Artificial intelligence (AI) compute is critical to AI progress, influencing where and how advanced AI systems are developed and deployed. Alongside data, algorithms and skills, AI compute has emerged as a strategic asset with profound implications for innovation and competitiveness. This Spotlight examines the evolving dynamics of AI compute across three key dimensions, emphasising its role in advancing AI progress. First, it analyses the compute layers of the AI stack, exploring opportunities and bottlenecks to system-level performance – from the advent of increasingly specialised AI chips to the role of software and model design in efficient scaling. Second, it delves into the economics of AI compute, including surging investment and emerging business models such as neoclouds. Finally, it outlines policy considerations, focusing on key questions to ensure appropriate access to compute as a foundation for shared AI progress.
OECD Digital Economy and Society Outlook 2026
Abstract
The development and use of artificial intelligence (AI) rely on specialised infrastructure known as AI compute. Along with other key enabling inputs such as skills, data and algorithms, AI progress hinges on AI compute as a strategic asset for developing, scaling and appropriately accessing AI systems. Policies that incentivise private sector AI compute investment through light touch regulation and enabling environments will help steer compute capacity. This, in turn, can help achieve both economic security and resilience objectives across the AI stack. Considerations around AI compute increasingly shape national and regional priorities, including the capacity to build and use AI systems adapted to local needs, languages, public policies and other domestic priorities.
This Spotlight discusses various dynamics affecting the future of AI compute and its role in advancing AI progress. It first examines anticipated technological advances and challenges across the compute layers of the AI stack (Figure 1.S.1). Second, it provides an overview of the economics of AI compute highlighting investment, appropriate access and market trends. Lastly, the Spotlight looks at key policy considerations to ensure the future of compute supports shared AI progress.
AI compute comprises a range of technologies, from AI chips (i.e. specialised AI accelerators/semiconductors) to data centres to cloud computing. It includes “one or more stacks of hardware and software used to support specialised AI workloads and applications in an efficient manner” (OECD, 2023[1]). It is distinct from general-purpose compute, such as central processing units (CPUs), in several ways. Although general-purpose compute can technically be used for AI workloads, it is significantly less efficient and effective. Training advanced AI systems requires a high level of parallelised computing, in which many calculations are carried out in parallel. This has led to increasingly specialised compute layers in the AI stack as AI applications, the number of parameters and dataset sizes grow (OECD, 2023[1]). For example, hardware is increasingly purpose-built for AI workloads, such as tensor processing units (TPUs) or increasingly AI-specialised graphics processing units (GPUs).
AI compute enables the training of AI systems (i.e. the creation or selection of models/algorithms and their calibration) and inference (i.e. using the AI system to produce an output) (OECD, 2024[2]). This results in AI compute requirements that vary according to user needs. AI compute can be owned and appropriately accessed in several ways: “on premise” (e.g. in a private or government-owned data centre), in “the cloud” (e.g. by purchasing compute from public cloud providers) or at “the edge” (e.g. through a mobile device). Such ownership and modalities for appropriate access can raise key policy questions, including around economic security and supply chain resilience. Large-scale AI compute providers include government-funded compute facilities, private compute clusters and public cloud providers that provide compute for public purchase.
While the performance of individual compute components in the AI stack, such as GPUs, are critical, the future of AI training and inference relies on several other factors. AI compute depends on system-level performance throughout the AI stack and on appropriate access to key enabling inputs such as energy, water and skilled labour (e.g. in computer science, programming, data engineering) (Figure 1.S.1). AI compute also depends on communications networks, which enable access to data, as well as their transmission and distributed processing. In particular, telecommunications networks constitute a fundamental component for enabling architectures such as distributed cloud, edge computing and data-intensive digital services. As such, compute performance also depends on the capacity of networks to support higher data volumes, low latency and high reliability for AI workloads.
Compared to around 2020 when AI compute was not well understood beyond specialised technical and policy communities (OECD, 2023[1]), AI compute is now firmly on the agendas of policymakers worldwide. Many countries view AI compute as strategic infrastructure needed within their borders. Consequently, they have invested significantly – from building data centres to securing AI chip supply chains. Such actions can be driven by various national policy goals, including doubling down on advantages in AI system development, or by national security or data localisation requirements. Policy action can also support national AI compute capacity, such as in economies with little domestic capacity and limited access to regional or international compute providers.
Precise and comparable measures of domestic AI compute capacity are not widely available. According to one estimate of public cloud availability (i.e. cloud compute available for public purchase), the United States and the People’s Republic of China (hereafter “China”) account for the majority share of regions with AI-capable public cloud compute, while emerging economies represent just 17% of these availability zones1 (Lehdonvirta et al., 2025[3]; OECD.AI, 2026[4]). As sizeable AI compute investments come online, the technology and policy questions raised in this Spotlight warrant attention. Actions today should support a future where appropriate access to AI compute lays the foundation for shared AI progress.
Notes: This figure is illustrative of the compute elements in the AI stack and not exhaustive (e.g. it does not include communications networks, AI model or data layers as its focus is on AI-specific infrastructure components of the AI stack). Edge computing devices are not pictured.
Source: OECD elaboration based on OECD (2023[1]), “A blueprint for building national compute capacity for artificial intelligence”, OECD Digital Economy Papers, No. 350, OECD Publishing, Paris, https://doi.org/10.1787/876367e3-en.
Memory, interconnection and thermal limitations: Can AI compute keep up?
While experts disagree on how scaling AI inputs – including compute – will yield future advances in AI capabilities (Gundlach et al., 2025[5]; International AI Safety Report, 2026[6]), AI compute today faces several technical limitations. These include limitations in memory, interconnection – the movement of data across the stack – power and heat management (discussed below). Overcoming these limitations depends on continued innovation across the AI stack, and particularly for its compute layers, contributing to advances in system-level performance.
Memory bandwidth and latency currently limit AI compute (“the memory wall”). AI compute faces a bottleneck caused by the widening gap between increasingly performant AI chips and limits in memory bandwidth and latency (Gholami, 2024[7]). While AI chips like GPUs are growing increasingly powerful and fast, memory technologies like Dynamic Random-Access Memory have not kept pace. Evidence demonstrates that accelerator performance2 has tripled every two years, while Dynamic Random-Access Memory has only scaled at 1.6x for the same period (Gholami, 2024[7]). This gap means that AI models lag in getting data in and out of memory quickly. This means that AI processors (chips) can sit underused, waiting for data to be retrieved from memory, which is especially relevant for deep learning and transformer-based models. The ability to move and access data through memory also directly affects energy consumption. Addressing the “memory wall” calls for technological innovation towards different computing paradigms, such as those leveraging techniques like compute-in-memory3 (Yu et al., 2021[8]) along with optimisation at the software and algorithmic levels.
