Artificial intelligence (AI) systems are defined in the OECD Recommendation on AI as (OECD, 2019[10]):
Competition in artificial intelligence infrastructure
a machine-based system that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments. Different AI systems vary in their levels of autonomy and adaptiveness after deployment.
Generative AI systems are now commonly known and available on hundreds of millions of consumer devices at the touch of a button, but for most, the underlying technology remains a mystery. The training and deployment of an AI system relies on a highly complex and capital-intensive global supply chain. For the first half of 2025, capital expenditure relating to AI infrastructure in the United States was estimated to contribute around 1.1-1.2% to GDP growth, a greater portion than spending on Internet infrastructure during the adoption of the Internet of the late 1990s to early 2000s (Mims, 2025[11]; Kedrosky, 2025[12]; Aliaga, 2025[13]).
It is important for competition policymakers to recognise that while recent digital innovations may not on the surface resemble 19th century technologies such as railroads and telephones, digital innovations are nonetheless dependent on a supply chain of physical infrastructure to function (Rahman, 2018[14]).
This section of the paper seeks to explain the layers of physical infrastructure that exist in the AI supply chain (in industry parlance, the technology “stack”) and outlines the current market features which may have implications for competition. This includes a strong focus on the supply chain for chips (the common name for microchips or integrated circuits) that are necessary to train and deploy AI systems. The chapter will also examine other digital infrastructure: namely the data centre and cloud computing layer of the AI supply chain, including the power and cooling systems needed to operate data centres, as well as the broadband network infrastructure that underpins the AI compute ecosystem.
The AI applications which are found on consumer devices today rely on complex models which have been trained on vast databases. Just a few years ago these AI systems struggled with even basic tasks. Now they can solve many complex problems, write software or create realistic images and videos (Samborska, 2025[15]). A lot of the recent improvements in AI capabilities have come from scaling up AI systems.1 These developments have required enormous improvements in computing power (Samborska, 2025[15]). The compute supply chain is therefore of crucial importance to the development of AI.
The provision of computing power to AI model developers and model users is typically provided by massive data centres and operated by cloud providers. These AI hardware operators rely on a range of key inputs including advanced compute hardware, networking (Mckinsey, 2025[16]), and energy (Chen, 2025[17]). Figure 1 shows a simplified supply chain. This chapter will begin by explaining the importance of integrated circuits (chips) to AI and explain the key parts of the chip supply chain (highlighted in green). The remainder of the chapter will discuss the other inputs involved in data centres which bring together the computing power to run AI models (purple and blue).
Note: This diagram is a significant simplification of the key steps in the AI infrastructure supply chain with key inputs for competition identified by this paper. It does not survey the other inputs to AI systems, such as data, algorithms, models and know-how.
Source: Adapted from Pilz (2023[18]) An Assessment of Data Center Infrastructure’s Role in AI Governance, https://www.konstantinpilz.com/data-centers/assessment.
Fundamentally, computer processing relies on integrated circuits (commonly referred to as microchips or chips) which contain transistors (switches) that enable a computer to make calculations (Intel, 2024[19]). Developments in AI are therefore inherently connected to innovations in chips (LeCun, Bengio and Hinton, 2015[20]).
The more transistors a chip contains, the greater its computational capacity. To increase performance, the chip industry has focused on packing more transistors onto each chip and shrinking their size. From 1950 to 2010, computing power roughly doubled every two years, a trend known as Moore’s Law, named after Intel’s founder who predicted this exponential growth (Schaller, 1997[21]; LeCun, Bengio and Hinton, 2015[20]). Since 2010, the growth in compute has accelerated dramatically, now doubling approximately every six months (Samborska, 2025[15]).
Improvements in chips helped AI move from simple rule-based systems to much more complex approaches (such as deep learning, inference and neural networks that underpin current AI developments). As AI has advanced to its current state, general purpose computing chips (such as Central Processing Units (CPUs)) which were central to previous computing developments became impractical. Instead, chips which could conduct simultaneous computations rose to the fore. Box 1 explains the key types of chips used in AI.
