NVIDIA Is Building 120,000 GPUs Into UK AI Factories. Each One Generates 700 Watts of Heat. Here Is Why FRP Is the Secondary Specification.
- Aug 12
- 11 min read
Updated: Aug 13
NVIDIA and partners are installing 120,000 Blackwell Ultra GPUs in UK data centres by end of 2026 — the largest AI infrastructure rollout in the country's history. Nscale and Microsoft are building the UK's most powerful supercomputer in Loughton. CoreWeave is establishing a new facility in Scotland. The total investment is up to £11 billion. An AI factory running 120,000 GPUs at 700 watts each generates 84 megawatts of heat that must be removed continuously, every hour of every day. The infrastructure that removes that heat — and the secondary materials that run through and around it — is being specified right now. FRP is the correct secondary specification for the most thermally intensive, electrically dense, and continuously operational facilities ever built in the UK.
Published by Reinforce Technology | 12 August 2026
Jensen Huang, NVIDIA's founder and CEO, described the current moment precisely: "AI is entering its next frontier and driving the largest infrastructure buildout in human history." In the UK, that buildout has a specific and documented scale. NVIDIA and partners including CoreWeave, Microsoft, and Nscale have committed up to £11 billion to install 120,000 NVIDIA Blackwell Ultra GPUs in UK data centres by end of 2026, representing the largest AI infrastructure rollout in British history (NVIDIA, 2025). Nscale and Microsoft are building the UK's most powerful supercomputer in Loughton, powered by more than 24,000 NVIDIA Grace Blackwell Ultra GPUs for Microsoft Azure services. CoreWeave is establishing a new data centre in Scotland powered by renewable energy. Stargate UK, the facility housing NVIDIA Blackwell Ultra GPUs for OpenAI's most advanced models including GPT-5, is operational in Nscale's UK data centres (NVIDIA, 2025).
The physical infrastructure of an AI factory is unlike any data centre that has come before it. A conventional enterprise data centre runs at 5 to 10 kilowatts per rack. A hyperscale cloud data centre handling web-scale workloads runs at 15 to 25 kilowatts per rack. An AI factory running NVIDIA Blackwell Ultra GPUs operates at 60 to 120 kilowatts per rack, with next-generation configurations targeting 200 kilowatts and above. At 120,000 GPUs drawing approximately 700 watts each, the aggregate power draw of the UK's planned NVIDIA GPU deployment is approximately 84 megawatts of continuous electrical load — equivalent to the baseload electricity consumption of a mid-sized UK town, operating continuously, 24 hours a day, generating heat at the same rate as the power consumed (IEA, 2026).
Every watt of that 84 megawatts must be removed as heat to maintain the operating temperatures that GPU performance and longevity require. Air cooling, the conventional thermal management approach for enterprise data centres, cannot manage heat densities of 60 to 120 kilowatts per rack within the physical constraints of standard data centre design. The UK's AI factory buildout is therefore driving the fastest and largest transition to liquid cooling in UK data centre history — direct-to-chip cooling, rear-door heat exchangers, and in the most advanced facilities, full immersion cooling — creating secondary infrastructure environments that are persistently wet, electrically dense, and designed for continuous operation without maintenance downtime across a 25-year asset life. FRP is the secondary specification that the combination of those demands requires.
What Makes an AI Factory Different From a Conventional Data Centre
The term "AI factory" is deliberate and accurate. NVIDIA uses it to distinguish facilities designed specifically for AI model training and inference from conventional data centres designed for general-purpose computing. The distinction is not primarily about the software workloads. It is about the physical infrastructure requirements that the hardware demands.
NVIDIA's Blackwell Ultra GPU architecture is designed for transformer model training at scale, the computational process behind large language models, image generation, and the reasoning systems that underpin GPT-5 and its successors. Training a model at the scale of GPT-5 requires running tens of thousands of GPUs simultaneously, fully utilised, for weeks or months. The heat generated by that sustained full utilisation is not comparable to the intermittent, variable loads of conventional enterprise computing. An AI factory GPU cluster operates at its thermal maximum continuously, and the cooling infrastructure must be designed and maintained to match that continuous maximum load rather than the average load that enterprise data centre cooling systems are sized for.
