Frontiers of AI

The trends discussed so far describe transformations already underway: systems in production, regulations in place, organizations adapting. This section, however, deals with a different dimension. The six trends it encompasses are not limited to the operational present; they delineate the strategic space that already shapes investment decisions, positioning and sovereignty, even if their full effects are still unfolding. 

AI geopolitics is redefining alliances and dependencies. AI-first organizations are anticipating unprecedented competitive models. Digital twins and advanced simulation are transforming how we design, experiment and decide. Ambient AI is blurring the boundary between environment and computation. Convergence with quantum computing is expanding the horizon of solvable problems. And AGI has ceased to be mere academic speculation to become an explicit strategic hypothesis in the world's leading laboratories.

Geopolitics and Technological Sovereignty of AI

AI as a strategic state asset

AI is no longer a sectoral technology but a strategic infrastructure comparable to energy, telecommunications or financial systems. What is at stake is not only economic competitiveness: it is the ability of states to maintain autonomy in critical decisions, from defense to financial supervision, to the management of essential infrastructures. In this context, control of foundational models, advanced semiconductors, data centers and specialized talent has become a major national power factor, and industrial policies, export controls and investment strategies already reflect this new reality.

The strategic value chain: where sovereignty is at stake

Technological sovereignty in AI is not an all-or-nothing situation: there are layers to it, and dependence can arise in any of them:

• Hardware: TSMC produces over 90% of the world’s most advanced chips, ASML is the sole global supplier of the EUV lithography needed to make them, and NVIDIA controls more than 85% of the GPU market for frontier AI model training. Three companies, three countries, one structural bottleneck.

• Infrastructure and models: three hyperscalers (AWS, Azure, Google Cloud) concentrate roughly two-thirds of the world’s total compute capacity; on top of it, OpenAI, Anthropic and Google DeepMind develop the most capable foundational models, with cumulative advantages in data, talent and investment that are hard to replicate.

• Talent: frontier research remains geographically concentrated, with international mobility turning migration policy into technology policy.

No one country or organization controls all these layers simultaneously. The strategic question is not how many layers are controlled, but which are mission critical and which are manageable through diversification, agreements or redundancy.

Three competing models

The United States, China and Europe have articulated structurally distinct responses, which are not mere differences of emphasis but geopolitical projects with profound consequences for global alliances, markets and standards.

The United States combines the primacy of the private sector in model development with increasing state intervention in the supply chain: the "CHIPS and Science Act" mobilizes tens of billions in domestic semiconductors, and export controls on advanced chips to China represent the greatest technological restraint among major powers since the Cold War. The strategy is clear: maintain the edge in frontier models and deprive competitors of the hardware needed to achieve it. 

China combines massive state investment, civil-military integration and an explicit strategy of technological self-sufficiency and comprehensive control of the digital ecosystem. Its stated goal is self-sufficiency across the entire value chain by 2030, from semiconductors to proprietary foundational models. U.S. export restrictions have accelerated this agenda: DeepSeek demonstrated in 2025 that Chinese innovation can produce competitive models with previous-generation hardware, complicating the logic of containment through chip control. 

Europe has staked regulatory leadership as a vector of geopolitical influence. The AI Act and GDPR have generated a real "Brussels effect": global companies adapt their products to European standards because the European market is too big to ignore. However, Europe maintains significant infrastructural dependencies: its most advanced foundational models are American, its cloud is largely foreign, and its sovereign computing capacity is limited. ASML is the notable exception: the Dutch monopoly in EUV lithography makes Europe an indispensable player in the global semiconductor chain. 

The rest of the world navigates between these three poles with very uneven capabilities. Some emerging countries are articulating strategies of their own with growing ambition: India is developing foundational models in local languages and negotiating its position in semiconductor supply chains; Brazil is leading AI governance initiatives in Latin America; the African Union is advancing continental frameworks for digital sovereignty. However, for most countries the choice between incompatible ecosystems remains more implicit than deliberate, with real risks of structural dependency that the literature calls "data colonialism": the mining of local data to train models that are deployed globally, without source countries capturing value or maintaining effective control over their digital infrastructure.