There are three main types of chips that are used in the training and deployment of AI systems, referred to as AI accelerator chips:
Graphical Processing Units (GPUs) – Originally designed to process images for video game graphics, these are the primary chips used for AI training. By completing computing tasks in parallel rather than sequentially, GPUs are well optimised for the type of computations needed to train AI even though they are still designed for general purpose computing needs. They are by far the most widely available and most used chips in the AI sector.
Field-Programmable Gate Arrays (FPGAs) – FPGAs are chips that can be programmed and reprogrammed to perform specific tasks after they have been manufactured, unlike GPUs which have fixed hardware structures. FPGAs are well-suited to AI computation because they can be tailored to the specific needs of AI applications. FPGAs have the advantage of being available “off the shelf” to be configured immediately but may be less cost effective and efficient in the long run compared to designing a chip for a specific need.
Application-Specific Integrated Circuits (ASICs) – ASICs are chips that are custom designed to perform a single function with high efficiency and speed.* ASICs are typically hard wired to suit a specific AI algorithm. ASICs are very hard to commercialise given the very high upfront design and manufacture costs, and the lower demand given their ability to function for only a specific application. Currently only the largest tech firms active in AI development have deployed ASICs into their AI offerings, such as Google’s Tensor Processing Units (TPUs), Google’s own AI-specialised chips.
Note: * It is important to note that the distinction between ASICs and GPUs is often blurry as GPUs were themselves a type of ASIC when originally designed for processing graphics but are used for a wider range of tasks
Source: Center for Security and Emerging Technology (2020[22]), AI Chips: What They Are and Why They Matter, Center for Security and Emerging Technology, https://doi.org/10.51593/20190014.
The following sections run through the key layers of the chip production process, starting with a brief overview of the value chain before discussing the different key firms operating at each level. There are multiple elements which come together in data centres to form the compute backbone for AI, this section is not exhaustive but seeks to highlight some of the key components in the value chain.
To produce the chips (shown in Box 1) which are key to AI’s development and operation, there is a complex supply chain which relies on an array of different actors playing different roles. These steps are shown in Figure 2. A full discussion of the value chain can be found in Mapping the Semiconductor Value Chain (OECD, 2025[8]), which is summarised below. The process can be segmented into three major steps: Design, Fabrication, and Assembly Test and Packaging (ATP). Each of these process steps depends on distinct inputs including materials such as silicon but also equipment, often from highly specialised suppliers.
Note: This diagram is a significant simplification of the key steps in the AI infrastructure supply chain with key inputs for competition identified by this paper. It does not survey the other inputs to AI systems such as data, algorithms, models and know-how.
Source: Adapted from (OECD, 2025[8]) Mapping the semiconductor value chain https://doi.org/10.1787/4154cdbf-en.
Design refers to creating the architecture and layout of a chip. Chip design is highly complex interdisciplinary process, involving thousands of engineers, years of research and development, and hundreds of millions of dollars of investment (Semiconductor Industry Association, n.d.[23]). Chip design is a high-value activity that sits upstream of manufacturing and downstream of fundamental research. It adds significant intellectual property (IP) value and can shape the performance, cost and energy efficiency of AI systems. Chip design for AI accelerators goes through regular design cycles with the latest chips typically released every one or two years, which can quickly make previous models redundant given the exponential improvements in technology.2 Fabrication refers to a key part of the manufacturing process in which highly specialised tools and processes are used to physically produce the chips. This process is extremely capital intensive with a high degree of economies of scale. After being fabricated, chips are cut and packaged individually onto circuit boards, placed into their frame and protective outer shell, and tested before being released for sale this is referred to as Assembly /Testing and Packaging.
At the different levels of the supply chain there are also different business models operating which we briefly define below.
Integrated Device Manufacturers (IDM) design, produce and sell chips. Traditionally, IDMs handled all three process steps – design, front- and back-end manufacturing. However, IDMs have increasingly outsourced parts of their production, following the so-called “fab-lite” business model.
Fabless companies design and sell chips, they outsource manufacturing to foundries (manufacturing facilities producing chips for other companies) and Outsourced Semiconductor Assembly and Test (OSAT) companies.