Moving data between chips is another bottleneck to AI compute capabilities (“the interconnect wall”). As AI models and systems scale in size and complexity, training and inference increasingly rely on distributed computing across multiple chips, nodes and racks, and the movement of data between them (Costa, 2025[9]). Even with increasingly powerful AI chips, slow interconnects can reduce workload throughput and increase AI model training time. Overcoming the interconnect wall involves advances in hardware such as through optical interconnects (Priyadarshi, 2025[10]) and software, like in-network computing (Silva et al., 2023[11]) or adaptive routing algorithms (Bonato et al., 2025[12]). Newer AI applications like agentic AI4 for autonomous network orchestration (Komaragiri and Sheelam, 2025[13]) are emerging as potential additional ways to optimise interconnects and networks more broadly.
Electricity and cooling also limit AI compute (“the power and thermal wall”). This limitation occurs when increasing AI accelerator performance is limited by how much power and heat data centres can reliably handle (SEMIVISION, 2025[14]). This is concerning as some industry analysts project AI chip power demand will increase by a factor of ten within the next two to three years. For example, an NVIDIA A100 chip in 2025 consumed about 400 watts (W), while NVIDIA’s R300 – projected to reach the market in 2028 – will demand 3 000 W (SEMIVISION, 2025[15]). Increases in power demands present major challenges for data centres, which must not only supply sufficient energy but also sufficient cooling to prevent overheating and system failure. This challenge is compounded by rising density in server racks, as AI chips are packed more closely together to optimise power density for training advanced AI systems. Innovations like liquid cooling (over traditional air cooling) may help mitigate some of these limitations (Ramakrishnan et al., 2025[16]).
Specialised AI chips and the future of semiconductor manufacturing
The increasing demand for AI chips is driving a new wave of innovation in semiconductor manufacturing, with the creation of powerful “superchips” designed and specialised for AI workloads. For example, there have been innovations in semiconductor advanced packaging, facilitating seamless integration of multiple compute and memory dies (a small piece of semiconducting material containing the integrated circuit built during wafer fabrication) into a single chip. This technique, which improves data transfer speeds and energy efficiency, has already been implemented in the last few generations of AI chips (AMD, 2023[17]; NVIDIA, 2025[18]). The rise in demand for specialised AI chips is also promoting investment and research into cooling solutions integrated into the chip package. Using microfluidic cooling systems, for example, brings liquid coolant directly inside the silicon where heat is most concentrated (Bolgar, 2025[19]; SEMIVISION, 2025[14]). Another advance is silicon photonics, which uses light instead of electricity to move data between chips and servers. This helps AI data centres overcome bandwidth, distance and power limits, as model sizes keep growing (IEEE, 2025[20]).
Cloud hyperscalers5 and other compute providers are also developing custom AI chips tailored to their specific needs. Hyperscalers like Amazon Web Services (AWS), Google Cloud and Microsoft Azure remain major customers of accelerator companies like NVIDIA and AMD. However, they are also investing significantly to develop their own chips (Morgan, 2023[21]) to optimise performance and reduce dependency on third-party suppliers. These accelerators are often designed for specific internal workloads, such as search, video processing or recommender systems, where customisation can yield significant efficiency gains. Such custom accelerators typically remain complementary rather than completely replacing more traditional GPUs (Moss, 2023[22]; Gartenberg, 2024[23]).
From chips to systems: Why system-level performance is critical
Advanced AI models are trained through the orchestration of an increasingly large number of specialised chips across multiple servers across multiple geographically distributed data centres. This trend underscores that the performance of individual chips, while important, is no longer sufficient to determine overall AI system performance. What increasingly matters is how technology components work together across the compute layers of the AI stack, which is facilitated by its interconnects.6 Interconnects in AI data centres broadly refer to all physical and logical pathways that enable data to move between compute elements. This ranges from nanometre-scale wiring that links compute and memory units in a chip to the high-speed networks connecting thousands of servers across clusters and data centres (Figure 1.S.2). To exchange information efficiently across every layer of communication (from processing units, to memory, storage and systems), advances in interconnect speed and latency are essential.
Collaboration across the compute layers of the AI stack and between industry players is needed to advance AI progress. The need for system-level optimisation has led to technological innovation across the compute layers of the AI stack. This includes co-design efforts in which chip manufacturers, cloud providers and networking companies collaborate to ensure that system-level performance scales with the demands of AI systems. It is critical that this collaboration between industry players continue to ensure sustained progress in AI (Moss, 2023[22]). Beyond system-level optimisation, at the governance level, it is also essential to link various elements of national compute infrastructure with related policy dimensions. Together, they form an integrated strategic approach – such as enabling policies and laws, skills, research and development, resource efficiency and other considerations (OECD, 2023[1]).
Examples of key technology components relying on interconnects across the compute layers of the AI stack
Source: OECD elaboration.
The rise of AI-ready data centres
Data centres are evolving to meet AI workload demands, prompting significant investment to reshape how they are designed, built and operated. Data centres are receiving extremely large private and public sector investments, which is raising concerns. Can data centres continue to outperform other sectors or will today’s investors be left holding expensive assets they cannot resell at competitive prices (GRI Institute, 2025[24])? Investments in data centres also raise key questions about resource security, particularly in areas facing energy grid constraints and water scarcity (OECD, 2022[25]).
Data centres are also part of an interconnected ecosystem in which networks play a fundamental role in their operation, interconnection and geographic distribution. This interaction is essential for the development of distributed architectures, such as edge computing and multi-cloud environments, where networks make it possible to bring computing resources closer to the end user and reduce latency. Network operators may play relevant roles as owners, operators or partners in data centre infrastructure.
An AI-ready data centre is purpose-built to support the unique demands of AI workloads. Data centres can sometimes be retrofitted to support AI stack needs, such as where facilities are shared for multiple purposes in rented colocation. However, companies are emerging that provide technical specifications tailored to AI – termed “AI-ready data centres” in this report. Compared to traditional IT infrastructure, AI-ready data centres must often embody several characteristics as noted below.
