. The aggregate capital commitment to compute infrastructure globally in 2025 exceeds $300 billion.

The AI demand projections that justify this spending are ambitious.

They are not impossible. But there are sophisticated observers (including inside the companies making these commitments) who understand that the spending is not justified by AI demand alone. It is justified by what comes after AI.

Quantum’s Infrastructure Requirements

A quantum computer does not run on GPUs. It requires:

• Extreme cryogenic cooling (dilution refrigerators capable of reaching 15 millikelvin)

• Electromagnetic shielding

• Precision microwave and laser control systems

• Low-latency classical co-processors for error correction

• High-speed interconnects between quantum processing units and classical HPC clusters

(We just gave you the next big things, not a single writer talk about it

— We detail the winners we identified at the end of the article).

The cryogenic requirement is the most dramatic. IBM has already built what it describes as a “super fridge”: a dilution refrigerator large enough to cool quantum systems of the scale needed for fault-tolerant computation. The cooling system alone for a large quantum data center will be one of the most demanding engineering challenges in the history of computing.

But here is the critical point: many of the infrastructure investments being made today; the power capacity, the physical space, the high-speed fiber, the advanced thermal management, the land adjacent to existing data centers, are not specific to GPUs. They are general compute infrastructure that will serve whatever compute paradigm arrives next. The companies building these facilities know this.

The Quiet Strategic Pre-Positioning

Silicon Valley’s most sophisticated investors and operators understand something they do not say publicly:

Quantum computing, when it arrives at commercial scale, will require the same physical infrastructure as classical HPC: power, cooling, connectivity, physical space, but it will not require nearly as much of it to deliver equivalent or superior computational performance for the problem classes where it excels.

This is a counterintuitive point worth unpacking carefully. A fault-tolerant quantum computer with 1,000 logical qubits will, for specific molecular simulation problems, outperform a classical supercomputer cluster occupying an entire building and consuming hundreds of megawatts of power.

The quantum computer will fit in a room and consume a fraction of the power, though it will require elaborate cooling infrastructure. The comparative efficiency, measured in FLOPS-per-watt for quantum-amenable problems, will be orders of magnitude superior.

What this means for the infrastructure thesis: the companies building massive compute infrastructure today are not just preparing for the AI moment. Or they are, but it won’t end well when the market will figure out all of these capex were here for nothing.

Good ending: they are building the physical environment into which quantum computing will eventually be co-located with classical HPC.

The design of modern hyperscale data centers, modular, power-flexible, thermally managed, high-connectivity, is compatible with quantum integration in a way that older computing infrastructure is not.

There is another dimension: quantum error correction requires classical co-processors operating with extremely low latency, sub-microsecond response times, to detect and correct qubit errors in real time.

IBM demonstrated in October 2025 that a key real-time error correction algorithm can run on conventional AMD FPGAs with a tenfold speed advantage.

This means the classical HPC infrastructure being built today is the co-processor layer for tomorrow’s quantum systems.

The data centers going up today are not just for AI. They are the substrate for a hybrid classical-quantum computing future that the insiders can see coming even if the public narrative has not caught up.

The Timeline — When Does This Actually Happen?

The Four Phases of Quantum Maturity

The quantum industry commonly describes its maturity in four phases, each with distinct characteristics and commercial implications:

Phase 1: NISQ Era (Current — 2027) Noisy Intermediate-Scale Quantum computers have between 50 and 1,000+ physical qubits but lack error correction sufficient for fully reliable computation. In this phase, quantum computers are useful for research and for hybrid quantum-classical workloads where quantum advantage is achievable on specific problem types despite noise.

Key milestones:

• Google’s Willow demonstrating below-threshold error correction (achieved December 2024);

• IBM Nighthawk demonstrating improved circuit complexity (Q4 2025); multiple demonstrations of narrow quantum advantage in real-world applications.

Phase 2: Early Fault Tolerance (2027–2030) The first fault-tolerant quantum systems arrive. IBM’s target is Starling, 200 logical qubits with 100 million gate operations, by 2029. Quantinuum targets Apollo, its first fault-tolerant system, by 2030. Microsoft targets a fault-tolerant system on its topological qubit architecture on a similar timeline. In this phase, quantum computers become reliable enough for commercial applications where the economic value justifies the cost and complexity. Drug discovery workflows begin genuinely incorporating quantum simulation. Financial optimization achieves demonstrable advantage.

