> **来源：[研报客](https://pc.yanbaoke.cn)** # Data Centre Trends Report 2026 Shifting up a gear # Contents 0 Foreword: More growth, greater efficiency 1 Data Centre Trends 2026 6 The 2026 Ten Trends 7 Trend 1: The need for speed 9 Trend 2: Tide turns for trade careers 11 Trend 3: Edge data centres: search for sites accelerates 13 Trend 4: Industrialisation revolution 15 Trend 5: Gas is back in fashion 17 Trend 6: Breakthrough cooling solutions 19 Trend 7: Rare earth elements become rarer 21 Trend 8: The race for faster permitting 23 Trend 9: New chips on menu 25 Trend 10: The year to get quantum ready 2 Industry Focus 28 Focus on EMEA 38 Focus on North America 42 Focus on Latin America 49 Delivery with AI 3 Conclusion 53 Shifting up a gear 54 Our Authors 55 About Soben 56 References # Foreword # More growth, greater efficiency As we move into 2026, the big question for the industry is: how do we pick up the pace in this accelerating sector? The last 12 months have been exciting, to say the least. Multiple new players have joined the game, with one of the biggest stories of the year being Oracle's meteoric rise to become the world's fourth largest hyperscaler. And while much of the mainstream talk has been about the multi-gigawatt ambitions for AI data centres, existing data centre markets continue to expand. Some countries are only starting to develop their cloud data centres, while others continue on that journey. Edge data centres are growing with the spread of 5G. And, once AI models have been trained in large, out-of-town data centres, we will need to find space for inference data centres – close to cities and industrial hubs – to run those models. In this report, we focus on areas where Soben, part of Accenture, is most active in the data centre market - although, there are new pins in our map almost every week. There are few regions in the world where data centre construction will not ramp up over the next 10 years. These multi-front growth ambitions have exacerbated the challenges that the sector was already facing. Perhaps the most wicked of these is how countries will power their growing data centre infrastructure. In the shorter term, the industry is already coming up with solutions, protecting constrained grids by adding more renewable energy sources and turning to natural gas as a supplementary power source. Looking only 12 months ahead, perhaps the biggest constraint is people. General contractors (GCs), mechanical and electrical contractors (MEP), construction and project managers, quantity surveyors and commissioning experts with the track record to handle these complex projects are in short supply. And, given the ambitions of many data centre developers to compress time spent on site, this is no market for novices. # A year for change We cannot rise to these challenges with business as usual. Change is already coming to every element of data centre development from the chips and equipment they deploy, to land searches, permitting and regulations; from standardisation and automation of design to AI-optimised scheduling and modular builds. In Europe, regulation is pushing greater energy efficiency in operation and low water usage. In other markets, 'sustainability' wears a different face, but the need to conserve power and water is just as great – out of necessity and to keep local politicians, communities and, in some cases, investors on side. Underpinning these shifts is the move towards digital construction and the use of AI to create more efficient processes. We cannot create experienced planners, construction professionals and electrical specialists overnight. But we can use the people we have more efficiently. Which brings me to our biggest story of 2025: Soben became part of Accenture in March. We have always been focused on finding ways to disrupt how capital projects are delivered – creating better outcomes for our clients in a rapidly-changing market. Our alliance brings together Accenture's vision of reinvention and digitalisation with Soben's industry expertise and commercial experience. It is already exciting to see how our teams are working together to combine delivery experience and new, AI-based tools to help our clients meet these challenges head on. 2026, we are ready for you! # Scott Smyth, December 2025 # Scott Smyth Founder and CEO Soben part of Accenture # DataCentre Trends 2026 Soben Part of Accenture # 2 Data Centre Trends Soben Part of Accenture # The need for speed Ambitious construction timelines Supply chains choosing projects Higher construction costs 03 # Edge data centres: search for sites accelerates >Roll-out of 5G driving growth >Inference AI data centres coming >New deals to secure real estate 04 # Industrialisation revolution Standardised design elements Modular goes hyperscale Foundations for build automation 02 # Tide turns for trade careers > Nvidia CEO highlights boom for construction trades >More Gen Zs consider construction > National interventions on training required # Gas is back in fashion Electricity grids cannot cope Natural gas to the rescue Power Couples: a hybrid approach # Rare earth elements become rarer >Further restrictions by China >Europe most at risk >Recycling ramps up 08 # The race for faster permitting >National and state level zoning > BYO energy and low water usage helps >Al does permitting leg work 09 # New chipson menu > Nvidia still dominates >Hypercalers developing their own chips >Global politics creates waves 10 # The year to get quantum ready >Hypercalers' secretive deployments >Industry organisation prepares standards >Commercial deployments closer # 01 # The need for speed BACK Speed equals revenue - whoever wins the speed race wins the revenue. Businesses still need to track their cap spending, but schedule is the priority over cost at the moment. # # Sophie Smith # VP - Americas Soben Part of Accenture > Ambitious construction timelines > Supply chains choosing projects >Higher construction costs The biggest trend for 2026 will be faster delivery. Both established players and newcomers are looking to shift left, with incredibly ambitious timelines for their projects. And some have plans to