-
High power density: today’s AI racks (e.g. server racks containing AI chips like GPUs) typically require between 100-200 kilowatts (kW) per rack compared to 10-15 kW for standard IT racks. This already requires strong and reliable power distribution and control. Future AI racks are projected to demand even greater power density, such as 1 megawatt (MW). By the end of this decade, such developments may require up to 10 times more power than today’s AI racks and up to 100 times more than traditional racks (Blake et al., 2025[26]; Continuum Labs, 2025[27]).
-
Advanced cooling: the high power density of chips like GPUs generates much more heat than traditional infrastructure. This can make traditional data centre air cooling techniques insufficient for the systems to function properly. AI-ready centres can use “direct-to-chip” or “immersion” liquid cooling, often following the specifications of AI chip vendors (Musipatla, 2024[28]). This can also affect the amount of water required (Li et al., 2025[29]).
-
Energy efficiency: given the significant power requirements of the AI stack, data centre operators typically aim for high energy efficiency to manage resources and save costs. One measure of energy efficiency is power usage effectiveness (PUE). The generally accepted theoretical minimum for PUE is 1.0 (the ratio of total energy used to the energy delivered to IT equipment). High-efficiency IT equipment strives to get as close to this ideal as possible1 (OECD, 2022[25]). AI-ready data centres typically aim for ideal PUE ratings at or below 1.2 (as required, for example, by Germany’s Energy Efficiency Act) compared to traditional or older data centres typically with higher PUEs (Lawrence, 2024[30]; GRI Institute, 2025[24]).
-
Heavy infrastructure: compared to traditional IT equipment, the AI stack is significantly denser because hardware is packed tightly to maximise computing power. AI racks can weigh up to 2 tonnes, requiring reinforced flooring and other structural upgrades (Smolaks, 2024[31]; GRI Institute, 2025[24]).
-
Scalable design and modularity: data centre components age at different rates (e.g. typically a 3-5 year replacement rate for servers and 10-30 years for infrastructure systems like cooling and generators) (Monzavi, 2022[32]). However, there is likely to be some level of disruption from technology innovations and new state-of-the-art demanded by advanced AI systems. As the AI stack evolves, scalable and modular data centre construction is essential to mitigate infrastructure obsolescence and adapt to the latest AI technology state-of-the-art.
1. PUE is “defined as the ratio of total facilities energy to IT equipment energy”, as illustrated by the equation PIE = Total Facility Energy / IT Equipment Energy (The Green Grid, 2012[63]). A PUE of 1.0 is a generally accepted mathematical theoretical limit. It is not achievable in practice as this would mean all energy would be delivered directly to IT equipment, which is not achievable in practice as there is energy required for auxiliary systems like cooling and lighting (The Green Grid, 2012[63]).
Source: OECD estimates with sources indicated in-text.
Promoting efficient AI compute
It will remain critical to ensure AI systems are supported by efficient AI compute to manage costs and key resources like energy, water and raw materials like critical minerals. Despite major increases in the workloads of data centres, their energy use has remained remarkably consistent, until recently. In 2014, data centres used about 194 Terawatt hours (TWh) or 1% of global energy demand (IEA, 2025[33]). By 2020, they still accounted for this same percentage (IEA, 2021[34]; OECD, 2024[35]) but rose to a 1.5% share of global energy demand by 2024 (IEA, 2025[33]). Global data centre electricity demand is projected to more than double by 2030 to around 945 TWh – about the electricity consumed by Japan in 2025 – to account for approximately 3% of global electricity consumption (IEA, 2025[36]; IEA, 2025[33]). AI is expected to be the primary driver of this surge, with energy demand from AI-ready data centres (see Box 1.S.1) projected to more than quadruple over the same period (IEA, 2025[36]).
Data centre operators and their suppliers must also address electricity grid management issues. For example, when an AI training run needs to synchronise results across many parallel tasks, a large number of GPUs may have to power down at the same moment. This sudden drop in electricity use creates sharp fluctuations in power demand, which can place stress on the electricity grid and waste energy. Such challenges underscore the need for increasingly efficient and modular AI compute and data centre design to manage key resources like energy and water. Software-hardware co-design can be essential for promoting efficient AI compute. This approach involves optimally designing both algorithms and chip architectures to enable efficiency in AI-specific applications (Weissberger, 2025[37]). AI itself is also being deployed in promising use-cases to improve efficiency in energy grid management and optimisation. Such applications could yield efficiency gains that help offset the mounting electricity demand driven largely by generative AI (IEA, 2025[33]). However, the International Energy Agency also notes significant barriers to adoption of AI in the energy sector that will need to be overcome to unlock these efficiency gains at scale.
A country’s domestic energy resources are also important factors, with jurisdictions able to leverage competitive energy sources reaping potential comparative advantages. Energy sources such as nuclear could play a growing role in the energy mix of countries. Methods to capture and re-use data centre waste-heat are also promising, offering an additional pathway to reduce overall energy demand and enhance compute efficiency. Waste-heat recovery systems can extract thermal energy from liquid-cooled or immersion-cooled environments and supply it to nearby buildings or district heating networks. Integrating such solutions into data centre design could both reduce cooling loads and provide renewable-like thermal energy to local communities (OECD, 2022[25]).
An increasing number of researchers and companies are also advancing potential solutions for orbital-based compute – putting data centres in space – to leverage the abundance of solar energy in space. By leveraging the continuous availability of solar power and free-space optical interconnects (using light to transmit data), such systems hold promise to address energy concerns raised by current terrestrial compute modalities (Beals, 2025[38]; Kübler et al., 2026[39]). However, such an idea is not new and remains unproven at scale given challenges of operating compute hardware in space. These include latency issues; handling space radiation and solar flares that damage hardware; operations and maintenance costs; and challenges in cooling hardware in no-gravity settings.
In addition to energy, compute providers also increasingly contend with challenges securing sufficient water for data centre buildouts and operations, making this input likely an additional bottleneck to expanding AI compute. Data centres use water both directly for cooling and indirectly through water-intensive electricity generation. Semiconductor fabrication also uses large amounts of water (OECD, 2022[25]).