Foundries operate fabrication plants (fabs) to provide manufacturing services to companies that design chips (fabless, Integrated Device Manufacturers (IDM) and system companies).
Systems companies also design chips and rely on outsourcing for manufacturing however unlike fabless companies they design them for their own products and services.
Outsourced Semiconductor Assembly and Test (OSAT) companies offer contract manufacturing services to external customers in relation to testing and packaging the chips
Semiconductor Intellectual Property (IP) vendors design and sell functional blocks (also referred to as “IP cores”) that are used by chip designers to shorten time to market and lower chip design costs. Examples include a block to enable the chip to connect to USB or ethernet port.
2.3.1. Suppliers of chips used in AI
As noted in Box 1, GPUs are currently the most used chips for running AI tasks in data centres. The GPU market is highly concentrated with high margins, frequent innovations and huge capital investment. Nvidia (a fabless company) has emerged as the market leader in the sector, with recent estimates suggesting that the firm has over 80% market share for GPU chips used for AI (Farooque, 2025[24]). Nvidia has gross margins of over 70% and has seen its revenues increase by 405% between 2023 and 2024 (NVIDIA, 2025[25]). In July 2025, Nvidia became the first public company to reach a USD 5 trillion market value, exceeding even major tech firms such as Microsoft and Apple (Mickle, 2025[26]; Montgomery and Robins-Early, 2025[27]).
AMD is Nvidia’s most direct rival in the GPU chip design space, but the firm was slower to focus on GPUs and they have struggled to catch up. This is driven by Nvidia’s first mover advantage, strong performance and the CUDA software that Nvidia built around its hardware (Pak, 2024[28]). Recently however AMD has announced some major supply arrangements including with OpenAI (AMD, 2025[29]).
Software plays a critical role in unlocking the full potential of GPUs for AI workloads. While GPUs provide the raw computational power needed for training and inference, it is the software stack; comprising frameworks, libraries, drivers, and compilers; that orchestrates and optimises this hardware for efficient parallel processing. Innovations in software enable better memory management, faster data throughput, and support for increasingly complex models, making it possible to scale AI systems effectively. Without robust and adaptable software, even the most advanced GPUs would be underutilized in AI applications.
Nvidia was the leader in this area and created the CUDA software environment, which is designed to optimise communication between AI training and deployment tasks and Nvidia GPUs. Given Nvidias strong market position, CUDA has in essence become the industry standard and Nvidia’s CEO Jensen Huang has described it as the ‘’operating system’’ for AI. There have been several attempts to replicate this software environment using opensource standards, for example, Modular.
Source: Gambacorta, L. and V. Shreeti (2025[3]), The AI supply chain, https://www.bis.org/publ/bppdf/bispap154.htm Pak, A. (2024[28]), The CUDA Advantage: How NVIDIA Came to Dominate AI And The Role of GPU Memory in Large-Scale Model Training, https://medium.com/@aidanpak/the-cuda-advantage-how-nvidia-came-to-dominate-ai-and-the-role-of-gpu-memory-in-large-scale-model-e0cdb98a14a0 (accessed on August 2025); Bradshaw, T. (2024[30]), Nvidia’s rivals take aim at its software dominance, https://www.ft.com/content/320f35de-9a6c-4dbf-b42f-9cdaf35e45bb (accessed on September 2025); Vipra, J. and S. Myers West (2023[31]), “Computational Power and AI”, https://ainowinstitute.org/publications/compute-and-ai.
Intel, the IDM which for many years has dominated CPU sales (as the key chip for running the Windows operating system) has attempted to design a rival GPU but has only achieved a very small market share (Cusumano, 2024[32]). Nvidia has recently invested in Intel and formed a partnership to produce customer CPUs that NVIDIA will integrate into its AI infrastructure platforms (Stokel-Walker, 2025[33]).
