{"title":"AI Servers: The 2026 Data Study","publisher":"Servnet (servnetuk.com)","url":"https://www.servnetuk.com/research/ai-servers","updated":"2026-07-05","licence":"Free to cite with attribution + a link to servnetuk.com/research/ai-servers","notes":"Hardware specs are from manufacturer datasheets (static). Prices/market figures carry a source and date and are indicative. No hardware street prices are published (quote-gated & volatile).","gpuSpecs":[{"gpu":"NVIDIA H100 SXM5","arch":"Hopper","vram":"80 GB HBM3","bandwidth":"3.35 TB/s","tdp":"up to 700 W","interconnect":"NVLink 4th-gen 900 GB/s","srcId":"h100"},{"gpu":"NVIDIA H100 PCIe","arch":"Hopper","vram":"80 GB HBM2e","bandwidth":"2.0 TB/s","tdp":"350 W","interconnect":"NVLink bridge 900 GB/s (2-GPU)","srcId":"h100-pcie","note":"NVLink bridge is 900 GB/s (reuses A100 bridges); the 600 GB/s figure belongs to the dual-GPU H100 NVL card."},{"gpu":"NVIDIA H200 SXM","arch":"Hopper","vram":"141 GB HBM3e","bandwidth":"4.8 TB/s","tdp":"up to 700 W","interconnect":"NVLink 4th-gen 900 GB/s","srcId":"h200"},{"gpu":"NVIDIA B200 SXM","arch":"Blackwell","vram":"180 GB HBM3e (DGX/HGX)","bandwidth":"8 TB/s","tdp":"~1,000 W (config. max)","interconnect":"NVLink 5th-gen 1.8 TB/s","srcId":"dgx-b200","note":"192 GB is the raw dual-die capacity; DGX/HGX B200 configures 180 GB/GPU. TDP is a configurable maximum."},{"gpu":"NVIDIA GB200 (superchip)","arch":"Grace + 2× Blackwell","vram":"372 GB HBM3e + 480 GB LPDDR5X","bandwidth":"16 TB/s (HBM)","tdp":"superchip","interconnect":"NVLink-C2C 3.6 TB/s per superchip","srcId":"gb200"},{"gpu":"NVIDIA L40S","arch":"Ada Lovelace","vram":"48 GB GDDR6 (ECC)","bandwidth":"864 GB/s","tdp":"350 W","interconnect":"No NVLink; PCIe Gen4 x16","srcId":"l40s","note":"No NVLink — an inference/graphics part, not a training GPU."},{"gpu":"NVIDIA A100 80 GB SXM","arch":"Ampere","vram":"80 GB HBM2e","bandwidth":"2.04 TB/s","tdp":"400 W","interconnect":"NVLink 3rd-gen 600 GB/s","srcId":"a100"},{"gpu":"AMD Instinct MI300X","arch":"CDNA 3","vram":"192 GB HBM3","bandwidth":"5.3 TB/s","tdp":"750 W (TBP)","interconnect":"Infinity Fabric; PCIe Gen5 x16","srcId":"mi300x","note":"The leading non-NVIDIA alternative."},{"gpu":"AMD Instinct MI325X","arch":"CDNA 3","vram":"256 GB HBM3e","bandwidth":"6.0 TB/s","tdp":"1,000 W (TBP)","interconnect":"Infinity Fabric; PCIe Gen5","srcId":"mi325x"}],"aiServers":[{"model":"PowerEdge XE9680","vendor":"Dell","gpus":"8× HGX H100/H200 SXM (or MI300X / Gaudi3 OAM)","form":"6U","cooling":"Air","srcId":"dell-xe9680","sells":true},{"model":"PowerEdge XE9640","vendor":"Dell","gpus":"4× H100 SXM (NVLink) or Intel Max OAM","form":"2U","cooling":"Direct liquid (facility water)","srcId":"dell-xe9640","sells":true},{"model":"PowerEdge XE8640","vendor":"Dell","gpus":"4× HGX H100 SXM5","form":"4U","cooling":"Closed-loop liquid + fans","srcId":"dell-xe8640","sells":true},{"model":"ProLiant DL380a Gen12","vendor":"HPE","gpus":"up to 8× double-wide PCIe (RTX PRO 6000, H200 NVL, H100 NVL, L40S)","form":"4U","cooling":"Air or direct