The role of AI model design, software and algorithms in advancing AI progress
Innovation in AI model design, software and algorithms are also key to advancing AI progress. Innovation at these levels can enable deployment of AI models in resource-constrained settings, such as on mobile edge devices, both increasing adoption and reducing costs. They point to a future where advances in AI system capabilities do not necessarily mean increases in the use of compute and key resources like energy and water.
The evolution of AI model design, training and inference techniques will continue to have implications for AI compute needs, affecting the relationship between the amount of compute used and the model’s performance. While training compute is often a significant one-time cost to train the model on extensive data, inference compute can compound as an ongoing cost every time the model is used (OECD, 2023[1]). For example, today’s advanced large language models deliver more accurate outputs if compute is appropriately increased during inference, i.e. when querying an already trained AI model. Whereas older models answered queries directly, more recent “reasoning systems” can “talk” themselves through answering queries via extended “Chains-of-Thought”. In so doing, they generate extensive discourse before producing an answer and thus require more compute to generate an output (International AI Safety Report, 2026[6]). The use of such systems is expected to rise as agentic AI becomes more prominent. While improving model accuracy, transparency and performance, such systems risk further increasing compute needs and costs at the inference stage (Feng et al., 2025[40]; International AI Safety Report, 2026[6]).
Model design can also help increase compute efficiency. Model pruning, quantisation, distillation and top-k routing in “Mixture-of-Experts” models are some techniques to improve compute efficiency. Pruning removes redundant model weights, while quantisation reduces numerical precision to shrink model size and speed up inference. For its part, distillation leverages outputs from more advanced models to less capable ones that can run more efficiently, while retaining most of the original model’s performance (Dong et al., 2023[41]). Top-k routing can help increase a model’s parameter count without a proportional increase in the compute used (Shazeer et al., 2017[42]). This has led several language models to garner global attention, such as DeepSeek-V2 and related models (DeepSeek-AI, 2024[43]), showcasing how AI capabilities can advance with varying compute resources and costs.
Beyond Moore’s Law: New computing paradigms, including quantum computing
New computing paradigms are being explored that go beyond traditional silicon-based computing technology. This is critical to the future of compute as Moore’s Law7 slows in response to the approach of conventional semiconductors to physical limits in miniaturisation.8 Yet advanced AI systems continue to require significant AI compute and energy to scale. New computing paradigms offer possibilities in alternative architectures and materials. This could sustain or accelerate compute performance when silicon can no longer deliver improvements at scale. To that end, it could leverage novel computing principles, optics or even neuroscience (e.g. brain-inspired architectures like neuromorphic computing) (Shastria et al., 2021[44]).
Of particular relevance, quantum computing is an emerging paradigm for information processing that uses the principles of quantum mechanics to carry out calculations that are difficult for classical computers or entirely beyond their reach. Quantum machine learning has the potential to significantly lift the computational ceiling of AI with more efficient training. Researchers increasingly combine the distinct fields of AI and quantum technologies into “quantum AI”, exploring how the two technologies could work together (Coccia, 2024[45]). Such advances are tightly linked to ongoing progress towards fault-tolerant hardware, i.e. theorised large-scale and stable processors that can reliably perform quantum computing, particularly through improvements in qubit9 quality and quantity (OECD, 2025[46]). As hardware capabilities evolve – from noisy intermediate-scale quantum computing to fully fault-tolerant systems – both algorithmic improvements and hardware manufacturing will have to evolve together to unlock commercial quantum computing applications (Acampora et al., 2025[47]).
Quantum AI applications face an additional challenge: the slow operational speeds of qubits fundamentally restrict the rate at which classical data can be transferred in and out of a quantum computer. This makes it unfeasible to convert large training datasets (classical data) into quantum states (e.g. superposition, entanglement),10 as well as to convert quantum outputs into bits for classical computers to read and interpret.
In the short term, the technical challenges for quantum AI applications are significant, but potential solutions are emerging. Many experts do not consider quantum machine learning technically viable in the next decade (OECD, 2025[46]). However, in the nearer term, hybrid quantum-classical computing approaches could contribute to AI applications. In these cases, a given problem is decomposed into distinct parts that can be efficiently tackled using quantum algorithms and others that are better suited to classical algorithms. In addition, new quantum-inspired AI algorithms are emerging that apply mathematical principles from quantum computing without needing actual quantum hardware. These approaches can improve the efficiency and performance of existing classical computing systems and pave the way for more powerful AI applications well before fault-tolerant quantum computers are available.
Investment is surging to build AI compute supply; will AI demand keep pace?
Investment in AI compute is booming. Between 2024 and 2025, AI firms working on IT infrastructure and hosting attracted the most venture capital (VC) investment – from USD 54.5 billion to USD 123.9 billion. This was almost half of all VC investment in 2025 (USD 276.5 billion). Between 2012 and 2025, VC investment in IT infrastructure and hosting reached a cumulative USD 291.2 billion, partially reflecting a rush to build AI compute infrastructure for scaling advanced AI systems (Figure 1.S.3) (OECD, 2026[48]).
In addition to VC investments, corporate internal investment in AI compute is significant. The largest cloud providers are making massive investments in AI infrastructure, as reflected in their capital expenditure trends. In addition to the big players, the emergence of AI-specific cloud providers, termed “neoclouds”11 offering “AI-compute-as-a-service” or “GPU-as-a-service”, is also shaping the AI compute market by providing a supply of AI-specific compute. Recent estimates suggest that as of 2023, for the EU27, Canada, the United Kingdom and the United States alone, investment in AI-related computer hardware totalled USD 54 billion. This estimate likely captures only a portion of the total spending on compute infrastructure. However, it suggests that public and private spending on AI-specific compute is significantly higher than that of private equity or venture capital investments, considering that VC investment in IT infrastructure and hosting totalled approximately USD 30 billion in 2023 (OECD.AI, 2026[49]).
Venture capital investments in AI by industries receiving the largest total investment, 2012-2025
Notes: Historical revisions and methodological improvements to the Preqin dataset can affect results; caution is advised when comparing different versions of the data. AI firms operating in IT infrastructure and hosting offer an indicative measure of AI infrastructure (AI compute) firms as Preqin’s methodology for classifying firms by industry takes a broad definition of IT infrastructure and hosting. See the methodological note for more information ( https://oecd.ai/en/preqin).
