The emergence of the world’s leading Frontier Intelligence Labs represents a profound change in the history of computing. For most of the twentieth century, artificial intelligence was principally concerned with reproducing particular forms of human cognitive performance: reasoning over symbolic representations, recognising patterns, understanding language, playing games, diagnosing problems or controlling machines. The contemporary frontier is considerably more ambitious. The leading laboratories are increasingly attempting to construct systems that can reason across domains, learn from experience, formulate and pursue objectives, use tools, interact with complex environments, discover new knowledge, cooperate with other intelligent systems and, potentially, participate in the development of succeeding generations of artificial intelligence. The central technological question is therefore shifting from whether machines can perform intelligent tasks to whether intelligence itself can be engineered as a general, scalable and potentially self-improving capability.
The term Frontier Intelligence Labs is consequently preferable to a narrower description based solely upon artificial intelligence. It encompasses laboratories pursuing substantially different conceptions of what advanced machine intelligence may become. OpenAI, Google DeepMind and Anthropic have developed increasingly general systems based upon large-scale learning, reasoning and agency. xAI and Meta Superintelligence Labs are pursuing increasingly ambitious general and superintelligent systems at extraordinary computational scale. NVIDIA Research Labs is advancing the computational substrate upon which much of the frontier depends. Safe Superintelligence is explicitly oriented towards the construction of superintelligence under stringent safety objectives, while Thinking Machines Lab is investigating highly capable systems designed to extend human judgement and agency. Recursive Superintelligence approaches the problem through recursive improvement, Ineffable Intelligence through experiential learning and superlearning, and AMI Labs through world models and alternative approaches to machine reasoning. Meanwhile, Mistral AI, Aleph Alpha, Cohere and AI21 Labs represent important European, Canadian and Israeli dimensions of the frontier, while Alibaba and Qwen, Baidu and ERNIE, DeepSeek, MiniMax, Tencent Hunyuan, Moonshot AI and Kimi, Zhipu AI, Baichuan AI and ByteDance Seed demonstrate the extraordinary breadth of China's frontier ecosystem. Sakana AI adds an important Japanese perspective centred upon evolutionary and collective approaches to intelligence.
These laboratories should not therefore be understood as twenty-five versions of the same organisation pursuing the same destination. They represent a collection of competing hypotheses about intelligence. Some believe that continued scaling of general-purpose models, combined with increasingly sophisticated reasoning and agency, may be sufficient to produce general intelligence. Others believe that world models, reinforcement learning, continual learning, evolutionary processes, multi-agent systems or recursive self-improvement will be necessary. The intellectual importance of the frontier lies precisely in this diversity. The next decisive advance may arise from greater scale, a new architecture, a new training paradigm, an unexpected interaction between several existing techniques, or a conceptual breakthrough concerning the nature of intelligence itself.
American Laboratories and Competing Paths to Advanced Intelligence
OpenAI occupies perhaps the most prominent position in the contemporary development of general machine intelligence. Its significance derives from the combination of frontier research, extraordinary computational resources, global deployment, large-scale experimentation and an explicit long-term orientation towards artificial general intelligence. The organisation's development from language models towards reasoning systems, multimodal intelligence, tool use and increasingly autonomous agents reflects a broader transformation in the conception of an AI system. The model is no longer simply a mechanism that produces an answer in response to a prompt. It is increasingly conceived as a persistent computational entity capable of decomposing objectives, selecting strategies, using external resources and undertaking extended sequences of intellectual operations.
This transition has enormous implications. A system that performs a task once is fundamentally different from a system that can perform the task repeatedly, evaluate its own performance, learn from failure and adapt its approach. Generality therefore depends upon more than breadth of knowledge. It depends upon the capacity to transfer reasoning strategies between domains and to operate successfully in circumstances that were not explicitly represented in training. OpenAI's importance lies partly in its attempt to integrate these capabilities within increasingly general systems. Its trajectory is consequently central to the broader transition from artificial intelligence towards general intelligence, autonomous intelligence and potentially superintelligence.
Google DeepMind represents a complementary but intellectually distinct tradition. It combines deep learning with reinforcement learning, scientific discovery, multimodal systems, robotics and large-scale computational experimentation. Its historical work demonstrated that machines could acquire strategies that had not been explicitly programmed and could discover solutions to difficult problems through interaction and optimisation. Its subsequent scientific systems demonstrated that machine intelligence could contribute to knowledge production itself. Google DeepMind is therefore particularly important because it treats intelligence not merely as a mechanism for producing useful outputs but as a potential engine of discovery.