The talent shortage, only one qualified candidate for every three quantum positions globally, according to 2025 estimates, begins to ease as university programs scale.

Phase 3: Quantum Advantage at Scale (2030–2035) Systems with thousands of logical qubits become available. PsiQuantum’s million-qubit target, DARPA’s US2QC (Utility-Scale Quantum Computer) program targeting practical utility-scale quantum by 2033, and multiple commercial systems enable broad commercial deployment.

Drug discovery timelines begin collapsing. Materials science applications begin delivering breakthrough results, new battery chemistries, better solar cells, catalysts for green chemistry. The cryptography transition is in late execution. The economic value starts becoming legible in earnings reports.

Phase 4: Full-Scale Quantum Supremacy (2035+) Universal, fault-tolerant quantum computers with millions of logical qubits operate at commercial scale.

Blue Jay (IBM’s 1,000+ logical qubit target following Starling), and equivalent systems from Google, Microsoft, and the photonic competitors, operate in hyperscale data centers alongside classical HPC.

Artificial intelligence training runs on quantum-generated data as a matter of course. The feedback loop between quantum simulation and AI reasoning begins driving scientific discovery at rates that make the previous decade’s progress look incremental.

Prospectives, Predictions, and the Science Fiction That May Not Be Fiction

The Conservative Scenario

In the conservative scenario, the timeline slips by three to five years.

Engineering challenges (error rates, qubit coherence, scaling cryogenics, manufacturing yield) prove harder, than roadmaps actually anticipate. Fault-tolerant systems arrive in the early 2030s rather than 2029. The commercial payoff is real but gradual, concentrated in high-value verticals: pharma, finance, logistics, cryptography.

The quantum computing market reaches $150 billion by 2035. AI continues to advance on classical hardware, partially closing some of the gaps that quantum promises to address. The quantum transition is important but not as sudden as the most enthusiastic projections suggest.

This scenario is the most likely single outcome, not because the physics is uncertain, but because complex engineering schedules are reliably optimistic.

The Base Scenario

Fault tolerance arrives on roughly current schedules.

IBM delivers Starling by 2029 or 2030. Multiple competitors deliver equivalent systems by 2031. The drug discovery acceleration begins in earnest: first quantum-designed drugs in clinical trials by 2032–2033. Materials science breakthroughs in battery chemistry arrive.

The cryptographic transition is messy but largely complete for critical infrastructure by 2035. The quantum computing market reaches $500 billion by 2035, with pharmaceutical, financial, and logistics value generation leading. AI models trained on quantum-generated data begin outperforming classical AI on scientific tasks. The feedback loop between quantum and AI begins to be legible…

This scenario implies massive value creation for companies in quantum hardware, software, and adjacent infrastructure, and massive disruption for companies whose competitive moats depend on information asymmetry in molecular simulation (large pharma, agrochemicals) or on cryptographic security assumptions (every financial institution, every government).

The Accelerated Scenario

Topological qubits, Microsoft’s bet, deliver on their promise.

The engineering of Majorana quasiparticles proves tractable.

A million-qubit topological quantum computer arrives by 2033, not 2040. Simultaneously, AI-driven materials discovery, running on early fault-tolerant quantum systems, identifies a room-temperature superconductor.

This single discovery (which has been predicted for decades and never arrived) would transform energy transmission, electric motors, computing hardware, and medical imaging simultaneously.

The economic impact of a room-temperature superconductor alone is estimated in the tens of trillions of dollars over decades.

In this scenario, quantum computing does not merely accelerate existing industries. It creates new ones, industries predicated on the ability to design matter at the quantum level.

Molecular manufacturing. Quantum-designed pharmaceuticals. Materials with programmable quantum properties. The timeline to what AI researchers call AGI, a system with genuinely general reasoning capability, compresses because quantum simulation gives AI access to ground truth about physical reality…

The Science Fiction That Awaits: A Speculative Coda

Let us allow ourselves, briefly, to go where the physics and the trajectory of technological progress actually point, even when that destination sounds like a novel.