Outside of GPUs, major tech companies (including Amazon, Google, Microsoft and Meta) have also begun to design their own ASICs. However, these ASICs have typically only been available through each firm’s own cloud services and are designed for a specific use case.3 Some commentators suggest demands for ASICs may increase as demand for AI inference increases, as inference workloads are less complex and more focused on speed and cost less than the training of models (UncoverAlpha, 2024[34]). There are also several smaller startups trying to produce specialist chips (Nicol-Schwarz, 2025[35]), although to date none have been able to gain significant market share. Lastly, companies like Alibaba, Baidu and Huawei from the People’s Republic of China (hereafter ‘China’) are also starting to produce their own AI accelerators (Gambacorta and Shreeti, 2025[3]). With recent restrictions on Nvidia’s use in China, it may be that the Chinese market dislocates from other markets. As of 2025, Huawei was estimated to have 28% of the market share of AI accelerators in China compared to Nvidia’s 54% (Olcott and Wu, 2025[36]).
As well as AI accelerator chips there are also a range of other chips important to AI compute. For example, memory chips are also an important and concentrated area with providers such as Micron, Sk Hynix and Samsung which have the majority of market share (Davies, 2025[37]). These providers are all IDMs. Two key types of memory chips are explained below (Davies, 2025[37]):
Dynamic Random-Access memory (DRAM), which have become more commodity-like in nature with designs following industry standards and differentiation mainly focusing on speed, capacity and power efficiency. To compete, suppliers must invest in large scale facilities and memory prices are highly cyclical.
HBM chips, more recently have come to the fore with new innovations helping to alleviate the memory limitations in storing and retrieving data efficiency for AI. As the HBMs lifecycle has shrunk, it has been reported that standards are struggling to keep pace, meaning that at least temporarily, there is increased product differentiation in the market as customers such as Nvidia have used custom HBB solutions (trendforce, 2025[38]). SK Hynix has reported to have a market share of over 50% of the market (Davies, 2025[37]).
2.3.2. Manufacturers of chips used in AI
Referred to as foundries in industry jargon, they are responsible for the front-end manufacturing of chips. The most advanced foundries are some of the most complex and expensive production facilities on the planet and take years to build (OECD, 2025[8]). Building cutting edge fabs requires enormous capital investment and achieving a satisfactory return investment is possible only for firms with very large scale (Varas et al., 2021[39]). Currently, few firms have the technical capacity to compete for contracts to fabricate AI accelerator chips
Taiwan Semiconductor Manufacturing Corporation (TSMC) is the market leader among chip foundries sales to fabless companies. Data from 2024 estimates that TSMC controls a solid majority of worldwide chip manufacturing contracts, with a market share above 60% (Bowman, 2024[40]). TSMC’s share further increases when looking at contracts for only the most advanced chips, reaching 90% market share (Bowman, 2024[40]). For AI accelerator chips, TSMC leadership in 2024 disclosed that 99% of the world’s AI accelerators are made with TSMC technologies (Wu, 2024[41]). Samsung is currently the second largest firm in terms of foundry revenue but also manufactures its own chips. It has been able to develop capacity to manufacture cutting edge chips by focusing its strategy on attracting orders to manufacture ASICs (i.e. the custom chips often developed by large technology companies that are optimised to work for a specific AI workload, such as Google TPUs) (Bowman, 2024[40]). Intel has reportedly struggled to keep pace with technical developments, at the expense of its market share (OECD, 2025[8]; Aguirre, 2024[42]). However, it has also reportedly started large-scale production of the most advanced chips to be manufactured in the US after a recent USD 32 billion investment (Acton, 2025[43]).
After being fabricated, chips are cut and packaged, placed into their frame and protective outer shell, and tested before being released for sale. This OSAT step in the chip value chain was in recent times undertaken by firms that were not the chip fabricators. ASE Group, Amkor Technology, JCET and Tongfu Microelectronics as the leading firms specialising in these OSAT services (OECD, 2025[8]). For more advanced chips (including AI accelerators), TSMC, Samsung and Intel have developed greater internal capacity and are all seeking to outsource less to OSAT firms. While there is less publicly available information on relevant market shares, potential scenarios in the longer-term could see the leading AI accelerator chip foundries compete for OSAT contracts for the chips they are fabricating, with the standalone OSAT suppliers focusing on less advanced chips for other use cases (OECD, 2025[8]; Mordor Intelligence, n.d.[44]).