liquid","srcId":"hpe-dl380a","sells":true},{"model":"Cray XD670","vendor":"HPE","gpus":"8× HGX H100/H200 SXM5","form":"5U","cooling":"Air, with liquid option","srcId":"hpe-cray-xd670","sells":true},{"model":"ThinkSystem SR675 V3","vendor":"Lenovo","gpus":"up to 8× PCIe, or 4× HGX H200 SXM (NVLink)","form":"3U","cooling":"Neptune hybrid (HGX variant)","srcId":"lenovo-sr675","sells":true},{"model":"ThinkSystem SR780a V3","vendor":"Lenovo","gpus":"8× HGX H100/H200/B200 (NVLink 900 GB/s)","form":"5U","cooling":"Neptune direct liquid + air","srcId":"lenovo-sr780a","sells":true},{"model":"SYS-821GE-TNHR","vendor":"Supermicro","gpus":"8× HGX H100/H200 (NVLink + NVSwitch)","form":"8U","cooling":"Air (liquid variants exist)","srcId":"smci-821ge"},{"model":"DGX H100 / H200","vendor":"NVIDIA","gpus":"8× H100 (640 GB) / H200 (1,128 GB) SXM","form":"8U","cooling":"Air","srcId":"dgx-h100"},{"model":"DGX B200","vendor":"NVIDIA","gpus":"8× Blackwell (1,440 GB HBM3e)","form":"10U","cooling":"Air","srcId":"dgx-b200"},{"model":"GB200 NVL72","vendor":"NVIDIA","gpus":"72× Blackwell + 36 Grace (rack-scale)","form":"Rack","cooling":"100% liquid","srcId":"gb200"},{"model":"G593-SD0","vendor":"Gigabyte","gpus":"8× HGX H100 SXM5","form":"5U","cooling":"Air","srcId":"giga-g593"}],"cloudGpuRates":{"asOf":"mid-2025","note":"On-demand list rates move frequently and vary by region — treat as indicative and verify live before deciding.","awsCut":"AWS cut on-demand NVIDIA GPU EC2 pricing by up to ~45% (44% off P5/H100), effective 1 June 2025.","rows":[{"provider":"AWS (p5.48xlarge, H100)","perGpuHr":"~$4.1","srcId":"aws-cut"},{"provider":"Google Cloud (A3-high, H100)","perGpuHr":"~$3.00","srcId":"aws-cut"},{"provider":"Microsoft Azure (NC H100 v5)","perGpuHr":"~$6.98","srcId":"aws-cut"}]},"marketStats":[{"value":"~$6.7 trillion","label":"Global data-centre capex needed by 2030 to meet compute demand (~$5.2tn of it AI-capable)","srcId":"mckinsey"},{"value":"~219 GW","label":"Projected global data-centre capacity by 2030 (nearly 3×), ~70% of new demand from AI","srcId":"mckinsey"},{"value":"61%","label":"Western-European CIOs who say geopolitics will raise reliance on local/regional cloud providers","srcId":"gartner-sov"},{"value":">75%","label":"Enterprises outside the US expected to have a digital-sovereignty strategy by 2030 (Gartner)","srcId":"gartner-sov"}],"sources":[{"id":"hgx","name":"NVIDIA — HGX Platform","url":"https://www.nvidia.com/en-us/data-center/hgx/","kind":"primary"},{"id":"nvlink","name":"NVIDIA — NVLink & NVLink Switch","url":"https://www.nvidia.com/en-us/data-center/nvlink/","kind":"primary"},{"id":"h100","name":"NVIDIA — H100 Tensor Core GPU","url":"https://www.nvidia.com/en-us/data-center/h100/","kind":"primary"},{"id":"h100-pcie","name":"NVIDIA — H100 PCIe Product Brief (PB-11133-001)","url":"https://www.nvidia.com/en-us/data-center/h100/","kind":"primary"},{"id":"h200","name":"NVIDIA — H200 Tensor Core GPU","url":"https://www.nvidia.com/en-us/data-center/h200/","kind":"primary"},{"id":"dgx-b200","name":"NVIDIA — DGX