The organisation's current research also illustrates the increasingly important transition beyond general intelligence. The question is no longer simply whether an artificial system can equal humans across a broad range of cognitive tasks. It is whether increasingly capable systems could accelerate the research and development of their own successors. Recent research concerning possible routes from general intelligence to superintelligence has considered scaling, new paradigms, recursive improvement and collections of cooperating intelligent agents. The possibility of multi-agent systems becoming more capable collectively than individual systems is especially important because it suggests that the future architecture of intelligence may be distributed rather than monolithic.
Anthropic represents a third major interpretation of the frontier. Its Claude systems have achieved highly advanced capabilities in reasoning, programming, analysis and increasingly autonomous task execution. The organisation has simultaneously developed an unusually prominent research programme concerning the safety and controllability of increasingly capable systems. This combination makes Anthropic strategically important: it is not an organisation attempting to solve safety instead of capability, but a laboratory investigating how increasingly powerful systems can be made useful and controllable.
The significance of this problem is growing rapidly. Frontier systems are beginning to operate for extended periods, interact with external software and undertake complex technical tasks with relatively little direct supervision. Recent research indicates that current agents can perform substantial portions of the engineering involved in artificial intelligence research while still struggling with the deeper judgement, creativity, backtracking and resource allocation required for genuinely open-ended research. The frontier is consequently moving towards autonomy, but the evidence does not yet establish that autonomous artificial research has reached the point at which systems can reliably replace human scientific judgement.
xAI has become another major participant in the American frontier. Its strategic importance derives from the scale of resources committed to model development, computation and infrastructure and from its explicit ambition to create highly capable general-purpose intelligence. Its trajectory demonstrates the growing importance of computational scale in frontier development. The contemporary intelligence laboratory requires not merely researchers and algorithms but access to enormous computing clusters, specialised hardware, energy and data-centre infrastructure. The ability to combine these resources at unprecedented scale has become a strategic capability in its own right.
Meta Superintelligence Labs represents an especially significant development because Meta has moved from primarily competing through widely distributed foundation models towards an explicit ambition concerning superintelligence. Meta's current strategy combines enormous computational resources, open-weight development and an ambition to produce highly capable systems available on a broad basis. Its approach therefore challenges the assumption that frontier intelligence must necessarily remain entirely within closed laboratories. Meta's recent public positioning has explicitly emphasised the democratisation of advanced intelligence, while maintaining the company's ambition to develop superintelligent systems.
Computational Infrastructure as a Foundation for Frontier Intelligence
NVIDIA Research Labs occupies a different but foundational position. It would be misleading to regard NVIDIA simply as another model developer. Its importance lies in the relationship between intelligence and computation. Advanced machine intelligence is constrained by the speed, efficiency, memory, networking and energy consumption of the systems on which it operates. Research into processors, algorithms, simulation, robotics, model architectures and computational infrastructure consequently becomes part of the frontier itself. NVIDIA's recent investment and partnership with Safe Superintelligence demonstrates how closely the computational infrastructure and intelligence-development ecosystems are becoming intertwined.
Independent Laboratories Exploring Alternative Superintelligence Architectures
Safe Superintelligence is perhaps the clearest institutional expression of the proposition that general intelligence should not be regarded as the ultimate objective. Founded by Ilya Sutskever, the laboratory has deliberately maintained a comparatively secretive research programme while concentrating its institutional identity around the pursuit of safe superintelligence. Its importance has recently increased substantially through a major partnership with NVIDIA, giving the organisation access to a dramatically expanded computational platform. The laboratory represents a significant philosophical departure from the conventional technology-company model because its defining objective is not simply to build a commercially successful artificial intelligence platform, but to solve the problem of constructing superintelligence while maintaining safety.
Thinking Machines Lab occupies a different position. Its mission is to build artificial intelligence that extends human will and judgement, and its research has emphasised adaptability, customisation and the distribution of advanced intelligence. Its decision to release open-weight models while simultaneously investigating the safety implications of increasingly capable systems illustrates the tension between openness and control. Open weights can distribute technical capability, enable independent research and make model assumptions more inspectable, but they can also make advanced capabilities more difficult to contain. Thinking Machines therefore represents an important strand of the frontier in which the relationship between human agency and machine capability is treated as a central design question.