Fragmentation and risk of decoupling

The world is moving toward partially incompatible technology ecosystems. Chip export controls, geo-fenced data, divergent technical standards and parallel infrastructures are shaping what some analysts call "technoblocks": spheres of technological influence with their own governance, security and value logics. The Chip 4 alliance (US, Japan, Taiwan, and South Korea) coordinates Western semiconductor strategy; China cultivates its own sphere through the Digital Silk Road. A complete decoupling between the US and Chinese ecosystems would force third countries and organizations to choose, with potentially prohibitive transition costs.  

The limits of absolute sovereignty

Total self-sufficiency in AI is economically unfeasible for most countries and organizations. Recreating from scratch the entire value chain (from mining critical minerals to developing foundational models) requires investments and scales that only two or three global players can sustain. Realistic technological sovereignty is not isolation: it is the effective ability to decide, diversify suppliers, negotiate terms and avoid strategic lock-in at the truly critical layers. For most countries, AI sovereignty in practice boils down to controlling the only asset over which they have effective jurisdiction: the data generated in their territory.

Implications for organizations

The geopolitical debate lands in concrete corporate decisions. Reliance on a single foundational model provider exposes organizations to risks of lock-in, changes in pricing or terms of service, and even regulatory restrictions arising from tensions between jurisdictions. Multi-jurisdictional data localization and regulatory compliance add further layers of operational complexity. 

Multi-model and multi-cloud strategies, previously justified on technical performance and cost grounds, now take on an additional strategic dimension: they are the organizational equivalent of diversifying sovereign dependencies.

AI-First and AI-Only Organizations

Definition and taxonomy

The expansion of agentic systems raises a question that until a few years ago was theoretical: can an organization function with AI as its central cognitive architecture, with human labor being the exception? To answer it precisely, it is useful to distinguish three stages that business practice often confuses:

• An AI-enhanced organization uses AI to improve existing processes; this is the predominant model today.

• An AI-first organization designs its processes and structure based on AI capabilities, assigning to human judgment only those tasks where its comparative advantage is clear.

• An AI-only organization operates core functions entirely without human labor: No consolidated examples exist as of this writing, and its viability in regulated environments remains a working hypothesis.

The state of the art: AI-first as an operational frontier

The most advanced examples of AI-first organizations come, significantly, from the pure technology sector, where the absence of regulatory constraints on automation and the digital nature of the product allow the model to be pushed to its current limits.

Midjourney, the AI image generation platform, earned over $500 million in revenue in 2025 with a staff of approximately 163, no marketing investment and no external funding. Development platform Cursor (Anysphere) reached $500 million in annual recurring revenue in May 2025, making it the fastest growing SaaS company in history with less than 50 employees. The revenue per employee ratio of these companies (over $3 million in both cases) is roughly 10 times higher than historical benchmarks for the technology sector and large global banking groups.

In the field of financial services, an advanced case is MYbank, a Chinese digital bank owned by Ant Group, which since 2015 has been operating under the principle of zero human intervention in SME credit approval. Its "310" model – three-minute application, one-second approval, zero human intervention – has served more than 50 million SMEs. The system uses cash flow forecasting models with over 95% accuracy and relies on satellite geolocation data for agricultural risk assessment. MYbank operates without a branch network or sales force, although it maintains engineering and management teams: it is an AI-first model in its core operations, not AI-only as a whole.

Why is there still no AI-only organization?

The gap between AI-first and AI-only is not just technological; it is also regulatory, legal and organizational. In regulated sectors, such as banking and insurance, current regulations require oversight and human responsibility for material decisions. The European AI Act classifies AI applications in credit, healthcare, life insurance, and critical infrastructure as high-risk systems, with explicit requirements for human oversight. Eliminating this oversight at the operational core of a financial institution would be incompatible with the European prudential framework.

In unregulated sectors, the current limit on AI-only operations is not regulatory but technical and organizational. Autonomous AI agents can execute complex tasks, but their error rate in extended workflows, inability to handle situations not contemplated in training and the absence of legal liability mechanisms equivalent to those of a legal entity mean that fully eliminating human labor from core operations would introduce operational risks that are not yet manageable. 

The case of Klarna illustrates the current limits: the company reduced its workforce from 7,400 to approximately 3,000 between 2022 and 2025 through hiring freezes and extensive automation, with one AI assistant managing the equivalent of 853 employees in customer service. Their trajectory helps define the practical threshold between tasks that AI can perform autonomously with sufficient quality and those where human judgment continues to provide differential value today.