deliver several projects in parallel. Back in 2018, it may have been possible to construct a 400MW data centre in around 12 months. Today, with a shortage of contractors with data centre experience, longer equipment lead times and supply side power constraints in many parts of the world, this is a huge ask. For new entrants to the data centre development market, who don't have established supply chains, it may be impossible. There are ways to accelerate delivery, all of which the leading players are looking to deploy. These include standardised design elements, more use of prefabricated elements and modular construction, and virtually constructing data centres using 3D design and BIM to de-risk builds before work starts on site. The key to deploying such measures is a mature and competent supply chain. In existing data centre markets, general contractors (GCs) and mechanical, electrical and plumbing (MEP) contractors with the right credentials are already flat out. In newer markets, contractors and their supply chains will need time and support to climb the learning curve. # Supply chain chooses There are many would-be data centre developers with big plans and enthusiastic investors, but this is not enough to secure the supply chains they need to deliver new buildings. With demand for companies and professionals outstripping supply, designers, consultants and contractors are being selective. Given a choice between a newcomer and an established player which has relationships and deals with the owner furnished, contractor installed (OFCI) equipment suppliers, it is not difficult to see which way a GC or MEP contractor would go. Established players also learned some tough – but useful – lessons during Covid. These lessons have helped transform the way that long-lead time items such as transformers, switch gear and generators are ordered, stockpiled, and distributed among projects. New entrants won't yet have the long-term relationships with suppliers that enable these strategies. In the US, we are seeing informal partnering-type arrangements, where pragmatic commercial decisions are being used to tie in GCs and MEPs with long pipelines of projects ahead. For instance, the pain of cost hikes and programme overruns that occurred as a result of the Covid pandemic are being shared between developer and GC, avoiding the time, cost and aggravation of disputes and court cases. # Rising costs A certain impact of accelerated schedules in a resource-constrained market is higher construction costs, and we are already seeing this in the market. For instance, a US hyperscaler recently accepted a $25\%$ uplift in tender price from a GC to deliver to a demanding timeline. This comes against a backdrop of already rising costs for data centre construction and construction generally. And there is also the impact of building more AI data centres to take into account when considering likely cost inflation. Given the higher cost of chips needed to process AI workloads and of the supporting equipment and advanced cooling systems needed, AI data centres come at a premium when compared to cloud data centres. Research by Soben, part of Accenture, found that while cloud data centres currently cost between $8 million and$ 10 million per MW, GW+ AI data centres are costing as much as $17 million per MW. Contrary to popular belief, this signals that economies of scale are not driving the savings some would expect. In parallel with working with mature clients to find new ways to deliver data centres faster, companies such as Soben, part of Accenture, must provide data-based reality checks to those who don't yet appreciate the constraints and complexities of the market. Pushing forward without these reality checks increases the risks of time and cost overruns, together with associated disputes, and leaves optimistic investors disappointed. # 02 Tide turns for trade careers It's absolutely vital that we get the message out to the next generation – we need skilled tradespeople to help create the infrastructure we need for a digital economy. # # Mark Smith # Delivery Lead I&CP Accenture > Nvidia CEO highlights boom for construction trades >More Gen Zs consider construction > National interventions on training required The narrative about construction trades jobs being unattractive to the next generation has been turned on its head. In September 2025, Nvidia CEO Jensen Huang told Channel 4 News in the UK that tradespeople were set to benefit as data centre infrastructure doubles every year for a decade. Hundreds of thousands of electricians, plumbers and carpenters will be needed, said Huang. “The skilled craft segment of every economy is going to see a boom,” he predicted<sup>1</sup>. Translated into messages such as 'Plumbers and electricians will win AI race' and 'The Future Millionaires' on TikTok and YouTube, Gen Z can't help but take notice of Huang's predictions. Other big AI voices have joined the call for more tradespeople. In October OpenAI made a submission to the US's Office of Science and Technology Policy, warning that tens of thousands of new tradespeople will be needed for data centre and energy projects. With OpenAI's Stargate project alone planning for 7GW of capacity in six huge data centres in Texas, New Mexico, Ohio, and Wisconsin, the company has estimated that $20\%$ of the US's current skilled trades workforce would need to be deployed on data centres over the next five years to meet demand<sup>2</sup>. There are similar challenges in Europe. While demand for mechanical, electrical and plumbing (MEP) trades is set to rise, ageing workforces mean that numbers are falling. In the UK, the number of electrical workers fell by $19.8\%$ between 2018 and 2024 according to Government statistics while the plumbing workforce shrank by $18\%$ over the same period<sup>3</sup>. Skilled electricians and electrical specialists are also in demand for power transmission projects - which are vital to the development of data centres. In its Blueprint for Success report, Accenture cites a European utilities company which needed five times more electricity transmission specialists than it was currently employing