2.3.3. Suppliers of key of inputs into AI chip production
There are a broad range of inputs into chip production. Many of these are markets which are highly concentrated, but for the purposes of this paper, we briefly highlight the key suppliers in two of the higher value elements of the supply chain, silicon lithography and electronic design automation.
Silicon lithography
Silicon lithography is the process of making the intricate patterns of circuits that form chips. Innovations in the chip industry to squeeze more transistors into increasingly small chips require more advanced methods of silicon lithography to create the patterns at this nanoscopic scale. In recent years, ASML, a firm based in the Netherlands has established itself as the leading firm in lithography, providing the lithography machines for all AI accelerator chip manufacturers (Center for Security and Emerging Technology, 2020[22]; Aguirre, 2024[42]; Narechania and Sitaraman, 2023[45]).
No other firm has been able to commercialise an alternative to ASML’s Extreme Ultraviolet (EUV) Lithography technology that is necessary to manufacture the latest generations of chips. The two other major firms active in lithography (Nikon and Canon) have not been able to operational EUV technology. They are unable to provide the machines vital for manufacturing AI accelerator chips as these require higher precision than they what is available from their machinery (Center for Security and Emerging Technology, 2020[22]; Aguirre, 2024[42]; Narechania and Sitaraman, 2023[45]).
ASML manages a vast and complex network of suppliers and on occasion has acquired or invested in several of their key input suppliers for these lithography machines. For example, buying the firms supplying light sources and error in light beam detection technology, as well as taking a 24.9% stake in a subsidiary of optics firm Carl Zeiss that manufactures optical parts for ASML (ASML, n.d.[46]; ASML, 2016[47]). More directly relevant to AI they have also recently they have also made a direct investment in the other end of the supply chain, taking a stake in French AI model developer Mistral (ASML, 2025[48]).
Electronic Design Automation
Electronic Design Automation (EDA) is “specialised software used by engineers to bring together semiconductor designs using IP cores and custom designs. It allows them to design, simulate and verify the design” (OECD, 2025[8]). EDA software is designed in collaboration with the development kits chip fabricators release to ensure designs can be manufactured in their foundries.
Regardless of who is designing and fabricating a chip, EDAs are key inputs given their ability to shorten time to market and lower chip design costs. Driven by demand for AI accelerator chips (as well as other chipsets like those for smartphones or automotive technologies), firms have seen high growth in recent years (Mordor Intelligence, n.d.[44]). Previous work from the (OECD, 2025[8]) has shown that three firms, Cadence, Synopsys and Arm “account for more than 60% of the global EDA market and 70% of the IP market”. This concentration among just a small number of players means they have become indispensable for AI accelerator chip designers and manufacturers working at the cutting-edge.
While advanced chips are one of the core technologies powering the AI revolution, the technology must be operated together with other key inputs including power and cooling as highlighted in Figure 3 below. This section discusses the data centre operators which some in industry have started referring to as ‘AI factories’ (Harris, 2025[49]).
Note: This diagram is a significant simplification of the key steps in the AI infrastructure supply chain with key inputs for competition identified by this paper. It does not survey the other inputs to AI systems such as data, algorithms, models and know-how.
Source: Adapted from Pilz (2023[18]), An Assessment of Data Center Infrastructure’s Role in AI Governance, https://www.konstantinpilz.com/data-centers/assessment.
2.4.1. Data centres and cloud computing
Data centres serve as the backbone for AI development, providing the computational infrastructure required for both model training and inference. During training, vast datasets are processed through high‑performance compute clusters equipped with the necessary GPUs or specialised AI accelerators. This phase is highly resource-intensive, demanding significant power, cooling and networking capacity to handle the parallel compute workloads. Once trained, models transition to inference, where they apply learned parameters to new data for predictions or decision-making. Inference still requires optimised hardware and low-latency environments (i.e. minimal delays in the time it takes for data to reach users) to deliver real-time responses at massive scale (OECD, 2025[5]).