Recursive Superintelligence represents a more radical hypothesis concerning the mechanism through which superintelligence may emerge. Its fundamental proposition is that artificial intelligence research itself may become increasingly automated. If an AI system can design experiments, generate hypotheses, implement algorithms, analyse results and identify improvements to its own successors, intelligence becomes part of the production process through which more intelligence is created. The significance of this feedback loop is potentially enormous. Human scientific research progresses at the speed of human cognition, collaboration and institutional organisation. An automated research system operating continuously at machine speed could potentially accelerate the process by orders of magnitude.
Present Limits of Automated Scientific Judgement
Yet the transition should not be exaggerated. Current empirical evidence suggests that frontier agents remain considerably weaker at open-ended scientific judgement than their impressive performance on narrower engineering tasks might suggest. Recent evaluations have found that agents can complete substantial engineering work autonomously but remain considerably less capable when required to make genuine progress on difficult open-ended research questions. Recursive intelligence is therefore better understood as an emerging research direction and strategic possibility than as an accomplished technological reality.
Ineffable Intelligence represents another fundamentally different route. Its proposition centres upon learning through experience rather than relying predominantly upon the absorption of existing human-generated knowledge. This distinction is profound. Much of contemporary artificial intelligence is built from enormous bodies of human-created information. Such systems can acquire extraordinary breadth because humanity has accumulated vast quantities of written, visual, scientific and computational material. But a system that learns only from existing information remains fundamentally dependent upon what has already been discovered. An intelligence capable of generating experience, acting upon an environment, observing consequences, formulating hypotheses and continually updating its understanding could potentially become a generator of new knowledge rather than merely a compressor and synthesiser of existing knowledge.
AMI Labs offers another alternative. Its emphasis on world models represents a direct challenge to the proposition that linguistic competence and scale alone will produce sufficiently general intelligence. Genuine intelligence requires an internal representation of the world, an understanding of causal relationships, an ability to predict consequences and the capacity to plan. A system that can describe a physical object linguistically but cannot reliably model its behaviour has only a partial form of understanding. AMI Labs therefore represents a research tradition in which intelligence is grounded in models of environments and the consequences of actions.
These alternative laboratories are particularly important because they demonstrate that the future of intelligence cannot safely be extrapolated from the dominant architecture of the present. A new architecture may ultimately prove more important than another order of magnitude of scale.
European Strategic Autonomy in Frontier Intelligence
Mistral AI is the most prominent European challenger in the frontier model race. Its importance extends beyond technical performance because it represents Europe's attempt to retain meaningful strategic autonomy in advanced machine intelligence. A small number of American organisations have accumulated extraordinary computational resources, talent and commercial distribution. European governments and institutions have consequently become increasingly concerned about technological dependence. Mistral provides an alternative centre of capability through frontier models, open-weight systems, reasoning and increasingly agentic applications.
The strategic significance of Mistral therefore exceeds its position as a model developer. It demonstrates that frontier intelligence can be a component of national and continental technological sovereignty. The question is no longer simply which model is best. Governments and major organisations increasingly need to consider where intelligence is developed, who controls its infrastructure, where data is processed, what legal framework governs it and whether critical institutions can remain operational if access to a foreign provider is withdrawn.
Aleph Alpha and Cohere illustrate this sovereign dimension from another direction. Their importance lies less in an explicit claim to be the single laboratory most likely to achieve superintelligence and more in their attempt to create strategically controlled intelligence infrastructure for governments and enterprises. This illustrates a crucial distinction between technical frontier capability and strategic frontier capability. A laboratory may not produce the world's strongest general-purpose model while nevertheless becoming indispensable to governments, defence organisations, financial institutions or other highly regulated sectors.
AI21 Labs occupies a related position within the broader frontier ecosystem, with a history of foundation-model and language-intelligence research. Its importance illustrates the breadth of the frontier: not every consequential laboratory must seek the largest possible model or make the most explicit claims about superintelligence. Some contribute through architectures, reasoning systems, model development and specialised capabilities that can subsequently become components of larger intelligent systems.
China’s Multipolar Frontier Intelligence Ecosystem
The development of China's Frontier Intelligence Labs is one of the defining geopolitical characteristics of the current period. China should not be treated as a single laboratory competing against the United States. It possesses a broad ecosystem of competing organisations, including DeepSeek, Alibaba and Qwen, Baidu and ERNIE, MiniMax, Tencent Hunyuan, Moonshot AI and Kimi, Zhipu AI, Baichuan AI and ByteDance Seed. Their collective significance derives from the depth of China's technological ecosystem and the ability of multiple organisations to pursue frontier research simultaneously.