to provide technical assurance to meet a fourfold increase in connection requests<sup>4</sup>. # Gen Z eyes construction There are signs that the tide is turning. In the US, an October 2024 survey of 2,091 adults by The Harris Poll found that $78\%$ had noticed a recent growing interest from young people in pursuing trade careers. And while $23\%$ of people who don’t currently work in the trades said they planned to move into them, this proportion rose to $42\%$ for Millennials and $50\%$ for Gen Zs $^{5}$ . In the UK, the UK electrical sector trained 1,000 more apprentices between 20-21 and 2024 than it had done over the previous three years – although this is a small increase, compared to the numbers needed<sup>6</sup>. In November 2025, a survey of further education colleges in the UK revealed that $56\%$ of them had waiting lists for construction-related courses starting in 2025. Construction had seen growth in enrolments for full-time courses, apprenticeships and adult education, with $77\%$ reporting a growth in the numbers of 16-to-18-year-olds enrolling on courses<sup>7</sup>. The salaries on offer should be a draw for anyone considering a career as a data centre electrician. The Royal Examiner reported in September that data centre electricians average $61,391–$93,341 per year, with top earners commanding $122,921–$156,466 annually. The caveat is that long working hours and working away from home are often a requirement for the job. # Hyperscalers step in Hyperscalers have long been aware of the risks their construction programmes face due to a lack of skilled tradespeople, with companies investing in training initiatives alongside data centre developments. For instance, Google's Skilled Trades and Readiness (STAR) Program is a paid five-week programme which prepares people for entry-level jobs in construction in locations such as South Carolina, Nebraska, Northern Virginia and Ohio<sup>9</sup>. Google also announced in May 2025 that it would provide an initial $10 \text{m}$ grant to the electrical training ALLIANCE, involving the International Brotherhood of Electrical Workers (IBEW) and the National Electrical Contractors Association (NECA). The grant will go towards both apprentice schemes and on programmes to upskill existing workers $^{10}$ . Suppliers are stepping in too. Schneider Electric offers training and upskilling through its Schneider Electric University<sup>11</sup>. Given the huge uplift in electricians and other tradespeople required to deliver the data centres planned, intervention from national governments is vital. This was a point underlined by OpenAI in an October 2025 open letter to the US's Office of Science and Technology Policy. It called for the federal government to fund workforce development at the state and local level[12]. # 03 # Edge data centres: search for sites accelerates 2025 has been all about the Gigawatt AI data centre plans. But the new wave of edge data centres is going to be just as exciting. # # Pieter Schaap Group Development Director Soben Part of Accenture > Roll-out of 5G driving growth >Inference AI data centres coming > New deals to secure real estate While the world is marvelling at the giga-watt ambitions of hyperscalers, the growth in edge data centres – small facilities close to cities, industrial areas and transport corridors – is predicted to be even more meteoric. Research and Markets predicted in November 2025 that the market for edge data centres would grow from $15.4 billion in 2024 to$ 39.8 billion in 2030<sup>13</sup>. In North America alone, the edge computing market will grow by 32.2% annually from 2025 to reach $2.2bn in 2030, according to Accenture<sup>14</sup>. There are several factors driving the need for more edge data centres. One of the biggest current drivers is the roll-out of 5G, well-advanced in many parts of the US and Europe, as well as China, South Korea, Japan, India and Singapore, continues in countries such as Canada, Brazil, Chile, Saudi Arabia, Oman and UAE $^{15}$ . The principle of 5G is speed, with promises that it will provide network latency as low as 1 millisecond, compared to between 50 and 100 milliseconds for 4G. To do this, computing and data storage must be moved as close as possible to where the action is, whether that be gaming, augmented reality or entertainment. That means building data centres close to cities and industrial areas. Smart cities, and the growing use of IoT-connected devices in homes, manufacturing facilities, healthcare and retail all require processing capabilities to be located close by. Statista estimates that the number of IoT connections worldwide will grow from 19.9 billion in 2025 to 60.6 billion in 2034, with growing numbers of devices used for connected vehicles, IT infrastructure, asset tracking and monitoring and smart grids<sup>16</sup>. There is another wave of change coming that will demand more edge data centres. Today AI models are being trained in huge data centres, requiring vast amounts of power, and often built in remote locations. Tomorrow, we will see those models being deployed in every imaginable application – and in some that we can't even imagine yet. Inference data centres will run the trained models, located within edge data centres where low latency is needed for the application. # Fleets of sites In selecting potential sites for edge data centres, proximity to fibre and last-mile resiliency is vital. As these facilities must be operational 24-7, power redundancy is also vital, which may be challenging to achieve within tight city footprints. It is also important to think about how these facilities will be built; modular solutions for edge data centre construction are growing fast, so any potential site should be assessed around the logistics of delivering a modular build<sup>17</sup>. Companies that already have 'fleets' of small sites are already well placed to benefit from the edge data centre explosion. When Nvidia announced it was to invest $1bn in Nokia in October 2025, working together to incorporate AI into Nokia's telecom networks and data centres, there was a telling comment from Michael Dell, the CEO of Dell. Dell said: “The telecommunications industry owns the most valuable real estate for AI — the Edge, where data is created...The operators who modernize their infrastructure today won’t just carry AI traffic – they’ll be the distributed AI grid factories that process it at the source, where latency matters and data sovereignty is critical.” 