2038: The First Quantum-Designed Antibiotic

Antibiotic resistance kills approximately 1.3 million people per year today and is projected to kill 10 million annually by 2050. The pipeline of new antibiotics has essentially run dry because classical computational methods cannot accurately predict which molecules will kill bacteria without being toxic to human cells…

→ A quantum simulator with 2,000 logical qubits runs a complete quantum mechanical simulation of the ribosome of a drug-resistant bacterial strain, something that would have required more classical compute than the entire global stock of supercomputers to approximate. In weeks, it identifies three candidate molecules never previously synthesized. Within two years, one is in clinical trials.

2041: The Nitrogen Fixation Breakthrough

Quantum simulation of nitrogenase, the enzyme that performs biological nitrogen fixation, reveals the precise electronic mechanism by which it operates at room temperature. A team of chemists uses this insight to design a synthetic catalyst that performs the same reaction at 10% of the energy cost of the Haber-Bosch process.

Global fertilizer production costs fall by 60% over the following decade. The agricultural implications are planetary.

2045: The Recursive Loop

A system combining a large fault-tolerant quantum processor with a classical AI model, trained predominantly on quantum-generated scientific data, begins autonomously proposing and computationally validating scientific hypotheses at a rate no human team could match. It is not conscious. It does not have goals of its own.

But it operates as a scientific discovery engine, running thousands of computational experiments per day, each grounded in quantum mechanical accuracy. The pace of materials science, pharmaceutical development, and fundamental physics research accelerates in ways that are difficult to describe without sounding hyperbolic. We are not in the age of AI anymore. We are in the age of quantum intelligence.

This is not science fiction as escapism.

It is science fiction as extrapolation: taking the trajectory of a technology that is already delivering results (quantum simulation of molecular systems, quantum error correction, quantum-AI hybrid pipelines) and projecting that trajectory forward assuming nothing unexpected happens to interrupt it.

The unexpected things that typically interrupt technology trajectories are: physics limits (which quantum computing has largely not encountered yet), engineering failure (possible, manageable), and lack of economic incentive (manifestly absent, there is more capital chasing quantum than at any point in history).

The trail of gunpowder is already lit.

Short transition: Amara’s Law

Amara’s Law

Amara’s Law, formulated by futurist Roy Amara, states that we tend to overestimate the short-term impact of new technologies while underestimating their long-term effects. This law reminds us of the importance of considering both context and the full time horizon when forecasting technological progress.

The Investor’s Map — Where the Value Accumulates

The Quantum Value Stack

Like every deep technology, quantum computing has a stack — layers of the technology that will create value at different times and for different types of investors.

Layer 1: Hardware (2025–2035, now through fault tolerance)

This is the riskiest and potentially most rewarding layer. Pure-play quantum hardware companies. They are volatile, pre-profitability, and dependent on technical milestones being met. They also represent the largest potential upside if any one of them emerges as the dominant hardware platform.

IBM and Google have quantum hardware programs but are not pure-play bets.

The more defensive hardware play is the enablement layer: companies that supply cryogenic equipment, precision microwave and laser control systems, and specialized semiconductors for quantum control.

These are typically private companies or divisions of larger industrial conglomerates.

Layer 2: Infrastructure (2025–2030)

This is the most accessible and, arguably, the most certain quantum trade available to investors today. The same infrastructure required for quantum computing, power, cooling, space, fiber, co-processors, is being built right now, under the AI narrative. Investing in data center infrastructure, power generation and transmission (particularly for industrial-scale cooling), and advanced semiconductor fabrication is simultaneously an AI play and a quantum pre-positioning play.

The REIT structures around data centers, power utilities serving data center clusters, and semiconductor equipment companies benefit regardless of which specific quantum technology platform wins.

Layer 3: Software and Algorithms (2027+)

Quantum software companies, those building the algorithms, middleware, error correction software, and application layers that run on quantum hardware, will capture significant value but primarily after hardware matures enough to deploy reliably.

Qiskit (IBM, open-source) and Q-CTRL (private) are the leading quantum software platforms today. Companies building quantum algorithms for specific verticals, Multiverse Computing for finance, Qubit Pharmaceuticals for drug discovery, are early stage but could become highly valuable as hardware matures.