The power density consequences are structural. A conventional hyperscale data centre might allocate 5 to 10 megawatts of cooling infrastructure for every 10 megawatts of IT load, relying on the diversity of workloads across thousands of servers to reduce peak heat density below nameplate maximum. An AI factory GPU cluster has no such diversity. Every GPU is fully utilised simultaneously. The cooling infrastructure must be sized for 100% utilisation across 100% of the installed GPU count, continuously, with no diversity derating. At the rack level, this means liquid cooling systems delivering coolant flows capable of removing 60 to 120 kilowatts from a single rack, compared with the 5 to 10 kilowatts that computer room air conditioning units manage in conventional facilities.
The secondary infrastructure of a liquid-cooled AI factory therefore operates in conditions that are categorically more demanding than the secondary infrastructure of a conventional data centre: higher coolant flows creating greater moisture exposure risk around fittings and manifolds, higher rack power creating more concentrated DC fault current hazards in overhead cable management, and higher physical density creating less clearance for maintenance access, making the maintenance-free performance of secondary infrastructure across 25 years a more acute requirement than in any previous data centre category.
The Liquid Cooling Infrastructure That 120,000 GPUs Require
NVIDIA's Blackwell Ultra GPU architecture specifies direct liquid cooling as the primary thermal management approach for full rack deployments. The GB300 NVL72 rack, NVIDIA's reference architecture for Blackwell Ultra deployments, uses a rear-door heat exchanger and direct liquid cooling to remove the heat from 72 Blackwell Ultra GPUs per rack at power densities of up to 120 kilowatts. At 120,000 GPUs across the UK deployment, this implies approximately 1,700 racks of this density and type, each requiring coolant supply and return connections from a central mechanical plant, with the interconnecting coolant distribution manifolds running through the overhead secondary infrastructure of the data centre floor.
The coolant distribution infrastructure of a Blackwell Ultra deployment is a new category of secondary engineering in UK data centre construction. Coolant manifolds running at 10 to 20 litres per minute per rack, at pressures of 2 to 4 bar, through overhead secondary distribution from a central chilled water plant, create persistent moisture exposure at every fitting, joint, and connection point in the distribution system. Condensation from cold coolant pipes in the warmer ambient air of the data centre floor adds to the moisture exposure of the secondary structural and cable management elements in the immediate vicinity of the coolant distribution. And the thermal cycling of the coolant system between operational temperature and ambient during shutdown events creates expansion and contraction stresses at the connection interfaces of the secondary structural support elements that carry the coolant manifolds.
FRP structural profiles for coolant distribution manifold supports, FRP cable trays routing the DC power distribution cables that supply 120 kilowatts to each rack, and FRP grating for access walkways between densely packed Blackwell Ultra rack rows provide the combination of non-conductivity, corrosion immunity in coolant-exposed environments, and maintenance-free performance across 25-year asset lives that the AI factory secondary infrastructure environment demands. No alternative secondary specification delivers all three simultaneously (IntechOpen, 2022).
The DC Power Environment at AI Factory Scale
The power distribution architecture of a Blackwell Ultra AI factory operates at voltages and current densities that create the most demanding DC cable management environment in any UK data centre category. NVIDIA's GB300 NVL72 rack draws up to 120 kilowatts at nominal supply voltages, with high-voltage DC bus architecture increasingly used at the rack level to reduce cable losses and improve power conversion efficiency. The DC bus architecture that AI factories use to deliver power at rack level operates at voltages of 380V to 800V DC, with DC bus cables routing from the rack power shelf to individual GPU baseboard management controllers across the rack interior.
The DC arc fault hazard at 380V to 800V DC bus voltage is specific and well-documented. DC arcs at these voltages sustain themselves without the natural zero-crossing extinguishing mechanism of AC arcs, and in a metallic cable tray environment can propagate along the tray infrastructure to adjacent racks or cable management segments. In an AI factory where adjacent racks are all operating at full power simultaneously, a DC arc fault propagating through metallic cable management infrastructure could initiate a cascading failure across multiple racks — interrupting model training runs that may have been executing for weeks and causing financial losses that vastly exceed the cost difference between FRP and steel cable management at the specification stage.