18 # 04 Industrialisation revolution Hyperscalers are really leading transformation in the construction industry, rethinking how we generate designs and treating construction sites more like factories. # # Luis Guadarrama # Lead Contracts Manager - Mexico Soben Part of Accenture >Standardised design elements >Modular goes hyperscale > Foundations for build automation The only way to achieve the speed and scale of data centre builds planned for the next decade is to move from a traditional, bespoke construction approach towards one that is industrialised. The big players already scale in the form of 'capacity blocks' of 30-to-50MW buildings, or they install halls one by one behind a moving screen within a large shell. Modularisation will super-charge this approach. Think platform-based design, where prefabricated elements and modules are the building blocks of a scheme's design and delivery onsite is a matter of assembly with fewer labour resources needed on the ground. The foundations for this approach are already in place, with prefabricated electrical rooms, power skids and modular halls plugging into some of today's developments. But now, with some hyperscalers looking 50-plus data centres ahead, there is an opportunity to take this approach to the next level. With these long pipelines of work, supply chains can be tied in early, with their components and equipment plugged into 3D BIM designs from the earliest phases. And constructing more and more elements inside manufacturing facilities, away from the variabilities of a site environment, means installations can be pre-tested and quality assured before they leave the factory. # Next-level modular The deployment of modular data centres is already growing fast in the edge data centre. Since these tend to be smaller facilities, often in already built-up areas, fabricating offsite and craning into place can make perfect sense. The modular data centre market is set to grow from $32.4bn to$ 85.2bn between 2024 and 2030, according to Research and Markets $^{19}$ . Multiple moves in the latter half of 2025 indicate that the market is shaping up to meet the surge in edge data centre development needed to service 5G, IoT and inference applications. In October, Northstar Enterprise + Defence announced that it was partnering with construction materials giant Owens Corning to create lightweight and highly insulated modular data centres, which can be quickly deployed with smaller foundations or on roofs of existing buildings20. US company Armada announced in July 2025 the launch of its largest ever modular data centre Leviathan, liquid cooled and with ten times the compute capacity of its biggest existing module and designed to be stacked or to create megawatt facilities. At the same times it announced a $131 million funding round with a raft of new investors joining existing ones which include Microsoft's venture arm M1221. With ambitions to super-accelerate development timelines, hyperscalers are now working on modular design and delivery for large-scale data centre developments. Details are, of course, under wraps, but the concept is to construct the steel building shell on site and slot in fully contained rack modules, like mini data centres in themselves. Data centre suppliers are developing off-the-shelf modular solutions for larger data centres. In August 2025, Vertiv launched One Core – for data centres between 5MW and 50MW - which sees cooling modules, MEP modules, racks power modules and corridor modules fitted into a steel shell[22]. In November 2025, NVIDIA unveiled the Omniverse DSX Blueprint, an open blueprint for AI data centres which it says allows faster design by virtually assembling equipment and systems from Siemens, Schneider Electric, Trane Technologies and Vertiv to create a digital twin $^{23}$ . Vertiv has claimed that this approach could slash delivery times in half $^{24}$ . # Future proofing One possible advantage to an advanced modular approach could be that upgrades and updates can be more easily managed, with new modules simply replacing old ones. Modular design can also enable staged delivery, where data centre operators can start small and add infrastructure over time as demand ramps up. Looking further ahead, an industrialised approach, with a kit of parts to plug into design could help the data centre industry increase automation. That certainly applies to the design process, but automation in manufacturing facilities – using robots for fitting and installation tasks – is on the far horizon too. # 05 Gas is back in fashion # You need a mix of cheap power, green power and always-available power. The ones that get the mix right will get where they need to go. Robert Kim Head of Growth I&CP Accenture > Electricity grids cannot cope > Natural gas to the rescue >Power Couples:a hybrid approach Power constraints in some of the world's leading data centre locations are well documented. Shortages in locations such as Virginia and North Carolina in the US, Querétaro in Mexico and the European cities of Frankfurt, London, Amsterdam, Paris and Dublin are slowing or stalling data centre developments there. And demand for power from data centres is set to rise dramatically. Research by Accenture indicates that power consumption by data centres in the US could increase to over $7 \%$ of the country’s total electricity consumption by 2028, rising to between $16 \%$ and $23 \%$ by $2033 ^ {25}$ . The focus has been on finding locations with plentiful supplies of power; proximity to renewable energy is a big draw. But, given the widespread need for data centre delivery at speed there was always going to be a shortfall. Looking ahead, balancing the growth of AI with its need for power is one of the biggest challenges the world faces. According to the International Energy Agency (IEA) data centres consumed around $1.5\%$ of global electricity in 2024. The IEA predicts that by 2035, that proportion will have doubled to $3\%$ , although it could be as high as $4.4\%$ if AI takes off faster than expected, and faster adoption of energy efficiency for data centre technology could result in that proportion falling to $2.6\%$ . In the US, where around $50\%$ of the world's data centres are concentrated, the IEA forecasts an $130\%$ increase in electricity consumption to 240 TWh by 2030[26]. # Back to gas While the world ponders this problem, data centre companies have found a new (old) solution to the limitations of electricity grids: natural gas. This was a trend that was emerging in the US at the beginning of 2025, with natural gas companies appearing at data centre events for the first time. Now there are multiple examples of natural gas powering data centres, either for primary or back-up power. In some cases, natural gas generators are being used as a temporary measure, while data centres wait for grid connections or upgrades. Although labelled a bridging solution it could be one that bridges for several years, rather than several months. Where natural gas infrastructure is already available, using natural gas for power generation can shorten the time to operation, and may take less time for permitting. One well-documented example is xAI, which is using natural gas generators for its Colossus data centre in Memphis, allowing it to get up and running in record time. However, xAI's permitting strategy was somewhat creative – achieving permits for the gas turbines after operation began $^{27}$ . A recent deal, announced in October 2025, will see Texas gas company Energy Transfer supply 2.3GW of natural gas capacity to Oracle Cloud Infrastructure's planned data centre developments. Power will be supplied through a modular natural gas system which can supply up to 200MW of power under a minor source air permit[28]. Accenture predicts that by 2030 gas will be supplying nearly $60\%$ of power demand for data centres in the US, leading to a rise of between $8\%$ and $11\%$ in overall gas use for power generation[29]. Other markets are following the US's lead. In Ireland, where there is now limited capacity in the grid, the Irish Times reported in June 2025 that 11 data centres were contracted to connect to the gas network, with four of those waiting for a connection. A further 15 were waiting for a decision on whether they could be connected<sup>30</sup>. In the UK, there were 86 requests for data centres to connect to the gas network in the 12 months to August 2025, according to gas industry association Future Energy Networks<sup>31</sup>. And Alberta, Canada, with its substantial natural gas reserves<sup>32</sup>, is experiencing a surge of interest from data centre developers with over 30 projects proposed as of September 2025<sup>33</sup>. # Hybrid solutions Natural gas is, of course, a fossil fuel. And although it produces fewer greenhouse gas emissions than diesel, it is still a polluter. Hyperscalers are not necessarily turning their backs on renewable energy. Some suggest hybrid solutions, which combine renewable energy and power generation from natural gas - a 'Power Couple' – could be a good way forward. One of the arguments for turning to natural gas is that it allows a more reliable power source, given the fluctuating load demands of AI power centres and variable power from renewable energy sources. There is also an argument that natural gas is a good transition solution until green hydrogen becomes a viable option, since gas pipelines and equipment could potentially be used for hydrogen. Rolls Royce, for instance, highlights the fact that its new range of gas engines for data centres have been designed to operate with a mixture of natural gas and hydrogen or with $100\%$ hydrogen, should that become possible<sup>34</sup>. In the medium term, nuclear power will also become a viable option, either to back up renewable sources, or as the sole power source. The last 12 months have seen a flurry of announcements about plans and partnerships to develop small modular reactors (SMRs) alongside data centre campuses. For instance, Amazon and SMR company X-energy have announced plans to build 12 SMRs near Richmond in Washington with operation mooted to begin sometime in the 2030s<sup>35</sup>. And in the UK, US company Holtec International and EDF UK are eyeing up a former coal-fired power station site where they could develop multiple SMRs on a 100-acre area to support data centres, within a similar timeframe<sup>36</sup>. # 06 Breakthrough cooling solutions These next-generation cooling technologies mean that data centres consuming vast amounts of water is definitely yesterday's news. > Cold plates the norm for AI > Immersion cooling advancing > Microfluidics breakthrough? # William Ferreira # Director - Brazil Soben Part of Accenture As vital to the future of AI as the advancement of chips is the evolution of cooling technology for those chips; higher processing capacity and denser racks creates greater quantities of heat than a cloud data centre does. With cooling systems specialists, hyperscalers and chip manufacturers all hard at work on R&D programmes to find new solutions, 2026 could be the year of a major breakthrough. Microsoft suggested it was getting close to that with the announcement in September 2025 that it had deployed a new system that removed heat three times faster than a cold plate, although to date it had only done this in the lab. Working with Swiss start-up Corinitis, it is developing microfluidics, where tiny channels are etched into the back of a chip so that cooling liquid can flow through the chips themselves to remove heat<sup>37</sup>. Others, including Nvidia and TMSC which makes Nvidia's chips have also been working on microfluidics $^{38}$ . It's an idea that was first aired in the 1980s with research work ramping up seriously around five years ago $^{39}$ . Engineers are also working to make existing cooling technologies more efficient, to reduce the power needed to run them. These include liquid cooling systems, both the