Data Centre GPU Market - By Function (Training, Inference), By Form Factor (PCIe Card, SXM Module, Rack-Scale), By Cooling Technology (Air, Liquid, Immersion), By End User (Cloud Service Providers, Enterprises, Government, HPC), By Region

The four largest US hyperscalers collectively disclosed 2025 capital expenditure guidance of USD 325 billion, representing approximately a 60 percent increase over their 2024 aggregate. Management statements from each hyperscaler have identified AI infrastructure, specifically data centre GPU procurement, as the primary driver of the increase. Amazon's disclosed capex of USD 105 billion includes data centre GPU server procurement, data centre construction, and power infrastructure, with CEO Andy Jassy stating publicly that AI demand was exceeding AWS's ability to supply it. The commercial incentive for hyperscalers to deploy GPU compute is quantified by their cloud GPU pricing: AWS charges USD 4.22 per hour for a single H100 SXM GPU, meaning a fully utilised H100 generates approximately USD 37,000 in annual revenue against a hardware cost of approximately USD 25,000 to USD 30,000 per GPU, a payback period below one year that justifies aggressive procurement even at current elevated pricing.

NVIDIA's Blackwell B200 delivers approximately 2.5 times the training throughput of H100 in NVLink 72 cluster configurations through architectural improvements in the transformer engine and the fifth generation NVLink interconnect. Hyperscalers who completed H100 cluster buildouts in 2023 and 2024 are now ordering Blackwell for their next frontier model training runs, creating an upgrade cycle demand that is additive to the base expansion demand from new capacity additions. The commercial stake in deploying the latest-generation GPU is high: a Blackwell-based training cluster completes the same model training run in 40 percent of the time of an H100 cluster, translating directly into faster product iteration cycles and earlier deployment of improved AI models. NVIDIA's Q4 FY2025 earnings disclosure that Blackwell revenue had exceeded H100 revenue in the quarter confirmed that the upgrade cycle was occurring at the pace that NVIDIA's capacity expansion enabled.

Enterprise organisations in financial services, healthcare, manufacturing, and professional services are deploying data centre GPUs for production AI applications including document processing, recommendation systems, and predictive analytics. This enterprise segment is growing from a smaller base than hyperscaler demand but is expanding the total addressable data centre GPU market beyond the concentrated hyperscaler procurement base. Dell Technologies reported in Q3 FY2025 earnings that enterprise GPU server revenue had grown significantly year-on-year, with growth attributed broadly across financial services and healthcare verticals. Microsoft's Copilot and Azure OpenAI Service create enterprise demand for GPU compute that flows to Microsoft's data centres but also drives enterprise customers to deploy private GPU infrastructure for sensitive data processing requirements. NVIDIA's enterprise software and services revenue, which requires enterprise customers to operate GPU hardware, reached USD 1.5 billion in FY2024, a proxy measure of enterprise data centre GPU deployment scale.

Several national governments have established sovereign AI computing infrastructure programmes that are driving data centre GPU procurement independent of commercial hyperscaler investment. Japan's government announced a JPY 2 trillion commitment to domestic AI infrastructure that includes GPU computing capacity, with SoftBank and NTT Data both recipients of government co-investment for Blackwell-based AI computing facilities. Saudi Arabia's Public Investment Fund and UAE's national AI strategy have committed to large-scale GPU deployments with Microsoft, NVIDIA, and other US technology companies as infrastructure partners. India's government AI mission has announced GPU cluster procurement for national AI research infrastructure. These national programmes represent demand that is geopolitically driven rather than commercially driven and is therefore less sensitive to return-on-investment calculations that govern hyperscaler procurement decisions.

The NVL72 Blackwell rack draws 120 kilowatts of power, compared to 10 to 15 kilowatts per rack for conventional compute. Conventional air-cooled data centre cooling infrastructure provides maximum rack cooling capacity of approximately 15 to 20 kilowatts, meaning that every data centre facility deploying Blackwell-class GPU racks requires liquid cooling installation before the hardware can be energised. The Uptime Institute's 2024 Global Data Centre Survey found that approximately 18 percent of existing data centre facilities have liquid cooling infrastructure capable of supporting GPU density above 50 kilowatts per rack. The capital cost of retrofitting air-cooled data centres with direct liquid cooling or immersion cooling is USD 3 to USD 8 million per megawatt of GPU capacity, a significant addition to the hardware cost that is currently being absorbed by hyperscalers with multi-billion dollar infrastructure budgets but represents a substantial barrier for enterprise data centre GPU deployments. These factors substantially limit data centre GPU market growth over the forecast period.