Sources: OECD.AI (2026[49]), “Live Data: Investments in AI and Data”, OECD.AI Policy Observatory, https://oecd.ai/en/data?selectedArea=investments-in-ai-and-data (accessed on 30 July 2026); OECD (2026[48]), Venture capital investments in artificial intelligence through 2025, https://doi.org/10.1787/a13752f5-en.
Governments, often together with private sector partners, have also announced unprecedented commitments to expand domestic AI compute capacity. The acceleration of public investment in AI compute since about 2024 reflects growing recognition of compute as a source of domestic competitive advantage. Governments are introducing dedicated AI compute strategies or embedding robust compute components within broader national AI strategies (OECD, 2026[50]). This reflects a policy shift towards viewing AI compute as a critical national asset to be strategically managed.
Three countries provide compelling examples of the shift towards appreciating AI compute as a national asset. In 2024, Canada launched its Sovereign AI Compute Strategy, committing CAD 2 billion to build sovereign compute capacity. This was complemented by funding and policies enabling small and medium-sized enterprises, as well as research institutions, to access AI compute through subsidies and credits (Government of Canada, 2025[51]). In 2025, the United Kingdom launched the UK Compute Roadmap, including up to GBP 2 billion to “deliver a holistic and user-centred compute ecosystem” and GBP 1 billion to expand the AI Research Resource. It also announced a new national supercomputing service (Gov.UK, 2025[52]). In 2025, the United States announced Stargate, a joint venture between OpenAI, SoftBank and Oracle that commits billions to building AI data centres. This move complements efforts to reduce regulatory barriers to building data centres as part of the 2025 United States AI Action Plan (OpenAI, 2025[53]; The White House, 2025[54]).
Cross-border technology partnerships are also becoming strategic tools. The European Union (EU) has co-ordinated a continent-wide approach to AI compute development through multiple interconnected initiatives. Building on the European High Performance Computing Joint Undertaking (EuroHPC JU) and its EUR 7 billion budget for 2021-2027, the EU has launched new programmes focussed on augmenting AI compute capacity. These include the EU AI Factories and AI Gigafactories, typically funded through sharing costs between EuroHPC JU and participating members (European Commission, 2026[55]).
Evolving business models and appropriate access modalities for AI compute
A new generation of cloud providers – neoclouds – has emerged to meet the compute-intensive demands of AI workloads. Unlike traditional hyperscalers, neoclouds are purpose-built for advanced AI, such as machine learning and high-performance computing. Their core business model offering is GPU-as-a-service, providing access to cutting edge AI chips with low-latency interconnects (Craske, 2025[56]). Such companies aim to provide faster access to AI chips at lower costs than traditional cloud providers. Examples of neoclouds include CoreWeave (the United States), Nebius (the Netherlands) and Nscale (the United Kingdom).
How AI compute is appropriately accessed, and whether it is available for public purchase, funded by government or privately owned, will shape innovation trajectories (Table 1.S.1). Evolving business models and modalities for appropriate access for AI compute also raise competition questions. Massive capital investments in AI compute infrastructure, networking and energy access give traditional hyperscalers a considerable advantage in market positioning (OECD, 2025[57]). Combined with deep integration into AI frameworks, extensive developer ecosystems and close partnerships with hardware vendors, these factors may create powerful path dependencies. While newer AI-specific services like those provided by neoclouds may emerge or expand, AI compute users will likely continue to rely at least in part on existing hyperscaler infrastructure available for public purchase.
Note: Information on AI compute business models and appropriate access modalities is illustrative and non-exhaustive. Public compute refers to compute services available for public purchase, and should not be confused with government-funded compute (sometimes referred to as public sector compute).
Source: OECD analysis based on Craske (2025[56]), Top 10: Neocloud Companies Transforming Global Data Centres, https://datacentremagazine.com/top10/top-10-neocloud-companies-transforming-global-data-centres; and Lehdonvirta et al. (2025[3]), Measuring domestic public cloud compute availability for artificial intelligence, https://doi.org/10.1787/8602a322-en.
AI compute capacity is a key factor shaping who can build and deploy advanced AI systems. Policymakers face several key considerations in planning and managing AI compute as a technically capable strategic asset that can facilitate appropriate access to efficient AI compute in the years to come. They must also consider how to appropriately widen access beyond advanced nations and large technology companies so AI can benefit all people.
Today’s compute investments hold promise to lay the foundation for future AI progress. Current global spending on AI compute is enormous. However, figures often reflect announcements and commitments rather than implementation. Executing on such commitments depends on key factors like supply chain resilience, energy/water availability and potential shifts in national priorities. Policymakers should consider measures to track the successful implementation of AI compute commitments, including how they promote prosperity for all. Such measures could help assess the long-term viability of large-scale public investments in compute infrastructure over time as AI compute business models and the technology itself advances. Analysis of competition dynamics in AI compute providers is also essential to ensure a vibrant AI compute supplier market that spurs innovation in compute products and services.
Natural resource efficiency and supply must be appropriately managed. The natural resource demands of training and using advanced AI systems are significant. This could create bottlenecks to AI progress as hyperscale data centre buildouts can take up to three years but securing multi-gigawatt power capacity can take over four years (US Department of Energy, 2024[58]; Data Center Frontier, 2025[59]). To mitigate this timing mismatch, policymakers should consider efficient resource management as part of AI compute planning. This could include leveraging economies of scale and efficiencies through establishment of “Special Compute Zones” that pool and leverage shared resources (Centre for Future Generations, 2025[60]). Planning could also be improved by investing in more precise measurements or estimation techniques for the natural resource needs of AI systems. In addition, policymakers could consider tax incentives, modernising building codes or public-private partnerships that encourage complementarity and efficiency in compute projects, such as where compute infrastructure and district heating networks coexist.
Optimising the AI stack is as important as scaling hardware performance. Training advanced AI models is often a one-time, high-intensity cost, while AI inference involves continuous workloads that grow with user adoption. Efficiency gains will come not only from hardware improvements, but also from optimisation across the AI stack and system-level performance, including at the model and algorithmic levels. Industry collaboration on establishing system-level performance benchmarks can help quantify AI compute so it can be reliably measured and appropriately managed (MLCommons, 2024[61]).