more commonly used cold plate - or direct-to-chip - systems and immersion systems which see the entire server immersed in a dielectric liquid. # Cold plate cooling ramps up The use of cold plates, where a cold liquid is circulated in a metal plate sitting on the chip, grew significantly between 2024 and 2025, a trend that is set to continue in 2026. According to Trendforce, the deployment of liquid cooling in Al data centres rose from $14\%$ in 2024 to $33\%$ in 2025, with $40\%$ of Al data centres expected to use them in 2026. As a result in this surge in demand, three of the four cool plate manufacturers which supply to the US have ramped up their production<sup>40</sup>. Microsoft, Google and Meta already use direct-to-chip cooling for the AI racks in data centres while colocation providers Digital Realty and Equinix are building facilities that are ready for the technology $^{41}$ . AWS reported in June 2025 that it had developed its own cold plate system, which it says is adaptable so that it can be added when its needed and has a more powerful and efficient coolant distribution unit than off-the-shelf systems $^{42}$ . Research work is also underway on the design of the channels within the cold plates $^{43}$ . Chip manufacturers are also looking into systems where the cold plates are directly bonded to chips $^{44}$ . # Immersion increases Immersion cooling has pros and cons when compared to cold plate. It is more energy efficient but capital costs are higher, and it would not be a practical option for retrofit. One-phase immersion cooling sees the coolant, usually hydrocarbon based, absorbing heat and circulating through heat exchangers. Two-phase immersion sees the coolant, liquid polyfluoroalkyl substances (PFAS), vaporise as it absorbs the heat to be turned back to liquid by a condenser. While two-phase is more efficient, PFAS contains potentially harmful chemicals that may face regulation in Europe and the US. Whereas 12 months ago, immersion cooling was rarely used, there are signs that in 2026, it will be deployed more widely. In June 2025, Shell launched a new immersion cooling liquid aimed at data centres, Shell DLC Fluid S3, which it claims can improve power usage effectiveness (PUE) by up to $27\%$ [45]. In November 2025, Vertiv introduced its CoolCenter Immersion system in Europe which it says can accommodate up to 240 kW per unit [46]. UK company XDS reports that it is building a series of immersion-cooled data centres in the Middle East: two 10 MW centres in Riyadh and Jeddah which its website says will be constructed by June 2026 and November 2026 respectively and two 7MW data centres with immersion cooling in Dubai which is says will be handed over in $2026^{47}$ . It is worth noting that – at least in the near future – data centres will deploy a mixture of air cooling and versions of liquid cooling. Evaporative cooling which uses vast amounts of water will be phased out; in a world where the changing climate is already causing water shortages, data centres that exacerbate this problem will not be accepted by neighbouring communities. The real challenge for data centre developers is how to futureproof their facilities so that the right cooling systems, with optimised energy efficiency, can be deployed in the right places. Designers and suppliers are turning to advanced modelling and digital to help them optioneer and navigate the changes that might occur. # 07 # Rare earth elements becomerarer Should rare earth shortages deepen, regional priorities may shift—and Europe could face significant supply vulnerability. # Claire Jones Group Marketing Director – Global Soben Part of Accenture >Further restrictions by China > Europe most at risk >Recycling ramps up As 2025 drew to a close, China played a strong hand against the US in its trade negotiations by further restricting its export of rare earth elements (REE) and technologies. Used in several elements of data centre developments and in connected infrastructure, such as fibre optic cables, the possibility of REE shortages in future years is a cloud on the horizon of data centre expansion programmes – particularly in Europe. According to the International Energy Agency (IEA), China was responsible for $60\%$ of the world's REE mining output in 2024, followed by Myanmar, Australia and the United States. It is even more dominant when it comes to separation and refining, delivering $91\%$ of global production. And it also dominates the manufacture of permanent magnets, which contain REE and are widely used in equipment for data centres, accounting for $94\%$ of global production<sup>48</sup>. In April 2025, China introduced export controls on seven REE (samarium, gadolinium, terbium, dysprosium, lutetium, scandium, and yttrium), related compounds, metals and magnets. Then in October 2025 it extended controls on five additional elements (holmium, erbium, thulium, europium and ytterbium) and added further restrictions related to items containing Chinese-sourced REEs and any internationally made products containing its REEs or manufactured using Chinese technologies. It had already banned the export of rare earth extraction and separation technologies in 2023. # Europe at risk? The concern is that, should shortages of available REEs begin to impact on data centre equipment, the US will have to look after its own. And that could leave Europe out in the cold. The good news is that the 12 controlled REEs do not include those used in the manufacturer of chips, according to Taiwan's Ministry of Economy. The bad news is that the company which makes Nvidia's GPU chips – and therefore the majority of GPUs used in data centres - Taiwan's TSMC, relies on suppliers that are dependent on China's REEs $^{49}$ . Restricted REEs are also used in other parts of a data centre. Dysprosium is used (with neodymium) in high-strength permanent magnets for server cooling fans; neodymium and terbium are used in hard disk drives; and erbium and ytterbium are used to amplify signals in fibre optics $^{50}$ . There is a further twist in the supply chain tail in that Dutch company ASML is the world's only supplier of extreme ultraviolet