Modern data centres are therefore evolving to balance these dual demands, integrating energy-efficient designs, access to clusters of AI accelerator chips, workload orchestration, and advanced networking to support the growing scale and complexity of AI applications. To reach the necessary scale of data centre compute and service provision, AI firms have typically relied on the data centres and services operated by the largest cloud computing suppliers (Narechania and Sitaraman, 2023[45]).
Cloud computing refers to a service model (OECD, 2014[50]):
for enabling on-demand network access to a shared pool of configurable computing resources (e.g. networks, servers, storage, applications and services) that can be rapidly provisioned and released with minimal management effort or service provider interaction (Mell and Grance, 2011[51]).
There are broadly three categories of operators with regards to AI cloud providers (Lehdonvirta et al., 2025[52]):
Government funded compute facilities – these are typically intended for academic or military use.
Private compute clusters – these are owned by for-profit companies that build and use the compute for their own business purposes. They consist of large numbers of AI accelerators mounted into interconnected computers deployed in data centres. A private cluster can be used to power the company’s own AI development or rented out to another company.
Public cloud providers that are also for-profit companies. They are called “public” not because of any government affiliation but because their services are in principle available on demand to the general public and thus shared by many customers.
Given the scale of data centre compute and service provision required to train and deploy AI services, the largest cloud providers are well placed to take significant market share in AI cloud provision. Google Cloud, Amazon Web Services and Microsoft Azure have been dubbed the hyperscalers and are subsidiaries of larger digital services companies (OECD, 2025[5]). As noted by the French (Autorité de la concurrence, 2023[53]),
All three belong to major digital companies that are among the world's largest market capitalisations. They already have a strong presence in digital services markets and have leveraged their considerable financial resources and internal needs to build up IT capacity worldwide and offer a large number of diverse cloud services, which have subsequently formed ecosystems.
Compared to other areas of the AI infrastructure supply chain, competition authorities have already conducted far more in-depth research into cloud market dynamics. Across the market studies conducted to date, the combined market share of the hyperscalers across national and regional cloud computing services markets is consistently significant (OECD, 2025[5]).4 The most recently reported market share estimates for the cloud market globally also suggest the largest three providers have over 60% share of the market (Richter, 2025[54]).
AI firms often use the hyperscaler cloud computing services, given the hyperscalers’ data centres ability to provide the necessary compute resources and networking capacity necessary for training and deploying AI (Stucke and Ezrachi, 2024[55]). Hyperscalers are also vertically integrated into firms that are active in AI development (OECD, 2025[5]). For example, Google operates its own data centres that power its own AI services (such as the Google Gemini AI assistant). Hyperscalers have also begun to strike partnerships or strategic collaborations with major AI firms. Typically, this involves the hyperscaler making a multi-billion dollar investment in the AI firm, with the AI firm agreeing to use the hyperscaler’s cloud services and data centres for training and deploying their AI models (OECD, 2025[5]). For example, in November 2024, Amazon invested an additional USD 4 billion into major AI firm Anthropic, with Anthropic announcing Amazon Web Services as their primary training partner (Amazon, 2024[56]).
However, there are other providers offering cloud compute capacity to AI. These include, for example, Oracle, which has recently signed a massive USD 300 billion deal to provide cloud computing power to OpenAI the model developer which developed the popular ChatGPT AI service (Tom's Hardware, 2025[57]). There are also other smaller providers such as such as CoreWeave, Crusoe, Nebius, and Lambda Labs, which apply a similar on-demand business model, but focus exclusively on providing AI compute services. CoreWeave has received investments from Nvidia, securing a multiple billion-dollar order that requires Nvidia to purchase any unsold CoreWeave capacity through to 2032 (Reuters, 2025[58]). In addition, some AI modellers such as Mistral have also sought to move up the chain and are beginning to look to develop compute capacity (Mistral AI, 2025[59])
2.4.2. Connectivity services and infrastructures, also referred to as “digital infrastructure”
As AI models grow in size and complexity, networking becomes a critical enabler of performance and scalability. Training large models like LLMs requires thousands of GPUs to work in parallel, exchanging data continuously at ultra-high speeds.