International co-operation and cross-sectoral partnerships are critical to scale AI compute efficiently, including for large data centres. National compute strategies are on the rise, but approaches that lack co-ordination risk inefficiency and fragmentation. Regional projects like the EU EuroHPC Joint Undertaking and proposals for a “CERN-like” global facility illustrate potential advantages to sharing compute strategic assets (Sargent et al., 2025[62]). Policymakers can leverage partnerships, joint investment frameworks and cross-border AI compute agreements facilitating appropriate access to meet national AI compute needs, while maximising efficiency in capital expenditure and natural resource use. Without co-ordinated efforts, policymakers risk facing redundant investments, inefficient infrastructure expansion and deepening imbalances that could cause resource shortages and disrupt supply chains. The key question is: how can countries better align their strategies to avoid these pitfalls? Further analysis is needed into potential co-operation models in practice.
References
[47] Acampora, G. et al. (2025), “Quantum computing and artificial intelligence: Status and perspectives”, arXiv, Vol. 2505.23860, https://arxiv.org/abs/2505.23860.
[17] AMD (2023), “Instinct™ MI300X Accelerator”, https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html (accessed on 19 October 2025).
[38] Beals, T. (2025), “Exploring a space-based, scalable AI infrastructure system design”, Google Research, https://research.google/blog/exploring-a-space-based-scalable-ai-infrastructure-system-design (accessed on 18 March 2026).
[26] Blake, M. et al. (2025), “NVIDIA 800 VDC architecture will power the next generation of AI factories”, Technical Blog, https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-power-the-next-generation-of-ai-factories (accessed on 19 October 2025).
[19] Bolgar, C. (2025), “AI chips are getting hotter: A microfluidics breakthrough goes straight to the silicon to cool up to three times better”, https://news.microsoft.com/source/features/innovation/microfluidics-liquid-cooling-ai-chips/ (accessed on 27 October 2025).
[12] Bonato, T. et al. (2025), “Flowcut switching: High-performance adaptive routing with In-order delivery guarantees”, arXiv preprint, https://arxiv.org/pdf/2506.21406.
[60] Centre for Future Generations (2025), “Special compute zones: Europe’s recipe for boosting strategic autonomy in AI infrastructure”, https://cfg.eu/special-compute-zones-europes-recipe/ (accessed on 19 October 2025).
[45] Coccia, M. (2024), “Converging artificial intelligence and quantum technologies: Accelerated growth effects in technological evolution”, Technologies, Vol. 12/5, p. 66, https://doi.org/10.3390/technologies12050066.
[27] Continuum Labs (2025), “NVIDIA GB200 NVL72”, https://training.continuumlabs.ai/infrastructure/servers-and-chips/nvidia-gb200-nvl72 (accessed on 19 October 2025).
[9] Costa, P. (2025), “Breaking the networking wall in AI infrastructure”, Microsoft Research Blog, https://www.microsoft.com/en-us/research/blog/breaking-the-networking-wall-in-ai-infrastructure/ (accessed on 10 October 2025).
[56] Craske, B. (2025), “Top 10: Neocloud companies transforming global data centres”, https://datacentremagazine.com/top10/top-10-neocloud-companies-transforming-global-data-centres (accessed on 19 October 2025).
[59] Data Center Frontier (2025), “Facing the post-AI energy crunch: New research correlates data center power delays with innovation strategies”, https://www.datacenterfrontier.com/machine-learning/article/55289418/facing-the-post-ai-energy-crunch-new-research-correlates-data-center-power-delays-with-innovation-strategies (accessed on 19 October 2025).
[43] DeepSeek-AI (2024), “DeepSeek-V2: A strong, economical, and efficient mixture-of-experts language model”, arXiv, Vol. 2405.04434v5, https://arxiv.org/pdf/2405.04434.
[41] Dong, P. et al. (2023), “HeatViT: Hardware-efficient adaptive token”, arXiv, Vol. 2211.08110, https://arxiv.org/pdf/2211.08110.
[55] European Commission (2026), “AI factories”, https://digital-strategy.ec.europa.eu/en/policies/ai-factories (accessed on 15 June 2026).
[66] Fassi, I. and D. Shipley (eds.) (2017), “Introduction to miniaturisation”, Micro-Manufacturing Technologies and Their Applications: A Theoretical and Practical Guide, pp. 1-22, https://doi.org/10.1007/978-3-319-39651-4_1.
[40] Feng, S. et al. (2025), “Efficient reasoning models: A survey”, arXiv preprint, https://doi.org/10.48550/arXiv.2504.10903.
[23] Gartenberg, C. (2024), “TPU transformation: A look back at 10 years of our AI-specialized chips”, Google Cloud Blog, https://cloud.google.com/transform/ai-specialized-chips-tpu-history-gen-ai (accessed on 10 October 2025).
[7] Gholami, A. (2024), “AI and memory wall”, IEEE Micro, Vol. 44/3, pp. 33-39, https://doi.org/10.1109/MM.2024.3373763.
[52] Gov.UK (2025), “UK Compute roadmap”, https://www.gov.uk/government/publications/uk-compute-roadmap (accessed on 10 April 2025).
[51] Government of Canada (2025), “Canadian sovereign AI compute strategy”, https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy (accessed on 4 September 2025).
[24] GRI Institute (2025), “Data centres: Too good to pass or liquidity nightmare?”, https://www.griinstitute.org/realestate/event/europe-gri-2025_4756 (accessed on 10 April 2025).
[5] Gundlach, H., J. Lynch and N. Thompson (2025), “Meek models shall Inherit the Earth”, arXiv, Vol. 2507.07931, https://doi.org/10.48550/arXiv.2507.07931.
[65] Hongwu, J., S. Huang and S. Yu (2023), “Compute-in-memory architecture”, in Handbook of Computer Architecture, Springer Nature Singapore, https://link.springer.com/rwe/10.1007/978-981-15-6401-7_62-1.
[36] IEA (2025), “AI is set to drive surging electricity demand from data centres while offering the potential to transform how the energy sector works”, https://www.iea.org/news/ai-is-set-to-drive-surging-electricity-demand-from-data-centres-while-offering-the-potential-to-transform-how-the-energy-sector-works (accessed on 10 April 2025).
[33] IEA (2025), “Energy and AI”, https://www.iea.org/reports/energy-and-ai/energy-demand-from-ai (accessed on 29 October 2025).
[34] IEA (2021), “Global trends in Internet traffic, data centre workloads and data centre energy use, 2010-2020”, Data and Statistics, https://www.iea.org/data-and-statistics/charts/global-trends-in-internet-traffic-data-centres-workloads-and-data-centre-energy-use-2010-2020.