lithography machines which are vital to TSMC's chip manufacturing operations. ASML said in October that it had sufficient reserves to allow production to continue, once China's restrictions came into force. # Planning for the future Though the fluid nature of global trade deals may see changes on restrictions in coming months, the dominance of China in the mining and processing of REEs is stirring the creation of new supply chains. Mining projects are being developed in the United States, Australia, Brazil, Tanzania and India and refining operations are expanding or planned in the US, Australia and Malaysia $^{51}$ . In the shorter term, recycling of REE – currently at below $1 \%$ of consumption $52 \text{‰}$ - becomes a more attractive option. Microsoft, as part of its strategy to become zero waste by 2030, is pursuing the recovery of rare earth elements and precious metals from end-of-life hard disk drives through its Circular Centers. With four up and running in the US, Ireland and Singapore, it has plans for four more in Cardiff, UK, New South Wales, Australia and San Antonio in Texas $^{53}$ . The challenge is that, in a data centre market that demands hundreds of new facilities tomorrow, new mines, processing and recycling operations take years to come to fruition and require significant amounts of investment. However, new players with deep pockets have stepped into the game. US mining companies report a surge in interest from Gulf state investors who are keen to play their part in establishing new supply chains for REE and other critical minerals such as copper, lithium, nickel and cobalt $^{54}$ . # 08 # The race for faster permitting Permitting speed is becoming a competitive differentiator as governments race to attract nationally strategic data centre investment. # Helena Mubiru Senior Marketing Manager EMEA Soben Part of Accenture > National and state level zoning > BYO energy and low water usage helps > AI does permitting leg work Now that data centres have been elevated from behind-the-scenes facilities to nationally important infrastructure, the race between national and state governments to attract new facilities to their regions is on. National and state governments are looking for ways to streamline the permitting process, often with tax incentives thrown in. Increased awareness among the public brings tensions too. With news stories and social media posts about data centre facilities causing droughts and power shortages, challenges from communities are becoming more frequent. Data Center Watch, which tracks opposition to data centres in the US, reports that between March 2024 and March 2025, $18 billion worth of data centre projects were blocked and a further $46bn of projects were delayed due to opposition from residents and activist groups55. Data centre developers, particularly hyperscalers, are also playing their part in easing the permitting process. With concerns over water shortages in many regions, closed loop and liquid cooling can mitigate permitting issues and reduce local opposition to schemes. 'Bring-your-own' energy eases strain on electricity grids and, in the case of natural gas, can lead to quicker permitting. # Getting it right Recent mega projects provide inspiration for how authorities and developers can together create the right recipe for a speedier start on site. Early zoning of land for industrial use or, better still, for data centres can prepare local communities. OpenAI's Stargate 900 MW AI campus is located on a site in Abilene, Texas, that had already been zoned for industrial use. The City of Abilene granted an $85 \%$ property tax abatement in return for the $3.5 billion investment and the project gained environmental permits for natural gas generators $^{56}$ . Originally earmarked for a crypto- mining campus, plans to develop the site for AI data centres were only announced in 2022, and the first data centre came online in September 2025 $^{57}$ . In Mississippi, AWS built one of its data centre campuses for Project Rainier, which will power Anthropic's AI model training. Located in two pre-zoned industrial parks, the deal also saw a reported \(30bn in incentives and saw permits granted for AWS to build dedicated solar parks and make upgrades to the grid. After a special session of the state legislature in January 2024\)^{58}\(, site preparation began in Spring 2024, with buildings expected to come online from 2026. The UK is looking to go down the early zoning approach with its strategy to create AI Growth Zones which have the land, power and local backing to host data centres. Over 200 locations have bid to become an AI Growth Zone with three announced by the Government by mid-November 2025: Culham in Oxfordshire $^{59}$ , Blyth and Cobalt Park in the Northeast of England $^{60}$ and Anglesey and Trawsfynydd in North Wales $^{61}$ . Aragon, Spain, has attracted a huge amount of data centre investment around its capital Zaragoza by easing permitting, adding to the appeal of the region's menu of plentiful green power and proximity to power cables $^{62}$ . However, the region has to some degree become a victim of its own success. Connections to the grid have become a bottleneck $^{63}$ and farmers have blamed data centres for water shortages $^{64}$ , a perception that AWS is looking to counter with a programme of projects to reduce water leaks in the area $^{65}$ . # AI streamlines permitting With the plethora of permitting and planning procedures that any data centre development must go through, and with the myriad variations between countries and states, some hyperscalers are turning to AI for help with the leg work. Microsoft has reported on its use of AI to speed up planning and permitting for nuclear projects, a process which could equally be applied to data centre projects. By using a combination of generative AI for automated drafting, copilot to help permitters find answers to queries among regulatory data and automated pre-submission reviews to identify gaps in information, Microsoft says it has made productivity gains of between 25 and $75 \%$ 66. On the permitting authority side, there are already several