DeepSeek has perhaps done more than any other Chinese laboratory to alter international perceptions of China's position. Its significance has rested not only upon model performance but upon computational efficiency and the ability to achieve highly competitive reasoning capabilities despite severe constraints on access to the most advanced processors. This has challenged the assumption that frontier intelligence can be achieved simply through the possession of the greatest possible computational budget. DeepSeek demonstrates the strategic value of algorithmic efficiency and engineering ingenuity.
Alibaba's Qwen programme represents a different model. Alibaba possesses cloud infrastructure, enormous commercial resources and a large domestic user base. Qwen therefore exists within a vertically integrated ecosystem in which models, computation, distribution and commercial applications can reinforce one another. Qwen should properly be understood as an Alibaba programme rather than an independent company, but it merits separate identification because it has become one of the most important Chinese foundation-model families.
Baidu and ERNIE represent another major branch of China's frontier. Baidu has deep experience in search, language processing and large-scale computing, giving it a strong foundation for the development of general-purpose intelligence. Its continuing investment in ERNIE demonstrates that the frontier is not necessarily determined by a single period of benchmark leadership. The ability to sustain research, attract talent and integrate intelligence into a major technological ecosystem may prove equally important over the longer term.
MiniMax and Tencent Hunyuan demonstrate the ability of major Chinese technology groups to sustain frontier research through enormous domestic markets and extensive infrastructure. Tencent's scale in communications, gaming and cloud computing provides an unusually broad environment in which intelligent systems can be trained and deployed. MiniMax, meanwhile, has established itself as a significant independent developer of general-purpose models. Their importance is therefore both technical and industrial: frontier intelligence is becoming integrated into the broader structure of Chinese technology.
Moonshot AI and Kimi represent another important research trajectory, particularly in long-context reasoning and increasingly agentic systems. Zhipu AI has become a major Chinese general-purpose intelligence laboratory through its GLM family, while Baichuan AI remains an important participant in China's foundation-model ecosystem. ByteDance Seed adds another powerful research organisation backed by one of the world's largest technology platforms. These laboratories collectively demonstrate that China's frontier is neither narrow nor dependent upon one dominant organisation.
The Chinese ecosystem is consequently strategically significant not because every organisation has reached parity with every American laboratory, but because the country possesses a large number of laboratories capable of learning rapidly from one another. Competitive pressure between domestic organisations encourages experimentation, talent movement and rapid diffusion of successful techniques. The resulting ecosystem may prove more resilient than one based upon a single dominant laboratory.
Sakana AI and Evolutionary Collective Intelligence
Sakana AI introduces an especially interesting Japanese contribution to the frontier. Its research programme has explored evolutionary methods, model combination, automated scientific discovery and collective approaches to machine intelligence. This is significant because it challenges the implicit assumption that the future of artificial intelligence will necessarily consist of one increasingly large model.
Biological intelligence provides an alternative analogy. Human intelligence does not exist only within individual brains; civilisation itself constitutes a vast distributed intelligence formed through communication, institutions, specialisation and cumulative knowledge. An analogous machine ecosystem could consist of numerous specialised intelligent systems that cooperate, compete, evolve and exchange information. The resulting system could possess capabilities that no individual component possesses independently.
Sakana AI therefore connects directly to the emerging concept of cooperative intelligence. If several agents possess complementary strengths, their collective performance may exceed the capabilities of the strongest individual agent. Evolutionary selection could provide another mechanism for discovering architectures and strategies that human engineers would not easily design. The frontier could consequently move from the development of a universal machine towards the development of an intelligent society of machines.
Beyond Model Scale Towards Multidimensional Intelligence
The twenty-five Frontier Intelligence Labs collectively demonstrate that the frontier is no longer adequately described by model size. Model scale remains important, but intelligence is becoming multidimensional. A highly capable system must increasingly reason, remember, plan, use tools, interact with environments, learn from experience, evaluate its own behaviour and coordinate with other systems. The decisive advance may therefore come from integration rather than from any single capability.