Predictions from industry leaders

The heads of leading frontier AI labs increasingly forecast that AI-only organizations may be on the immediate horizon. Sam Altman, CEO of OpenAI, stated in 2024: "We're going to see ten-person companies with billion-dollar valuations very soon [...] There's a bet in my chat group of executive friends about when the first one-person company valued at a billion dollars is going to exist, which would have been unimaginable without AI. And now it's going to happen". Asked in May 2025 about when that scenario would materialize, Dario Amodei, CEO of Anthropic, replied, "2026".

Amodei develops the argument in his January 2026 paper, where he describes the functional equivalent of "a country of geniuses in a data center," i.e., 50 million agents more capable than any Nobel Laureate, operating at between ten and one hundred times human speed; and he estimates that 50% of entry-level jobs could be disrupted within one to five years. The same paper notes that Anthropic already runs most of the code it produces using AI, approaching full operational autonomy in software development.

Can existing organizations become AI-only?

The underlying strategic question is not whether AI-only organizations will exist, but how they will come to exist. The answer is counterintuitive: they are unlikely to emerge from the transformation of existing organizations. Clayton Christensen documented in The Innovator's Dilemma that incumbent companies are structurally incapable of adopting disruptive technologies from within: their processes, incentives and customer bases are optimized for the incumbent model, and any internal disruptive initiatives compete at a permanent disadvantage for resources and management attention. The transition to AI-first exacerbates this logic: an organization with tens of thousands of employees has its processes designed for that human scale. Those processes are not redesigned; they are replaced.

The pattern emerging in Asia points to an alternative path: the creation of new entities, with their own brand and no operational heritage, that compete freely until they reach critical mass and cannibalize the original company. Ping An, the world's largest insurer by premiums written, incubated between 2013 and 2022 eleven independent technology subsidiaries (including OneConnect, Lufax and Ping An Good Doctor), five of which were listed as standalone entities. DBS Bank created Digibank as a separate digital bank operating with one-fifth of the resources per customer of a conventional bank, with its lessons feeding back into the parent organization’s architecture . The mechanism is identical: a new, non-legacy entity that scales without the constraints of the parent organization and, if the experiment fails, is closed without dragging down the original company.

In Europe and, to a lesser extent, in the United States, this mechanism encounters structural frictions that go beyond AI regulation. In Europe, the introduction of workplace AI systems may require consultation or negotiation with works councils, and in some countries their explicit agreement. Collective bargaining agreements in employment-intensive sectors incorporate clauses limiting automation. Employee data protection under GDPR adds additional complexity. The result is an asymmetry with unintended strategic consequences: western regulation makes it difficult for existing organizations to build the AI-first entities that would eventually challenge them.

Once technology reaches the point where an AI-only organization is viable, the question of who will build it first will likely come down to geography.

Digital Twins and Simulation of Human Behavior

From Aerospace Engineering to Universal Simulation

The concept of the digital twin has a precise date and place of birth. In October 2002, Michael Grieves presented at a forum of the Society of Manufacturing Engineers what he called "Conceptual Ideal for Product Lifecycle Management": the idea that any physical object could have a digital correlate that would dynamically represent it throughout its life cycle, synchronizing in real time the state of the real object with its virtual representation. The term "digital twin" was later coined by John Vickers, NASA's chief engineer, who formalized the concept in the agency's 2010 technology roadmap. NASA's definition in that document remains the most accurate available: "a multiphysics, multiscale, probabilistic simulation of a vehicle or system that uses the best available physical models, sensor updates, and fleet history to mirror the life of its physical twin."

The starting point is important because it reveals the implicit premise that has guided the development of digital twins for two decades: a digital twin works well when the system it models obeys known, deterministic physical laws. A gas turbine, an aircraft fuselage, an electrical grid: complicated systems with many components, but in principle fully modelable if sufficient computational power and sensor data are available. Under that premise, the technology matured steadily. Today, digital twins of physical assets are operational in fields such as advanced manufacturing, energy, infrastructure and aviation, delivering measurable reductions in unplanned maintenance times and significantly accelerating product design cycles.

The epistemological boundary: complicated systems versus complex systems

Expanding the concept beyond the physical domain has revealed a limit that is not technological but epistemological. Michael Batty, the most recognized academic authority on computational modeling of cities, explains it precisely: digital twins work in complicated systems (many parts, but determinable behavior in principle) and encounter structural difficulties in complex systems, where the overall behavior emerges from the interaction of agents and cannot be deduced from the properties of their individual components. A city, an economy, a financial market, or a human organization are complex systems in this precise technical sense.