Layer 4: Application (2030+)

The ultimate beneficiaries of quantum computing will be incumbents in industries transformed by it: pharmaceutical companies that adopt quantum simulation early will develop better drugs faster. Financial institutions that implement quantum optimization will achieve superior returns. Logistics companies that implement quantum routing will have structurally lower costs.

These are traditional industry plays supercharged by a new capability, the quantum alpha accrues to the companies that use it first, before competitors can match them.

The Risk Map

Timeline risk: The history of quantum computing is littered with predictions that proved too optimistic. Every generation of quantum researchers has believed they were “almost there.”

The engineering challenges, decoherence, error rates, scaling cryogenics, manufacturing yield, have consistently proved harder than anticipated. Investors who price in the aggressive scenario will be disappointed if the conservative one materializes.

Platform risk: There are at least four distinct qubit modalities being pursued seriously (superconducting, trapped ion, topological, photonic), and it is genuinely unclear which will dominate at scale.

This is the 1970s personal computer market before IBM set the standard: multiple incompatible platforms, uncertain winner, significant risk of backing the wrong architecture.

Nationalization risk: Quantum computing is now explicitly a matter of national security for the United States, China, Europe, and a growing list of other nations. Governments are pouring money in (the US NQI invested $2.5 billion from 2019–2024; China has committed approximately $140 billion through its national venture fund), imposing export controls, and potentially restricting commercial access.

The quantum supply chain (rare materials, specialized equipment, talent) is subject to geopolitical risk in ways that previous computing generations were not.

Cryptographic disruption risk (and opportunity): The “harvest now, decrypt later” threat is real.

Companies with large amounts of sensitive encrypted data (banks, healthcare providers, defense contractors, governments) face an existential risk if they do not begin transitioning to post-quantum cryptography now.

The companies that provide post-quantum cryptography solutions (PQShield, evolutiQ, SandboxAQ) are positioned for certain near-term value creation regardless of when full quantum computing arrives.

The Geopolitical Dimension — The New Space Race

China’s Quantum Ambition

China has committed approximately $140 billion (RMB 1 trillion) through its national venture fund for quantum technology development. This is not primarily a commercial play. China’s quantum ambition is driven by national security imperatives: the ability to crack adversary encryption, the ability to build quantum communication networks that are uncrackable, and the strategic advantage that quantum simulation provides for materials science, pharmaceuticals, and AI.

China’s quantum communication program is already operational: a 2,000-kilometer quantum key distribution backbone connecting Beijing and Shanghai has been operational since 2016. Chinese researchers have published significant results in quantum computing, including demonstrations with neutral atoms and photonic systems. The geopolitical framing of quantum computing in China is explicit: it is a strategic technology, like nuclear, aerospace, and AI, and it will receive whatever resources are required.

The US Response

The National Quantum Initiative Act, passed in 2018 and renewed in subsequent years, has invested $2.5 billion in quantum through 2024, establishing Quantum Leap Challenge Institutes and the National Quantum Virtual Laboratory. DARPA’s US2QC program is explicitly targeting a utility-scale quantum computer by 2033, a government-funded moonshot with a specific deadline.

The US advantage in quantum is primarily in talent, institutional research (MIT, Caltech, Harvard, Chicago, Yale all have world-leading quantum programs), and commercial ecosystem (IBM, Google, Microsoft, and the rich startup landscape).

The US Chips and Science Act includes provisions for quantum workforce development. The White House has accelerated quantum policy, including executive actions focused on federal adoption and the migration to post-quantum cryptography.

Europe’s Quantum Flagship

The European Union’s Quantum Flagship program, with a €1 billion commitment, coordinates quantum research across member states. Europe’s strongest players include Quantinuum (UK), IQM (Finland), Pasqal (France), Alice & Bob (France), and Quandela (France). France in particular has made quantum computing a national priority, with significant government co-investment in French quantum startups.

The United Nations designated 2025 as the International Year of Quantum Science and Technology, a signal that the global scientific and policy community views this technology as the defining technical challenge of the coming decade.

The Questions That Matter Most

Can We Scale Error Correction?