examples of governments starting to use AI to speed up their side of the process, initially for more straight-forward residential applications. These include Austin, Los Angeles and Honolulu in the US $^{67}$ and Qatar – where officials say that approval times have reduced from 30 days to 120 minutes $^{68}$ . # 09 # New chips on menu Hyperscalers are no longer willing to rely on a single supplier; in-house and custom AI chips are reshaping the competitive landscape. # Joe Cusick # Operations Lead I&CP Accenture > Nvidia still dominates > Hyperscalers developing their own chips >Global politics creates waves Nvidia continues to dominate the market when it comes to the GPU (graphics processing unit) chips that are used for many AI training and inference tasks. Estimates put its share of the market at around $80\%$ . Given the predicted increase in AI data centre developments, there has been concern that there simply won't be enough chips to go round, although 2025 has seem some significant changes for Nvidia due to trading issues between the US and China. Whereas sales to China had accounted for between $20\%$ and $25\%$ of Nvidia's data centre revenue, the company is now expecting it to be $0\%$ $^{69}$ . Looking to reduce their reliance on one supplier, hyperscalers and other chip manufacturers have been hard at work developing their own chips for AI processing. We expect some of these new chips to be ready for deployment over the next 12 months. # Hyperscaler's in-house chips Google already produces its own range of chips, TPUs (Tensor Processing Units). While Amazon makes its Tranium AI chips for in-house use. There were reports mid-2025 that OpenAI was planning to deploy Google's TPUs. This was later debunked by OpenAI who said that it was merely testing the TPUs $^{70}$ in its AI labs $^{71}$ . Then in October 2025, we learned that OpenAI has been working with Broadcom in a $10 billion deal which will see custom chips produced for OpenAI to be used in its own facilities[72]. Word is that production of OpenAI's chips will begin in 2026. We had been expecting the arrival of Microsoft's Maia 200 chips in 2025. However, after additions to the chip requested by OpenAI reportedly made the chip unstable in simulations, Microsoft has now pushed back mass production of the new chip until 2026 $^{73}$ . AMD is working hard to win market share from Nvidia, planning to launch its next-generation GPU chip in 2026. AMD and OpenAI announced a partnership agreement in October 2025 which could see OpenAI deploy 6GW of AMD's GPUs from the second half of 2026 $^{74}$ . The UK's Arm Holdings is also nipping at Nvidia's heels. In October 2025, Meta announced that it would be using Arm Neoverse GPUs for some of their AI workloads due to their higher energy efficiency $^{75}$ . # China catches up Perhaps even hotter on the heels of Nvidia are chip developers in China, with Nvidia CEO Jensen Huang telling the UK's Financial Times in November 205 that China will win the AI race with the US<sup>76</sup>. China has made some strong power moves involving chips, with its government issuing guidance requiring any data centres receiving state funds to use only chips made in China and ordering any data centres that were less than $30\%$ complete to remove all installed foreign chips and cancel orders<sup>77</sup>. Chinese chip manufacturers are rushing to fill the void. Huawei is the current market leader, but there are start-ups working hard to take Nvidia's place too such as Cambricon Technologies, Moore Threads, Biren Technology, MetaX, Enflame and Hygon Information Technology<sup>78</sup>. Once developed, these chips will almost certainly be exported to other countries that are not limited by trade restrictions. US president Donald Trump threw another spanner into the global chip supply chain works in November 2025. In an interview with CBS, he suggested that Nvidia's most advanced Blackwell chips would only be sold to American companies $^{79}$ . # 10 # The year to get quantum ready Commercial quantum data centres are no longer theoretical; early deployments in New York, Tokyo, Barcelona and London signal the start of a new era. # Dakini Montañez Americas Marketing Manager Soben Part of Accenture > Hyperscalers' secretive deployments > Industry organisation prepares standards >Commercial deployments closer There is consensus that quantum computing will be big. But little consensus on just how big. McKinsey says that it could generate up to $97 billion in revenue from$ 4billion in 2024 $^{80}$ . Bain & Co reckons that it could reach between $5 billion and $15 billion by 2035 $^{81}$ . The World Economic Forum estimates that quantum computing could generate $622 million in value in the financial services sector alone by 2035 $^{82}$ . Stories released publicly show that we are moving towards commercial quantum data centres. In September 2025 Oxford Quantum Circuits (OQC) and Digital Realty, partnering with Nvidia, launched a quantum-AI data centre in New York. Based at Digital Realty's JFK10 site, the facility combines OQCX's GENESIS quantum computer and Nvidia's GH200 Grace Hopper superchips $^{83}$ . As well as Digital Realty's JFK 10 data centre, OQC has also supplied quantum computers to Equinix's TY11 data centre in Tokyo, in CentreSquare's colocated data centres in the UK $^{84}$ – it has LHR2 in London and LHR3-A in Reading $^{85}$ . In November, quantum computing company Qilimanjaro announced the opening of its Quantum Data Center in Barcelona. With 12 quantum computers, the centre aims to give companies, researchers and universities in Europe access to through the cloud<sup>86</sup>. # Special environments The consensus on next steps for quantum computing and data centres is that we will see hybrid arrangements where AI and quantum racks are co-located, since high-performance computing will be needed to carry out supporting processes. The quantum racks would be in dedicated rooms or areas that can provide the right environment. As technology stands today, quantum computers require special treatment: extremely cold and stable environments, isolated from mechanical vibration and shielded fro