Philosopher Stefano Moroni expresses the argument in even more direct terms: the limitations of urban digital twins are not temporary (they will not disappear with more data or greater computational power), but instead arise from the intrinsically emergent nature of social systems. The unpredictability of detail in a complex system is not an information deficit; it is a property of the system. This has immediate practical implications: a digital twin of a manufacturing plant can reliably predict when a bearing will fail; a digital twin of a city can approximate aggregate traffic trends, but cannot reliably predict the effect of a housing policy on ten-year residential segregation patterns. The distinction is not one of degree, but of nature.

This epistemological boundary has defined the ceiling of the digital twin field for decades. It is precisely here that large-scale language models introduce a discontinuity that deserves attention.

The tipping point: human behavior becomes modelable

In April 2023, a team of Stanford researchers published a paper that launched a radically new line of work. Joon Sung Park and his coauthors created 25 computational agents (each endowed with an identity, a persistent memory, a set of social relationships, and a reasoning ability based on a language model), and placed them in a simulated environment equivalent to a small city. The agents woke up, ate breakfast, went to work, formed opinions, initiated conversations and coordinated collective activities without these behaviors having been explicitly programmed: they emerged from the interaction between each agent's individual memory, their ability to reflect on past experiences and their model of the social environment. The paper won the Best Paper Award at the 2023 ACM Symposium on User Interface Software and Technology. The scientific community recognized that something qualitatively new had happened.

The underlying reason is that large-scale language models have absorbed, during training, an extraordinary amount of recorded human behavior: conversations, decisions, reasoning patterns, emotional responses, implicit social norms. They have not learned the laws of human behavior explicitly (no one knows them with that precision), but they have developed a statistically dense approximation that, under controlled conditions, generates plausible behaviors. For the first time, the premise that prevented modeling complex social systems has been partially lifted: not because human behavior has ceased to be emergent, but because there is now a behavior generator rich enough to populate a simulation with credible agents. 

The natural extension of this work came in November 2024. The same Stanford team published the results of a different scale experiment: 1,052 real people, interviewed in depth about their lives, attitudes, and experiences, were turned into agents that replicate their responses and behaviors in standardized surveys and social experiments. The generative agents replicated real individuals' responses on the General Social Survey with 85% accuracy (statistically comparable to the individual's own natural variability when answering the same survey two weeks later) and produced comparable results in replicating personality traits and in social science experiments. What began as a conceptual demonstration with fictitious characters in 2023 became, by 2024, an empirically validated methodology with real people.

Current applications and future horizon

The implications of this leap are cross-sectoral. In market research, the startup Simile – founded by Joon Sung Park together with Michael Bernstein and Percy Liang, the co-authors of the founding paper, and backed in February 2026 with $100 million by Index Ventures with participation from Fei-Fei Li and Andrej Karpathy, builds digital twins of real people to help companies simulate their customers' behavior before launching a product, modifying a pricing policy or redesigning a user experience. In a public demonstration, the platform correctly predicted eight out of ten answers to the questions asked by analysts in a simulated call. The global market research industry, valued at $142 billion, is facing structural disruption: what today requires weeks of fieldwork can be executed in hours on synthetic populations. 

In public policy and urban planning, a more nuanced but equally transformative use is emerging: not the digital twin as a predictive oracle, but as a scenario laboratory where the consequences of different interventions can be explored before committing real resources. In financial regulation, this approach has direct application in the stress testing of adverse macroeconomic scenarios and in the simulation of market behavior in the face of regulatory interventions. 

The trajectory of this domain suggests a qualitatively different scale ahead. If today it is possible to simulate a thousand real people with high fidelity, the question on the immediate horizon is what happens when that number reaches a million, a hundred million, an entire society modeled in real time. The applications in public policy, economic regulation and institutional design will be of a different order of magnitude than market research: not anticipating what product a consumer will buy, but predicting how a population will respond to a tax reform, a health crisis or a change in monetary policy before that intervention is implemented in the real world. This capability has no historical precedent, nor, as yet, governance framework to regulate it.

Ambient AI and Invisible Computing

The interface is the environment

Ambient AI – or ambient intelligence – is AI that operates without being invoked. Unlike conventional systems, which respond to an explicit instruction from the user, ambient systems continuously observe the context, infer needs and act proactively. The interface disappears not because it has been improved, but because the system no longer needs it: the environment itself becomes the point of interaction. Computing becomes "invisible" in the literal sense: embedded in objects, spaces and processes without the user perceiving it as such.