The US BIS export control rules from October 2023 and their 2024 updates prohibit export of NVIDIA Blackwell B200, Hopper H100, H200, and AMD MI300X to China without specific license. NVIDIA disclosed a USD 5.5 billion inventory and customer commitment charge in Q3 FY2024 related to the China export restriction, representing the direct financial impact of losing access to Chinese hyperscaler and cloud provider GPU procurement. The Chinese data centre GPU market, served by Baidu, Alibaba, Tencent, and ByteDance, is redirecting procurement to Huawei Ascend 910B and 910C, creating a permanent bifurcation of the global data centre GPU ecosystem. These factors substantially limit data centre GPU market growth over the forecast period.

New data centre construction in major markets including Northern Virginia, Dublin, Amsterdam, and Singapore is constrained by electricity grid connection timelines that have extended to 3 to 7 years as existing grid capacity approaches saturation from the concentration of data centre load. Dominion Energy, the primary utility for the Northern Virginia data centre corridor, disclosed in 2024 regulatory filings that it had USD 3 billion in queued data centre connections with lead times extending to 2030 and beyond. The grid constraint means that data centre GPU capacity expansion is ultimately limited by electricity infrastructure build-out that operates on decade timescales rather than by semiconductor manufacturing capacity that operates on year timescales. Microsoft, Google, and Amazon are each pursuing alternative power strategies including nuclear power purchase agreements and on-site generation to accelerate grid-independent data centre development. These factors substantially limit data centre GPU market growth over the forecast period.

Google's Trillium TPU v6, Amazon's Trainium2, Microsoft's Maia 100, and Meta's MTIA v2 are each absorbing AI compute workloads that would otherwise be served by NVIDIA or AMD GPU procurement. Google's TPU programme, now in its sixth generation, has reached the scale where a significant fraction of Google Cloud AI workloads run on TPUs rather than GPUs, representing a growing gap in merchant GPU addressable market. The hyperscaler custom ASIC programmes require 3 to 5 year development cycles and hundreds of millions of dollars per chip generation, limiting the pace of market share shift, but the directional trend toward custom silicon at the largest GPU buyers creates a headwind for merchant data centre GPU revenue growth that compounds over time. These factors substantially limit data centre GPU market growth over the forecast period.

Based on function, the global data centre GPU market is segmented into training and inference. The training segment leads by revenue because frontier model training runs require the highest-specification NVL72 Blackwell configurations at USD 2 to USD 3 million per rack, and the aggregate value of training cluster procurement exceeds inference infrastructure investment. The inference segment is expected to register the fastest growth rate as commercial AI applications at scale require dedicated inference GPU infrastructure that grows with user adoption and is additive to training cluster investment.

Based on form factor, the global data centre GPU market is segmented into PCIe add-in card, SXM module, and rack-scale systems. The SXM module segment leads because NVLink-interconnected SXM GPU configurations deliver the memory bandwidth and inter-GPU communication bandwidth required for large language model training, and hyperscaler procurement is weighted toward SXM configurations for training workloads. The rack-scale segment, represented by NVIDIA NVL72 and equivalent OEM products, is expected to register the fastest growth rate as training cluster scale increases and per-rack GPU count grows from 8 to 72 GPUs.

Based on cooling technology, the global data centre GPU market is segmented by air cooling, liquid cooling, and immersion cooling. Air cooling currently dominates the installed base because the majority of existing data centre infrastructure is air-cooled. Liquid cooling is expected to register the fastest growth rate because Blackwell-class GPU racks at 120 kilowatts require liquid cooling and new GPU data centre construction is being specified with liquid cooling as the baseline. Immersion cooling represents the most thermally efficient option at 97 percent heat removal effectiveness and is being evaluated by several hyperscalers for next-generation GPU cluster deployments.

Based on end user, the global data centre GPU market is segmented into cloud service providers, enterprises, government and national labs, and HPC centres. The cloud service provider segment leads because hyperscaler training cluster procurement represents the single largest GPU procurement by any buyer category. The enterprise segment is expected to register rapid growth as enterprises move production AI workloads onto private GPU infrastructure for data governance, latency, and cost reasons.