Generality concerns the range of problems an intelligence can solve. Autonomy concerns its capacity to pursue objectives without continuous human supervision. Adaptability concerns its ability to respond to unfamiliar environments. Continual learning concerns its ability to acquire new knowledge after deployment. Reflection concerns its capacity to evaluate its own reasoning. Recursion concerns its ability to improve the mechanisms through which intelligence itself is generated. Cooperation concerns the ability of multiple intelligences to operate collectively. Superintelligence concerns the possibility of performance substantially exceeding human capability across broad and consequential domains.
These properties should not be treated as a simple ladder. They interact. A highly general system may remain dependent upon human instruction. A highly autonomous system may lack sufficient world understanding. A reflective system may be unable to alter itself. A recursive system may require external evaluation. A collection of cooperating agents may be collectively powerful but individually limited. The future architecture of intelligence will therefore probably be heterogeneous.
This explains why the laboratories in this survey should not be ranked simply according to the performance of their latest models. The laboratory that develops the best model in 2026 may not develop the architecture that matters most in 2030. Frontier research is characterised by uncertainty precisely because the underlying problem remains unresolved.
Automated Research and the Prospect of Recursive Intelligence
Among the most consequential possibilities is the emergence of automated artificial intelligence research. If machines become capable of generating useful hypotheses, designing experiments, writing and evaluating code, interpreting results and proposing improved systems, then the rate at which intelligence improves could accelerate dramatically. The possibility of automated artificial intelligence development is increasingly being treated as a distinct strategic issue because it could alter the relationship between research capacity and technological progress.
Yet the distinction between engineering automation and genuine scientific intelligence remains crucial. Current systems can already perform impressive coding and technical tasks, but open-ended scientific research requires judgement about which problems matter, which hypotheses are promising, when an experiment has failed, when to abandon an approach and how to recognise a genuinely novel result. Recent empirical work suggests that these abilities remain substantially less developed than the ability to execute technical instructions.
Recursive Superintelligence is nevertheless important because it identifies a plausible threshold. Once an artificial system becomes sufficiently capable of improving artificial intelligence itself, the economics of research could change fundamentally. Human researchers would no longer be the sole source of innovation in machine intelligence. Intelligence would become both the product and the productive factor.
Safety, Alignment and Control at the Frontier
The rise of Frontier Intelligence Labs also creates a corresponding challenge concerning safety and control. As systems become increasingly autonomous, the assumptions underlying conventional software safety become less reliable. A traditional program executes a defined sequence of operations. An intelligent system can select strategies, formulate intermediate goals, use external resources and adapt to unexpected circumstances.
The problem is not necessarily that these systems possess malicious intentions. Rather, increasingly capable systems can discover strategies and behaviours that were not explicitly anticipated by their developers. The longer the planning horizon and the greater the autonomy, the greater the potential gap between the behaviour developers expect and the behaviour a system discovers to be effective.
Architectural Safety, Evaluation and International Governance
This makes safety an architectural problem rather than merely a regulatory problem. Evaluation, monitoring, interpretability, containment, alignment, access control and robustness must develop alongside capability. The challenge becomes especially acute if systems begin contributing materially to the development of their own successors. The development process then becomes recursive: increasingly intelligent systems help create increasingly intelligent systems, potentially reducing the time available for human institutions to understand the resulting trajectory.
The emergence of common safety concerns across competing laboratories is itself significant because it indicates that the technological frontier is becoming an international governance problem. The future of frontier intelligence will therefore depend not only upon technical breakthroughs but also upon the ability of governments, institutions and laboratories to develop credible mechanisms for evaluating and controlling increasingly capable systems.
Frontier Laboratories as Strategic Institutions
The twenty-five organisations in this survey represent a new class of institution. They are simultaneously research laboratories, technology companies, infrastructure operators, intellectual communities and strategic assets. Their work affects not merely software development but the future productivity of economies, the speed of scientific discovery, national technological sovereignty and the balance of geopolitical power.
Their competition is therefore not simply a contest for market share. It is a contest between different theories of how intelligence can be produced. OpenAI, Anthropic and xAI represent powerful approaches based upon large-scale general-purpose systems, reasoning and agency. Google DeepMind combines these approaches with reinforcement learning, scientific discovery and a broad research tradition. Meta Superintelligence Labs adds an ambitious open-weight dimension. NVIDIA Research Labs operates at the computational foundation of the entire ecosystem. Safe Superintelligence prioritises the problem of safe superintelligence itself. Thinking Machines Lab emphasises systems that extend human will and judgement. Recursive Superintelligence investigates recursive improvement. Ineffable Intelligence investigates experience-driven learning and superlearning. AMI Labs investigates world models. Sakana AI investigates evolutionary and collective intelligence.