This inversion (from the user going to the system to the system coming to the user) is made possible today by the convergence of three simultaneous developments: the miniaturization of models capable of running on edge devices while maintaining a continuous state (locally updated cumulative user memory) without relying on cloud connectivity (edge AI and TinyML), the densification of physical and biometric sensor networks, and the ability of LLMs to reason about heterogeneous and ambiguous context in real time. None of the three is new on its own; their simultaneous maturity is what makes Ambient AI capable of moving from concept to operational deployment. 

Current state of deployment

The most documented example of Ambient AI in operation are ambient AI scribes in clinical settings: systems that continuously listen to the patient-physician conversation, infer the clinical context without explicit instruction, and automatically generate encounter documentation. A UCLA randomized clinical trial evaluated two platforms – Microsoft DAX and Nabla – across 238 physicians from 14 specialties and over 72,000 encounters: it reduced documentation burden and improved indicators of professional burnout.

The system was not invoked iteratively during the consultation: it listened, inferred, wrote. It is still a bounded form of ambient intelligence (limited context, clear purpose, defined episode). Mature Ambient AI will operate beyond individual consultations, at the scale of the entire hospital, correlating longitudinal patterns with no user-determined start or end point.

The physical and digital future

Today's deployments are the tip of a broader transformation. In the coming years, Ambient AI will extend to physical and digital environments and make current cases look rudimentary:

Physical environments

Adaptive workspaces. The environment infers the occupant's attentional state from heart rate, rhythm variability, movement patterns, and reconfigures temperature, light, and noise level to optimize cognitive performance without conscious user intervention.

Anticipatory industrial maintenance. Systems will not alert when equipment fails: they will detect the behavioral pattern preceding the failure early enough to reorganize production. The disruptive event disappears from the operating horizon. 

Individually profiled wearables. Next-generation devices will not compare the wearer's vital signs to population averages, but to his or her own physiological history. The alert will be triggered before the symptom becomes conscious to the wearer.

Reactive urban infrastructure. Transportation, lighting, and waste management networks that self-adjust in real time to inferred usage patterns, without explicit centralized planning or human intervention in the loop.

Invisible home care. Systems that continuously monitor elderly people or those with chronic conditions, detect anomalies in routines (sleep patterns, mobility, feeding) and activate alert or intervention protocols without the user having requested anything.

Digital environments

Development environments that anticipate the problem. Proactive programming assistants will evolve into systems that, before the developer identifies the error, will have mapped out the likely solution space and presented options at the cognitively appropriate time.

Attention management, not just information management. Systems will not serve information when it is available, but when the user is in a position to process it: modeling attentional state throughout the day and calibrating the timing of interruption.

Continuous organizational context. Systems will know at all times the status of projects, pending communications and ongoing decisions, and will surface relevant information to each team member without being requested.

Autonomous resource negotiation. Environmental agents will manage tasks on behalf of the user (scheduling, budgeting, arranging access to services) within defined parameters, without requiring explicit approval for every low complexity decision.

Implications that technology does not solve

Ambient AI does not only raise privacy questions; it introduces a broader set of tensions that current governance frameworks have not resolved.

The first is the nature of error. In an invoked system, error is visible: the user asked for something, the system responded badly. In an ambient system, the error may not be perceived because there was no explicit request against which to compare the response. The UCLA ambient scribe recorded clinically significant inaccuracies in a proportion of encounters: in an invisible system, the error detection mechanism has to be deliberately designed, because it does not emerge naturally from the interaction.

The second is the asymmetry of power between those who design the environment and those who inhabit it. In a hospital, an office or a public building, users do not choose whether the environment is intelligent: they inhabit a space whose inferences about their behavior have been configured by a third party. Tshilidzi Marwala, Chancellor of the United Nations University, puts it precisely: Ambient AI has an appetite for data – intimate, behavioral, biometric – that renders conventional notions of informed consent structurally inadequate. The European AI Act, designed for invoked systems with bounded functions, does not provide a satisfactory response to these continuous observation environments. 

The third is cognitive dependency. A system that proactively manages the user's attention, information flow and interruptions not only assists its work: it shapes its cognitive architecture. The question posed in 2003, "is context-aware computing taking control away from the user?" has gone unanswered for decades. The scale at which Ambient AI poses it today turns what was an academic question into a design issue with immediate operational consequences. 