B200","url":"https://www.nvidia.com/en-us/data-center/dgx-b200/","kind":"primary"},{"id":"gb200","name":"NVIDIA — GB200 NVL72","url":"https://www.nvidia.com/en-us/data-center/gb200-nvl72/","kind":"primary"},{"id":"l40s","name":"NVIDIA — L40S","url":"https://www.nvidia.com/en-us/data-center/l40s/","kind":"primary"},{"id":"a100","name":"NVIDIA — A100 datasheet","url":"https://www.nvidia.com/en-us/data-center/a100/","kind":"primary"},{"id":"dgx-h100","name":"NVIDIA — DGX H100/H200 User Guide","url":"https://docs.nvidia.com/dgx/dgxh100-user-guide/introduction-to-dgxh100.html","kind":"primary"},{"id":"mi300x","name":"AMD — Instinct MI300X data sheet","url":"https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html","kind":"primary"},{"id":"mi325x","name":"AMD — Instinct MI325X data sheet","url":"https://www.amd.com/en/products/accelerators/instinct/mi300/mi325x.html","kind":"primary"},{"id":"dell-xe9680","name":"Dell — PowerEdge XE9680","url":"https://www.dell.com/en-us/shop/ipovw/poweredge-xe9680","kind":"vendor"},{"id":"dell-xe9640","name":"Dell — PowerEdge XE9640","url":"https://www.dell.com/en-us/blog/dell-poweredge-xe9640-gpu-diversity-density-and-cooling/","kind":"vendor"},{"id":"dell-xe8640","name":"Dell — PowerEdge XE8640 spec sheet","url":"https://www.delltechnologies.com/asset/en-us/products/servers/technical-support/poweredge-xe8640-spec-sheet.pdf","kind":"vendor"},{"id":"hpe-dl380a","name":"HPE — ProLiant Compute DL380a Gen12","url":"https://www.hpe.com/us/en/compute/hpe-proliant-compute/dl380a-gen12.html","kind":"vendor"},{"id":"hpe-cray-xd670","name":"HPE — Cray XD670 QuickSpecs","url":"https://www.hpe.com/psnow/doc/a50004292enw","kind":"vendor"},{"id":"lenovo-sr675","name":"Lenovo Press — ThinkSystem SR675 V3","url":"https://lenovopress.lenovo.com/datasheet/ds0151-lenovo-thinksystem-sr675-v3","kind":"vendor"},{"id":"lenovo-sr780a","name":"Lenovo Press — ThinkSystem SR780a V3","url":"https://lenovopress.lenovo.com/datasheet/ds0182-lenovo-thinksystem-sr780a-v3","kind":"vendor"},{"id":"smci-821ge","name":"Supermicro — SYS-821GE-TNHR","url":"https://www.supermicro.com/en/products/system/gpu/8u/sys-821ge-tnhr","kind":"vendor"},{"id":"giga-g593","name":"Gigabyte — G593-SD0","url":"https://www.gigabyte.com/Enterprise/GPU-Server/G593-SD0-LAX1","kind":"vendor"},{"id":"mckinsey","name":"McKinsey — “The cost of compute” (2025)","url":"https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-cost-of-compute-a-7-trillion-dollar-race-to-scale-data-centers","kind":"analyst"},{"id":"gartner-sov","name":"Gartner — geopolitics & digital sovereignty survey (Nov 2025)","url":"https://www.gartner.com/en/newsroom/press-releases/2025-11-12-gartner-survey-reveals-geopolitics-will-drive-61-percent-of-cios-and-information-technology-leaders-in-western-europe-to-increase-reliance-on-local-cloud-providers","kind":"analyst"},{"id":"aws-cut","name":"AWS — up to 45% GPU EC2 price reduction (Jun 2025)","url":"https://aws.amazon.com/blogs/aws/announcing-up-to-45-price-reduction-for-amazon-ec2-nvidia-gpu-accelerated-instances/","kind":"primary"}]}