AI systems rely on distributed cloud computing environments, often spanning multiple data centres across regions. This creates an even greater need for robust digital infrastructure and global connectivity, which includes fibre networks, Internet Exchange Points (IXPs) and submarine cables (Filippucci et al., 2024[60]). In the past two decades, innovations in the digital services sectors have created far greater demand on this backbone infrastructure. For the largest technology companies, investing in and/or developing backbone infrastructure is necessary to ensure efficient and reliable data distribution. More recent developments such as the widespread adoption of cloud computing, and the need for a global network of data centres for AI compute has only increased demand for large technology firms who can afford to vertically integrate across several parts of backbone infrastructure (Filippucci et al., 2024[60]).
The OECD (2024[61]) has studied in detail the activities and investments of the five largest digital services firms by market capitalisation (i.e. Alphabet/Google, Amazon, Apple, Meta/Facebook and Microsoft). Since 2024 the level of capital investments made by these players has also continued to accelerate (Thomas, 2025[62]) . These firms have made substantial investments in digital infrastructure, including submarine cables, terrestrial cables (e.g. fibre), cloud infrastructure and data centres. Such investments aim to help them move data traffic between countries, as well as between their data centres and end users.
They have become particularly important investors of submarine cables, a critical component of backbone infrastructure. Prior to 2012, these firms accounted for less than 10% of total used capacity of submarine cables. By 2021, this figure had climbed to 69% (OECD, 2024[61]). Alongside investments through consortiums, technology companies independently financing submarine fibre cables have become more common in recent years. Nearly half of trans-Pacific cable investments scheduled to begin operation between 2023 and 2025 are backed or funded by these major technology companies (OECD, 2024[61]).
Figure 4 below shows a visual representation of the Internet backbone infrastructure underpinning Google’s Dunant submarine cable connecting its US and European data centres and cloud computing networks.
Note: The diagram is a simplified version that does not explain more detailed aspects of submarine cables such as different cable sheaths at different depths, or the power and transmission equipment needed to send information along fibre optic cables.
Source: Adapted from Jayne Stowell (2018), “Delivering increased connectivity with our first private trans-Atlantic subsea cable”, The Keyword (Google Blog), https://blog.google/products/google-cloud/delivering-increased-connectivity-with-our-first-private-trans-atlantic-subsea-cable/.
Large technology firms have also invested in ways to make sure data moves smoothly and reliably across the Internet from their data centres to users, including their AI compute-based products and services. They do this by making special agreements, called peering agreements, to connect their networks with others. They also work on systems that help store and deliver data faster, called content delivery networks (CDNs), often by partnering with communication companies or Internet providers. Some technology companies run their own CDNs to support their services and those of other businesses. In addition, firms have formed agreements with Internet service providers to place servers closer to users, which helps deliver content (including AI based) more quickly and efficiently (OECD, 2024[61]).
2.4.3. Access to energy and water for data centres
As discussed above AI’s transformative potential is underpinned by its computational intensity. As more advanced AI systems are trained and deployed, their reliance on data centres and high-performance computing infrastructure has made energy supply and generation infrastructure a strategic input.
While estimates vary based on geography, the AI model being trained or queried, and the data centre infrastructure, it is clear that AI systems (namely the data centres where they are trained and deployed from) require a great deal of energy (Filippucci et al., 2024[60]; OECD, 2019[63]). Indeed, within the industry, data centre capacity is often measured in terms of power (megawatts) (Competition and Markets Authority, 2025[64]).
An illustrative example of the energy demand can be drawn from estimates relating to ChatGPT and other OpenAI products. The MIT Technology Review estimated that OpenAI’s GPT-4 model (launched in 2023 and deployed to the public through products such as ChatGPT and Microsoft Copilot) used 50 gigawatt-hours of electricity to train (O’Donnell and Crownhart, 2025[65]). This equates to the same power usage of the population of San Francisco for three days. Analysis by investment firm Blackstone estimates compared AI outputs to a search engine query, with one ChatGPT query consuming roughly 10 times as much power, and a video generation request involves 10 000 times as much power (equivalent to charging a typical smartphone roughly 119 times) (Klimczak, 2024[66]).