The fourth tension is causal accountability. In invoked systems, traceability is relatively straightforward: there is an instruction, a response, an attributable decision moment. In environmental systems, the causal chain is blurred. If an anticipatory maintenance system reorganizes production and that reorganization conditions subsequent human decisions, the boundary between technical agency and human agency is not clear. Current regulation, including the AI Act, assumes predictable finality and ex ante risk assessment; Ambient AI introduces emergent finality and continuous adaptive behavior, directly challenging existing compliance mechanisms. 

Ambient AI doesn't just change how we work or how we take care of ourselves: it changes the sequence between need and awareness. A system can know what we need before we do. The possibility that this capability may be deployed at the scale that the field’s trajectory suggests raises questions that reach beyond technology and regulation to something more fundamental: what it means to make our own decisions in an environment that has already anticipated them.

Interaction between AI and Quantum Computing

Two distinct technologies, an intersection that matters

AI and quantum computing are independent technologies with completely different principles, time horizons and use cases. AI is already operational on an industrial scale; quantum computing is still, for the most part, an advanced research field with very limited deployments. Their interaction, in both directions, has concrete implications for any organization that relies on digital systems.

A classical computer solves problems by testing options one at a time or in parallel, but always within a space of possibilities that grows in a manageable way. Some problems exceed this capacity: optimizations with thousands of interdependent variables, simulations of molecular systems, or certain mathematical problems that form the basis of modern cryptography. A quantum computer works in a fundamentally different way: instead of trying options one by one, it can simultaneously explore a space of possibilities of a dimensionality that no classical system can represent. For this specific class of problems – not all problems – the performance difference is not incremental, but many times greater than what classical computers can achieve.

The problem is that building a quantum computer that works reliably has proven extraordinarily difficult. Quantum information is extremely sensitive to environmental perturbations – temperature, vibrations, electromagnetic interference – and errors accumulate rapidly. For decades, the field advanced much faster in theory than in hardware. That partially changed in December 2024. 

The 2024 milestone and what it means

Google published results from its Willow processor in Nature. Willow is the first system to show that as more computational components are added, errors decrease rather than increase. This outcome had been theoretically predicted since 1995, but no system had previously managed to realize it. The significance lies not in the performance figures (which are impressive though based on artificial benchmarks), but in what it means for the trajectory of the field: the obstacle that for thirty years had prevented scaling these systems reliably has been overcome in the laboratory.

The gap between that achievement and a quantum computer with commercial applications remains considerable. Google's own researchers put that horizon at around the end of the decade. But the direction is no longer in dispute: the central problem was in error correction, and that problem now has a proven solution. What remains is engineering for scale, not a scientific leap in a vacuum.

Three ways this affects AI

The intersection of quantum computing and AI operates on three distinct planes, with different urgencies.

The first is the acceleration of machine learning. Training a large-scale AI model is essentially a mathematical optimization problem: finding the values of billions of parameters that minimize prediction error across an enormous search space. This is precisely the kind of problem where quantum computing offers a theoretical advantage. It has been formally demonstrated that fault-tolerant quantum systems could substantially speed up the training of large AI models, reducing both computational time and energy use. Achieving this requires hardware that does not yet exist at sufficient scale. When it does becomes available, however, it could radically alter the economics of AI model training, which is currently dominated by organizations able to afford massive-scale GPU infrastructure. 

The second plane is quantum machine learning itself: using quantum processors to run ML algorithms more efficiently. Here the literature is more cautious and describes both the promises and the actual obstacles. Quantum computers do not automatically outperform classical systems in all learning tasks, and the advantage they offer is far from universal. In many cases, classical systems with access to data can perform as well as quantum systems, even on problems specifically designed to favor quantum approaches. In other words: data, used well, can offset the quantum advantage in many situations. The hype about quantum AI as a universal accelerator is ahead of the evidence; real applications will be problem-specific rather than broadly transformative.

The third plane flips the perspective on impact: instead of quantum computing serving AI, it poses a threat to the security infrastructure that underpins all digital systems, including AI systems. All the cryptography that protects digital communications today (banking transactions, identity authentication, secure channels between systems) rests on mathematical problems whose difficulty is assumed to be unassailable for classical computers. A sufficiently powerful quantum computer would solve them directly. This plane is the most urgent, because some of its effects are already emerging.