[20] IEEE (2025), “Trends and challenges in optics for AI/ML data center scale-out networking”, https://events.vtools.ieee.org/m/483371 (accessed on 19 October 2025).
[6] International AI Safety Report (2026), “International AI Safety Report 2026”, DSIT 2026/001, https://internationalaisafetyreport.org (accessed on 17 March 2026).
[13] Komaragiri, V. and G. Sheelam (2025), “Agentic AI for autonomous network orchestration: A new frontier in Telecommunications”, EKSPLORIUM-BULETIN PUSAT TEKNOLOGI BAHAN GALIAN NUKLIR, Vol. 46/1, pp. 221-241, https://eksplorium.com/index.php/journal/article/view/23.
[39] Kübler, F. et al. (2026), “From concept to prototype: Realizing an in-orbit AI-system utilizing optical computing”, CEAS Nature Space Journal, https://doi.org/10.1007/s12567-025-00612-z.
[30] Lawrence, A. (2024), “Resiliency vs low PUE: Could regulation spark innovation?”, https://www.datacenterdynamics.com/en/opinions/resiliency-vs-low-pue-could-regulation-spark-innovation (accessed on 19 October 2025).
[3] Lehdonvirta, V. et al. (2025), “Measuring domestic public cloud compute availability for artificial intelligence”, OECD Artificial Intelligence Papers, No. 49, OECD Publishing, Paris, https://doi.org/10.1787/8602a322-en.
[29] Li, P. et al. (2025), “Making AI less ’thirsty’: Uncovering and addressing the secret water footprint of AI models”, arXiv, Vol. 2304.03271, https://arxiv.org/abs/2304.03271.
[61] MLCommons (2024), “Benchmarks”, https://mlcommons.org/benchmarks (accessed on 19 October 2025).
[32] Monzavi, B. (2022), “Taking a look at the lifespan of data centre components”, https://blog.cirkla.tech/2022/05/13/taking-a-look-at-the-lifespan-of-a-data-centre (accessed on 17 March 2026).
[21] Morgan, T. (2023), “How AWS can undercut Nvidia with homegrown AI compute engines”, https://www.nextplatform.com/2023/12/04/how-aws-can-undercut-nvidia-with-homegrown-ai-compute-engines (accessed on 10 October 2025).
[22] Moss, S. (2023), “Training Google’s Gemini: TPUs, multiple data centers, and risks of cosmic rays”, https://www.datacenterdynamics.com/en/news/training-gemini-tpus-multiple-data-centers-and-risks-of-cosmic-rays (accessed on 10 October 2025).
[28] Musipatla, V. (2024), “High-performance AI: Why data centers are turning to liquid cooling technology”, International Journal of Computer Techniques, Vol. 11/3, https://ijctjournal.org/wp-content/uploads/2025/05/High-Performance-AI-Why-Data-Centers-are-Turning-to-Liquid-Cooling-Technology.pdf.
[18] NVIDIA (2025), “NVIDIA GH200 Grace Hopper Superchip”, https://www.nvidia.com/en-us/data-center/grace-hopper-superchip (accessed on 19 October 2025).
[50] OECD (2026), “Securing AI compute: Government policy approaches”, OECD Policy Briefs, No. 67, OECD Publishing, Paris, https://doi.org/10.1787/7f96dc99-en.
[64] OECD (2026), “The agentic AI landscape and its conceptual foundations”, OECD Artificial Intelligence Papers, No. 56, OECD Publishing, Paris, https://doi.org/10.1787/396cf758-en.
[48] OECD (2026), “Venture capital investments in artificial intelligence through 2025”, OECD Policy Briefs, No. 50, OECD Publishing, Paris, https://doi.org/10.1787/a13752f5-en.
[46] OECD (2025), “A quantum technologies policy primer”, OECD Digital Economy Papers, No. 371, OECD Publishing, Paris, https://doi.org/10.1787/fd1153c3-en.
[57] OECD (2025), “Competition in artificial Intelligence infrastructure”, OECD Roundtable on Competition Policy Papers, No. 330, OECD Publishing, Paris, https://doi.org/10.1787/623d1874-en.
[2] OECD (2024), “Explanatory memorandum on the updated OECD definition of an AI system”, OECD Artificial Intelligence Papers, No. 8, OECD Publishing, Paris, https://doi.org/10.1787/623da898-en.
[35] OECD (2024), OECD Digital Economy Outlook 2024 (Volume 2): Strengthening Connectivity, Innovation and Trust, OECD Publishing, Paris, https://doi.org/10.1787/3adf705b-en.
[1] OECD (2023), “A blueprint for building national compute capacity for artificial intelligence”, OECD Digital Economy Papers, No. 350, OECD Publishing, Paris, https://doi.org/10.1787/876367e3-en.
[25] OECD (2022), “Measuring the environmental impacts of artificial intelligence compute and applications: The AI footprint”, OECD Digital Economy Papers, No. 341, OECD Publishing, Paris, https://doi.org/10.1787/7babf571-en.
[4] OECD.AI (2026), “Live data: AI compute”, OECD.AI Policy Observatory, https://oecd.ai/en/data?selectedArea=ai-compute (accessed on 17 March 2026).
[49] OECD.AI (2026), “Live Data: Investments in AI and Data”, OECD.AI Policy Observatory, https://oecd.ai/en/data?selectedArea=investments-in-ai-and-data (accessed on 30 July 2026).
[53] OpenAI (2025), “Announcing the Stargate project”, https://openai.com/index/announcing-the-stargate-project (accessed on 4 September 2025).
[67] Powell, P. and I. Smalley (2025), “What is hyperscale?”, https://www.ibm.com/think/topics/hyperscale#:~:text=Hyperscalers%2C%20an%20outgrowth%20of%20hyperscale,typical%20on%2Dpremises%20data%20centers (accessed on 27 October 2025).
[10] Priyadarshi, S. (2025), “Advancing AI scalability and performance with optical Interconnects”, IEEE Communications Magazine, Vol. 63/11, pp. 22-27, https://ieeexplore.ieee.org/abstract/document/11098644.
[16] Ramakrishnan, B. et al. (2025), “Understanding the impact of data center liquid cooling on energy and performance of machine learning and artificial intelligence workloads”, Journal of Electronic Packaging, Vol. 147/2, https://doi.org/10.1115/1.4067136.