Mistral AI, Aleph Alpha, Cohere and AI21 Labs demonstrate that frontier intelligence also has a sovereign and institutional dimension. China's laboratories demonstrate that the frontier is increasingly multipolar. DeepSeek demonstrates the importance of efficiency; Alibaba and Qwen demonstrate the power of integrated infrastructure; Baidu demonstrates the continuing importance of established technology platforms; Tencent and MiniMax demonstrate the scale of Chinese commercial ecosystems; Moonshot, Zhipu, Baichuan and ByteDance demonstrate the depth of China's independent and corporate research base.
The result is not a single race towards one predetermined machine. It is a contest between architectures, institutions and philosophies of intelligence.
Competing Institutions and the Future Architecture of Intelligence
The world's leading Frontier Intelligence Labs are collectively conducting one of the most consequential technological experiments in human history. Their objective is increasingly to transform intelligence from a biological phenomenon into an engineered capability that can be reproduced, scaled, distributed, combined and potentially improved through further intelligence.
The twenty-five laboratories considered here occupy different positions within this transformation. OpenAI, Google DeepMind, Anthropic, xAI and Meta Superintelligence Labs represent the principal American centres of large-scale frontier development. NVIDIA Research Labs provides much of the computational foundation upon which the wider ecosystem depends. Safe Superintelligence concentrates explicitly upon the safe development of superintelligence, while Thinking Machines Lab explores highly capable systems intended to extend human judgement. Recursive Superintelligence investigates the possibility that intelligence will eventually participate directly in the production of more intelligence. Ineffable Intelligence proposes experience and continual learning as a route towards radically more capable systems. AMI Labs challenges the dominance of language-centred approaches through world modelling. Mistral AI, Aleph Alpha, Cohere and AI21 Labs demonstrate the strategic importance of independent and sovereign intelligence capabilities. Alibaba and Qwen, Baidu and ERNIE, DeepSeek, MiniMax, Tencent Hunyuan, Moonshot AI and Kimi, Zhipu AI, Baichuan AI and ByteDance Seed constitute a formidable Chinese ecosystem. Sakana AI contributes a distinctive Japanese exploration of evolutionary and collective approaches.
Their differences may ultimately prove more important than their similarities. Some laboratories are betting that scale and reasoning will be sufficient. Others believe that machines must acquire richer models of the world. Some emphasise experience, continual learning or reinforcement. Others see recursive self-improvement as the decisive threshold. Still others are investigating cooperation among multiple intelligent systems. These approaches need not be mutually exclusive. Indeed, the most powerful future systems may combine them.
A plausible long-term trajectory therefore extends beyond the conventional conception of artificial intelligence. Artificial intelligence may provide the foundation; general intelligence may provide breadth; autonomous intelligence may provide sustained agency; reflective intelligence may provide self-evaluation; recursive intelligence may provide self-improvement; cooperative intelligence may provide collective capability; and superintelligence may represent the resulting condition in which machine cognition substantially exceeds human capability across domains of strategic significance.
The central uncertainty is not whether these developments will occur in exactly this sequence. They almost certainly will not. The deeper uncertainty concerns which mechanisms will prove sufficient to transform increasingly capable artificial systems into genuinely general intelligence, and whether the transition from general intelligence to superintelligence will be gradual, discontinuous, collective, recursive or some combination of all four.
What is already clear is that the frontier has moved. The leading laboratories are no longer merely attempting to make machines answer questions more accurately. They are attempting to create systems capable of asking better questions, deciding which problems deserve attention, discovering new knowledge, acting upon the world, learning from their own experience, cooperating with other intelligences and potentially contributing to the creation of their successors. The technological frontier is consequently becoming a frontier of intelligence itself.
For this reason, Frontier Intelligence Labs is more than a convenient description. It identifies an emerging institutional category at the intersection of computing, cognitive science, mathematics, engineering, scientific discovery and industrial strategy. These laboratories are not simply building the next generation of software. They are testing whether intelligence can become a scalable technological resource, whether cognition can be industrialised, and ultimately whether the creation of increasingly powerful intelligence can itself become an accelerating process.
The answer remains unknown. But the organisations examined in this paper are now among the principal institutions determining it.