The threat that is already active

The strategy known as "harvest now, decrypt later" consists of capturing encrypted communications today with the intention of decrypting them when quantum computing matures sufficiently. State actors with advanced intelligence capabilities have been applying it for years. Data requiring confidentiality for decades (medical records, trade secrets, regulatory communications or sensitive financial information) are being compromised now, regardless of when the quantum system capable of decrypting them arrives.

The most rigorous institutional response is that of the US NIST, which in August 2024 – after eight years of work and more than eighty proposals from research teams around the world – published the first three standards for quantum-resistant cryptography. The new algorithms are based on mathematical structures for which no efficient quantum attack is known. NIST urges organizations to begin migration immediately; the deadline for U.S. federal systems is 2035. For financial entities with long-lived data, that deadline is not the point for starting the transition: it is the final limit for completing the migration.

The horizon of the next decade

Over the next few years, the intersection between AI and quantum computing will shift from being a technology foresight topic to an element with operational impact on two simultaneous fronts.

The offensive front (the acceleration of AI capabilities) will come with the maturity of fault-tolerant hardware, predictably towards the end of the decade. Initial access will be via cloud services, replicating the trajectory that high-performance computing followed with GPUs: first accessible only to the best-funded, then democratized through competition among vendors. Organizations that have developed AI competencies by then will be better positioned to take advantage of that acceleration when it arrives.

The defensive front (cryptographic migration) admits no delay. The readiness window narrows as quantum processors scale, and cryptographic infrastructure migration in complex organizations takes years. Inventorying vulnerable assets, prioritizing by data lifetime, and transition planning are tasks that must begin now, not when the relevant quantum computer is operational and it is too late.

Artificial General Intelligence (AGI) as a Strategic Horizon

What is AGI, and has it arrived yet?

The term "artificial general intelligence" (AGI) first appeared in 1997, at a nanotechnology conference in Palo Alto. Its author, Mark Gubrud, a doctoral candidate at the University of Maryland, was not describing a desirable technological goal: he was warning of a risk. In his paper "Nanotechnology and International Security" , Gubrud defined AGI as systems capable of "rivaling or surpassing the human brain in complexity and speed, acquiring, manipulating and reasoning with general knowledge, and being usable in any phase of operations where human intelligence would be required." The definition went unnoticed for nearly a decade, until Ben Goertzel and Shane Legg rescued it and popularized it as a technical label in titling their collective book Artificial General Intelligence. The term had been born as a cautionary tale, but became a mission.

In February 2026, Nature published almost simultaneously two texts crystallizing arguably the most relevant debate in contemporary technology. The first, signed by four UC San Diego academics - specializing in philosophy, machine learning, linguistics and cognitive science – states unambiguously that AGI already exists: today's LLMs pass the Turing test, win gold medals in mathematical Olympiads and collaborate with humans in proving theorems. The second, published a fortnight later as correspondence in the same journal, replies that this conclusion is only possible by redefining the concept beyond recognition: the classical definition of AGI formulated in 2007 requires robustness under novelty, transferable generalization and goal autonomy, and current systems do not meet these requirements. The fact that top researchers, with access to the same systems and data, reach opposite conclusions reflects that "general intelligence" is a continuous concept without precise thresholds.

Fei-Fei Li herself, who built the foundation for modern computer vision and works in spatial intelligence, bluntly admits as much: "I struggle with this definition of AGI, to be honest". Dario Amodei, CEO of Anthropic, goes further: he openly declares that he does not like the term, preferring to speak of "powerful AI": systems with intellectual capabilities comparable or superior to those of a Nobel Prize winner in most disciplines.

That is the right perspective for organizations. The strategically relevant question is not philosophical – have we reached AGI – but functional: when can a system fully autonomously perform complete cycles of high-cognitive value work in all domains? That threshold has already been crossed in several sectors.  

The current state: simultaneous brilliance and fragility

Andrej Karpathy, co-founder of OpenAI and former head of AI at Tesla, coined the concept of "jagged intelligence" to describe the current condition of LLMs: systems that solve mathematical Olympiad problems and fail to determine which number is higher, 9.11 or 9.9; that are fluent in dozens of languages and suffer from what he calls "anterograde amnesia": inability to consolidate learning between sessions. Fei-Fei Li describes them as "wordsmiths in the dark: eloquent but inexperienced, knowledgeable but ungrounded": eloquent but without embodied experience in the physical world.