FRP cable trays in AI factory DC power distribution environments are non-conductive throughout, with volume resistivity of 10¹² to 10¹⁶ Ω·m. A DC arc fault in a cable within an FRP cable tray is contained at the fault location and cannot propagate through the tray infrastructure to adjacent cable management. The financial and operational consequences of DC arc fault containment failure in an AI factory — lost model training time, damaged GPU hardware, potential facility downtime — are the most severe in any UK data centre application category, making the non-conductive specification argument for FRP cable management in AI factories more commercially compelling than in any conventional data centre environment (IntechOpen, 2022).
The Loughton Supercomputer, Stargate UK, and CoreWeave Scotland
The three principal UK AI factory deployments announced by NVIDIA and partners represent three distinct data centre infrastructure environments where FRP secondary specification is directly relevant.
The Loughton supercomputer — the UK's most powerful, featuring more than 24,000 NVIDIA Grace Blackwell Ultra GPUs for Microsoft Azure — is a purpose-built hyperscale AI facility in Essex whose thermal and power density requirements will make it among the most liquid-cooling-intensive facilities in Europe on completion. The secondary cable management, coolant distribution structural supports, and access walkways of a facility of this scale and density represent a substantial secondary infrastructure specification decision whose consequences compound across the 25-year operational life of a facility that Microsoft is building as permanent sovereign AI infrastructure (NVIDIA, 2025).
Stargate UK, the Nscale facility housing NVIDIA Blackwell Ultra GPUs for OpenAI's GPT-5 and successor models, is an operational AI factory whose cooling infrastructure is being commissioned as the GPU deployment scales to full capacity in 2026. The real-time operational pressure to bring AI inference capacity online — with OpenAI's commercial model training and serving revenue depending on the availability of this infrastructure — creates a construction programme where the secondary specification decisions on cable management and cooling support are being made at pace. Getting them right now, with FRP, is significantly more achievable than retrofitting after the facility is fully operational with 60,000 GPUs running continuous inference loads.
CoreWeave's Scotland facility, powered by renewable energy, adds a geographic dimension to the UK AI factory buildout that extends the liquid cooling and high-density power distribution infrastructure requirements beyond the established London and South East data centre corridors. Scottish atmospheric conditions — higher rainfall, more frequent temperature extremes between summer and winter, and in coastal locations marine atmospheric chloride — create a more aggressive outdoor secondary infrastructure environment for the external cooling plant of the facility than the milder urban environments of most UK hyperscale data centres. FRP structural profiles and cable management for the external cooling plant infrastructure of a Scottish AI factory site provide corrosion-immune performance in the outdoor conditions of Scotland's climate that galvanised steel secondary infrastructure would accumulate maintenance liabilities against from the first years of operation (NVIDIA, 2025).
The Infrastructure Buildout Timeline and the Specification Window
NVIDIA and partners have committed to having AI factories operational by end of 2026. That is a four-month construction and commissioning window for the balance of the GPU deployments not yet operational. In data centre construction programmes operating under this kind of timeline pressure, the secondary infrastructure specification decisions — cable tray systems, coolant distribution structural supports, access walkways — are being made now and procured now, not at a future date when there is more time to evaluate alternatives.
The consequence of a secondary infrastructure specification decision made under timeline pressure with a suboptimal outcome — galvanised steel cable management in a liquid cooling environment that generates coolant moisture exposure, or steel cable management in a DC bus power distribution environment where arc fault propagation is a documented risk — is a maintenance liability that accumulates across 25 years of continuous AI factory operation. At the scale of 120,000 GPUs generating commercial revenue for Microsoft Azure, OpenAI, and CoreWeave's enterprise customers, the operational disruption cost of secondary infrastructure maintenance access in a fully loaded AI factory is many orders of magnitude larger than the purchase price difference between FRP and steel at the specification stage.