The central technical question in quantum computing is whether error correction can be made efficient enough that the overhead, the ratio of physical to logical qubits required, becomes manageable at scale.

IBM’s qLDPC (quantum low-density parity-check) codes, announced in 2025, claim to reduce error correction overhead by approximately 90% compared to conventional approaches. If this scales, the number of physical qubits required for 200 logical qubits might be 20,000 rather than 200,000. That is still an enormous number, but it changes the engineering calculus dramatically.

Microsoft’s topological qubit approach claims to reduce error rates inherently, by making the qubit encoding itself more resilient, rather than relying as heavily on software error correction. If Majorana qubits prove stable at scale, the overhead of error correction drops dramatically.

Alice & Bob’s cat qubit approach achieves the same result differently: by biasing the error model so that one type of error (bit flips) is suppressed by several orders of magnitude, only one type of error (phase errors) needs to be corrected. Simpler correction means lower overhead.

All three approaches represent genuine architectural innovation. The question is which one, if any, scales beyond the laboratory. The answer is almost certainly that multiple approaches will find their niches: superconducting for maximum clock speed and classical-quantum integration; trapped ion for highest fidelity per qubit; photonic for distributed quantum networks; topological for maximum qubit density when the materials science matures.

Will Quantum Compute and AI Actually Converge?

The short answer is: yes, the physics demands it.

The longer answer is: the timeline and mechanism are uncertain.

The clearest near-term convergence is quantum-generated training data for classical AI, Quantinuum’s Gen QAI framework is the first commercial example. This requires relatively modest quantum capability (achievable today on current hardware for narrow problem types) and delivers immediate AI value.

The medium-term convergence is quantum co-processors for AI training and inference, using quantum algorithms to accelerate specific computations within an otherwise classical ML pipeline. This requires fault-tolerant hardware and will likely emerge commercially in the 2030s.

The long-term convergence, quantum computing enabling genuinely general AI by providing accurate simulation of the physical world, is the thesis most responsible for the “matchstick” framing. It is also the one most distant in time and most dependent on the full quantum stack being realized. But the logical chain is sound, and the directional evidence is accumulating.

Now, we will see who will actually win, according to us. We identified around 20 names. We also give you all the specifics terms/words to know, the key numbers of the revolutions and key ressources to read.

Who Will Win This Revolution? How To Win As An Investor?

This is the question investors most want answered, and it is the one that is genuinely uncertain. The most honest prediction:

• IBM wins the enterprise ecosystem play. Its developer community, partner network, and roadmap clarity make it the most reliable bet for broad commercial deployment when fault tolerance arrives. But it is not the only one…

Read more

Welcome in this seventh edition of Future Cognitive Capital.

The best newsletter to talk about future cognitive business that will represent the future of data driven compounding.

Yes, that a term I invented : Data compounder & Data Moat

Here’s the company we will going through today:

As editions go by, you get sharper on these subject an so, it become easier to understand what we are talking about and so, I can go straight to the point.

However Quantity < Quality so expect a very interesting article since we we didn’t post any new file since November 16th.

The company of today is the perfect example of misunderstanding data compounder, I have been advocating this company since months but the market got it wrong; is the perfect example of a massive proprietary data base, multimodal architecture and other interesting notions we break down here.

I am super proud of this analysis and the opportunity it represent as the company was a bit bitten down recently….

If you want more than a nice story and actually want to invest in the small group of companies building the cognitive infrastructure of the economy, this is what Future Cognitive Capital is for. Inside you get a concentrated, real-money portfolio (+350% since January 2024 with a handful of names), full 50-point FCC scores on each company, weekly Signal Updates on the AI/data/compute stack, and access to the 200-page “AI as an Investment” handbook that sits behind all of it. I cap FCC at 100 paying subscribers and raise the price as the research library grows, so if you want a seat in this lab instead of reading from the outside, you can join now.

• Data As An Investment, Chapter II, out 15th of January.

Stops right here the teasing, begins right here the thesis.

Read more

You know how it works, and so, I will be more focus on insights, and try not to be too much tech oriented. If you struggle to understand why I will say some things, I highly invite you to read the last two posts to get some context.

Enjoy this third F2C company file.

Read more