And yet these same systems already write contracts, analyze credit risk, synthesize scientific literature, generate and audit complex code, or produce regulatory summaries. They do this autonomously, at a speed and scale that no human team can match. The question is no longer when that capability will arrive; it is what we do with what has already arrived and how we prepare for what comes next.

The causal chain: what's next

The ongoing transition follows a logic of cumulative escalation. The first stage is from tool to agent: systems move beyond responding to instructions and instead pursue goals through autonomous cycles of action, observation and correction. The second stage is from agent to environmental infrastructure. As Karpathy accurately puts it: LLMs are the new electricity. AI ceases to be a technology that we "adopt" and becomes a technology that "happens to us": invisible operating fabric of the systems we use, a condition of the environment rather than a tool in it.

The third stage is the one most underestimated by conventional analysis: the recursive loop. Models are already used to improve other models – generating synthetic training data, optimizing architectures, or producing research hypotheses. This creates a dynamic in which the speed of AI improvement depends on AI intelligence itself. Amodei calls it "the end of the exponential": not the point where the curve flattens, but where the acceleration becomes over-exponential, because the accelerating agent is the system being accelerated.

The absorption gap

Two curves advance at radically different speeds. The technical curve – exponential, self-accelerating through the recursive loop – compresses into years what used to take decades. The organizational absorption curve – process redesign, role reconversion, governance infrastructure building, institutional change management – advances more slowly, with considerable friction: legacy systems, organizational difficulties, cultural resistance, late-arriving regulation, or shortage of talent capable of integrating these capabilities into real operations.

The gap between the two curves is the determining variable of the next decade. Competitive advantage will not come from access to the best models, which will become progressively commoditized, nor from their cost, which will continue to fall exponentially. It will come from the speed and rigor with which an organization is able to redesign itself to operate with autonomous agents effectively and responsibly. This pattern is empirically documented: the most significant productivity gains do not appear where AI replaces tasks, but where it reorganizes entire processes and redefines human-machine collaboration.

Redefining the human role

The right question, therefore, is not which jobs will disappear, but what does a human do that an AI system cannot do even though it is faster, cheaper, and more consistent. The usual answer (creativity, empathy, leadership) is true but insufficient. There are dimensions that conventional analysis underestimates:

― Responsibility with real consequences. AI systems cannot be taken to court or lose a reputation built over decades. In financial, healthcare, legal and regulatory environments, human presence is a structural requirement.

― Certification and public faith. The notary who certifies, the doctor who signs, the auditor who countersigns: these acts are worthy not because of the information processing involved, but because of the personal and institutional responsibility behind them.

― Interpersonal relationship. There are contexts where what is needed is not the right answer but the presence of another person: grief, conflict, care. Replacing that presence with a more efficient system does not solve the problem.

― Formulation of worthwhile questions and moral judgment. When AI performs well at what it is asked to do, the value shifts to those who decide what to ask: who sets the goals, who identifies which problems deserve attention or who decides when there are conflicting values.

― Democratic legitimacy. Decisions affecting communities require deliberation and accountability that cannot be delegated to opaque systems, however precise they may be.

What these dimensions have in common points to something deeper than a distribution of tasks. Throughout history, humans have simultaneously been the agents who act and the subjects who are accountable for the outcomes of those actions. AI systems structurally dissociate this unity between action and responsibility. What is redefined, then, is not just what we do, but our position in the causal chain: less in execution, increasingly in intention, judgment, and accountability.

The real systemic risk: concentration

The emerging systemic risk is not that of the machine that rebels. It is that of the unprecedented concentration of cognitive capacity in a small number of actors (laboratories, corporations, states) whose advantage is self-amplifying by the same recursive loop that accelerates overall progress. Amodei writes that, if an authoritarian state were to achieve, thanks to AI, offensive dominance in cybersecurity or biology before the rest, the geopolitical consequences would be asymmetric and irreversible. Hassabis proposes, as a response, a CERN-inspired model of international collaboration: multilateral governance of the final steps towards general AI systems.

The institutional response - regulatory, corporate, international - is currently far behind the speed of the problem. AGI as a strategic horizon does not require organizations to resolve the philosophical debate over whether it has arrived. It requires that they act with the awareness that its consequences are already unfolding: in the systems they operate today, in the work cycles being redefined today, and in the governance decisions that are being made – or circumvented – today.