In their market study into cloud computing, evidence gathered by the UK (Competition and Markets Authority, 2025[64]) showed that:5
The energy supply to run and cool IT equipment in data centres is the largest portion of operational costs for running a data centre (compared to factors such as rent, maintenance, equipment, depreciation and labour costs).
Access to energy may present challenges for establishing new data centre capacity, as continuous and reliable energy sources are not always present, especially in areas with existing high energy demand.
Against this backdrop, firms including Amazon, Microsoft and Google have acquired stakes or entered into long-term supply commitments with energy suppliers to meet the power requirements of their current and future data centres (Competition and Markets Authority, 2025[64]). This includes investing in reactivating nuclear energy plants and natural gas generators (da Silva, 2024[67]; Sherman, 2024[68]; Weise and Metz, 2025[69]).
Energy efficiency also creates additional incentives for the largest firms in the sector to vertically integrate across operating data centres, designing and manufacturing AI accelerator chips. Custom designing Application-Specific Integrated Circuits (ASICs) described above in Box 1 can increase the efficiency of AI compute, allowing firms to reduce their reliance on more heavy general AI chips such as the GPUs manufactured by Nvidia. For example, Google’s development of its custom Tensor chips focuses on enhancing energy efficiency (Netherlands Authority for Consumers and Markets, 2022[70]); and Amazon’s partnership with Annapurna Labs to manufacture chips for its AI data centres aims to reduce energy consumption (Weise and Metz, 2025[69]).
Operating compute in data centres, particularly those involving large-scale model training and inference, also generates significant heat. Data centres for AI compute also rely on water-intensive cooling systems (in addition to the indirect water usage relating to electricity generation). This is an often-overlooked aspect of the AI infrastructure supply chain, with far less data and academic literature available (OECD, 2022[71]). Estimates from (Mytton, 2021[72]) note that data centres are already among the top-10 water consuming industries in the United States, and “often cluster in similar geographic areas and many rely on scarce water supplies, particularly in the western United States” (Siddik, Shehabi and Marston, 2021[73]).
Access to abundant and reliable water sources can influence where data centres are located, creating regional advantages and investment flows for firms able to locate data centres in optimum locations in terms of water access. Access to sufficient water sources may also enable more innovative and efficient operation of AI data centres, by allowing firms to reduce their energy consumption. The French (Autorité de la concurrence, 2023[53]) market study highlighted firms such as OVHcloud and Scaleway, which are operating data centres with water based cooling systems that can save up to 40% in energy consumption compared to conventional air conditioner based cooling systems. Technology on cooling is developing rapidly, and as technologies such as closed loop systems develop access to large amounts water may become less important.
← 1. While stock prices and investment plans of the compute industry suggest huge growth in compute needs, reports that Chinese model developer DeepSeek was able to produce an effective AI model with far less compute highlights that technological developments are inherently uncertain (Davern and Pinnuck, 2025[172]).
← 2. Previously GPU releases were typically every 2-4 years, but Nvidia has recently committed to an annual release cycle.
← 3. Recently there have been reports Google is looking to also offer its TPUs to other AI cloud providers. Eg see https://www.datacenterdynamics.com/en/news/google-offers-its-tpus-to-ai-cloud-providers-report/
← 4. Market share of hyperscalers as reported in competition authority market studies: range from 40-62% for AWS, 10-35% for Azure and 5-10% for Google cloud (Autorité de la concurrence, 2023[53]; Netherlands Authority for Consumers and Markets, 2022[70]; Ofcom, 2023[80]; Japan Fair Trade Commission, 2022[81]; Korea Fair Trade Commission, 2022[82]).
← 5. Similar findings were made in the cloud computing sector market studies conducted by the French Autorité de la concurrence (2023[53]) and the Netherlands Authority for Consumers and Markets (2022[70]).