[62] Sargent, M. et al. (2025), “Assessing the value of a CERN-like multinational research organization for AI”, https://www.rand.org/pubs/perspectives/PEA4024-1.html (accessed on 19 October 2025).
[15] SEMIVISION (2025), “Key players power requirement roadmap for AI”, https://substack.com/@semivision/note/c-154711241 (accessed on 10 May 2025).
[14] SEMIVISION (2025), “TSMC x Nvidia: Breaking the thermal wall: How advanced cooling Is powering the future of computing”, https://tspasemiconductor.substack.com/p/tsmc-x-nvidia-breaking-the-thermal?triedRedirect=true (accessed on 10 October 2025).
[44] Shastria, B. et al. (2021), “Photonics for artificial intelligence and neuromorphic computing”, Nature Photonics, pp. 102-114, https://www.nature.com/articles/s41566-020-00754-y.
[42] Shazeer, N. et al. (2017), “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer”, Google Research, https://research.google/pubs/outrageously-large-neural-networks-the-sparsely-gated-mixture-of-experts-layer.
[11] Silva, R. et al. (2023), “In-network computing – challenges and opportunities”, Internet Technology Letters, Vol. 7/3, p. e487, https://doi.org/10.1002/itl2.487.
[31] Smolaks, M. (2024), “AI has a weight problem”, https://intelligence.uptimeinstitute.com/resource/ai-has-weight-problem (accessed on 19 October 2025).
[63] The Green Grid (2012), “PUE™: A comprehensive examination of the metric”, White Paper, No. 49, The Green Grid, https://nikom.in/assets/pdf/0a58778d-fc96-4482-8c46-13abe76b015c.pdf.
[54] The White House (2025), America’s AI Action Plan, The White House, Washington, DC, https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf.
[58] US Department of Energy (2024), Recommendations on Powering Artificial Intelligence and Data Center Infrastructure, Secretary of Energy Advisory Board, Washington, DC, https://www.energy.gov/sites/default/files/2024-08/Powering%20AI%20and%20Data%20Center%20Infrastructure%20Recommendations%20July%202024.pdf.
[37] Weissberger, A. (2025), “OpenAI and Broadcom in $10B deal to make custom AI chips”, https://techblog.comsoc.org/2025/09/06/openai-and-broadcom-in-10b-deal-to-make-custom-ai-chips (accessed on 10 October 2025).
[8] Yu, S. et al. (2021), “Compute-in-memory chips for deep learning: Recent trends and prospects”, IEEE Circuits and Systems Magazine, Vol. 21/3, pp. 31-56, https://doi.org/10.1109/MCAS.2021.3092533.
← 1. Research by the OECD in collaboration with Oxford University Innovation identified the availability of public cloud-based compute resources relevant to AI across cloud regions. Data presented on OECD.AI represent an updated approach, automating the extraction of data on GPU availability at a more granular level using availability zones. A cloud region is a geographic area designated by the cloud provider. In practice, each cloud region contains one or multiple cloud availability zones, which in turn may host one or more physical data centres. For further information, please visit: www.oecd.ai/ai-compute.
← 2. Gholami (2024[7]) measures peak server hardware floating-point operations per second i.e. the maximum number of floating-point operations per second that an AI accelerator can theoretically perform in peak conditions.
← 3. Compute-in-memory is an emerging “non-von Neumann” (i.e. classical) computing paradigm that can help achieve performance and energy efficiency. It performs certain computational tasks in place and thus avoids the bottlenecks posed by shifting data between, for example, processing and memory (Hongwu et al., 2023[65]).
← 4. Agentic AI generally refers to “systems composed of multiple co-ordinated AI agents that can break down tasks, collaborate and pursue complex objectives autonomously over extended periods. Agentic AI systems are designed to operate in more open-ended, less predictable physical or virtual environments and to function with minimal human supervision” (OECD, 2026[64]).
← 5. Hyperscalers can be defined as “hyperscale data centres primarily used to deliver and manage mega-sized applications. Hyperscalers are largely cut from the same cloth as traditional data centres – but geared to a much larger scale than typical on-premises data centres” (Powell and Smalley, 2025[67]).
← 6. At the smallest scale, in the chip itself, compute and memory dies communicate through die-to-die interconnects. These are the densest, most power-efficient connections. Each node consists of several chips, where intra-node interconnects tie CPUs and GPUs together to enable scale-up training. Nodes then communicate via inter-node interconnects (high-speed networking fabric between servers), enabling scale-out training across many machines. Multiple servers (or nodes) fill a rack, connected by a local rack network. Several racks form a pod, joined by a high-performance backend fabric; multiple pods make a cluster, linked by a cluster-level backbone; and clusters together form a data centre, connected by a data-centre fabric split into backend (GPU-to-GPU AI workloads) and frontend (user, storage, and external ingress) networks.
← 7. “Moore’s Law refers to the observation made by Gordon Moore in 1965 that the number of transistors on a microchip doubles approximately every two years, leading to an exponential increase in computing power. This trend has been a driving force behind the many innovations in electronic devices and the progress of computing capabilities over the past several decades. However, Moore’s Law has begun to slow down primarily due to physical limitations in transistor scaling” (OECD, 2025[46]).
← 8. Miniturisation is the process of shrinking the size of a device, component or system without losing its capabilities and sometimes even improving performance (Fassi and Shipley, 2017[66]).
← 9. “Unlike classical computers that use bits to represent either a 0 or a 1, quantum computers use quantum bits (qubits), which can represent 0, 1 or any value in between through the use of superposition. Moreover, entanglement allows qubits to become interconnected: when qubits are entangled, the state of one qubit becomes intricately correlated with the state of another. The unique abilities of superposition and entanglement are expected to allow quantum computers to solve certain complex problems that are challenging or even intractable for classical computers” (OECD, 2025[46]).
← 10. Quantum states include superposition, when a “particle can be in an indeterminate, probabilistic combination of multiple places or states at the same time until it is measured” and entanglement, when “[p]articles can become so intricately linked that they cannot be measured or described independently of one another, regardless of how far apart they are” (OECD, 2025[46]).
← 11. Neoclouds are emerging compute providers that prioritise the needs of AI workloads, offering specialised AI infrastructure often at a lower cost than traditional providers (e.g. hyperscalers).