NVIDIA and partners are building the largest AI infrastructure in UK history. 120,000 Blackwell Ultra GPUs. £11 billion of investment. 84 megawatts of continuous heat load requiring liquid cooling at densities that have no precedent in UK data centre history. FRP cable trays, structural profiles, and grating are non-conductive in the DC bus power environment, corrosion-immune in the coolant-exposed liquid cooling infrastructure, and maintenance-free across the 25-year asset lives of the facilities that will run Britain's sovereign AI capability into the 2050s. The specification decisions being made on these facilities right now will determine whether that secondary infrastructure performs across those 25 years or generates the maintenance events that AI factory operators cannot afford to schedule.
Reinforce Technology FRP Products for AI Factory Infrastructure
Reinforce Technology supplies FRP cable trays, structural profiles, grating, and handrail systems for AI data centre and AI factory infrastructure across the UK. Available in polyester and vinyl ester resin systems for coolant-exposed and general data centre environments. Non-conductive throughout, eliminating DC arc propagation risk and earthing and bonding requirements. Maintenance-free across 25-year asset lives in the persistently wet, electrically dense secondary infrastructure environments of liquid-cooled AI factory facilities.
Contact us to discuss your AI factory project and the correct FRP secondary infrastructure specification for your cooling configuration, power density, and operational horizon.
Final confirmation of suitability for any specific AI data centre application, including cable management fire performance requirements and structural loading for coolant distribution support systems, remains the responsibility of the appointed project engineer. Reinforce Technology provides technical guidance and material recommendations based on information supplied to us, but specification sign-off should always sit with the qualified professional responsible for the design.
References
IEA (2026) Electricity 2026. International Energy Agency. Available at: https://www.iea.org [Accessed: 12 August 2026]. [Global data centre electricity consumption forecast to reach 945 TWh by 2030; AI primary driver of growth; power usage to quadruple within five years for AI-optimised facilities].
IntechOpen (2022) 'Fibre-Reinforced Polymer (FRP) in Civil Engineering', in IntechOpen Engineering Series. Available at: https://www.intechopen.com/chapters/84203 [Accessed: 12 August 2026]. [Non-conductive; volume resistivity 10¹² to 10¹⁶ Ω·m; corrosion-immune in moisture-exposed environments; maintenance-free across 25-year design life; 75% lighter than steel].
NACE International (2016) International Measures of Prevention, Application and Economics of Corrosion Technology (IMPACT). Houston, TX: NACE International. Available at: http://impact.nace.org/economic-impact.aspx [Accessed: 12 August 2026].
NVIDIA (2025) NVIDIA and United Kingdom Build Nation's AI Infrastructure and Ecosystem to Fuel Innovation, Economic Growth and Jobs. Available at: https://nvidianews.nvidia.com/news/nvidia-and-united-kingdom-build-nations-ai-infrastructure-and-ecosystem-to-fuel-innovation-economic-growth-and-jobs [Accessed: 12 August 2026]. [Up to £11 billion investment; 120,000 NVIDIA Blackwell Ultra GPUs by end 2026; Nscale and Microsoft Loughton supercomputer with 24,000+ Grace Blackwell Ultra GPUs; CoreWeave Scotland renewable-powered facility; Stargate UK for OpenAI GPT-5; Jensen Huang quote on largest infrastructure buildout in human history].
ScienceDirect (2025) 'Sustainable composites for metal replacement: Environmental assessment and material selection of fiber-reinforced polymer across industries', ScienceDirect, doi: 10.1016/S2667-3789(25)00051-3. Available at: https://www.sciencedirect.com/science/article/pii/S2667378925000513 [Accessed: 12 August 2026]. [Pultruded GFRP manufacturing emissions approximately 60 to 70% lower per tonne than primary steel, cradle-to-gate].
Younis, A., Ebead, U. and Judd, S. (2018) 'Life cycle cost analysis of structural concrete using seawater, recycled concrete aggregate, and GFRP reinforcement', Construction and Building Materials, 175, pp. 135-364. doi: 10.1016/j.conbuildmat.2018.04.183.
Comments