NVIDIA DGX A100 Maintenance
side by side.

DGX A100 · Ampere generation

DGX A100 320GB (8x A100 40GB) · DGX A100 640GB (8x A100 80GB)

DGX H100 / H200 · Hopper generation

DGX H100 (8x H100 SXM5 80GB) · DGX H200 (8x H200 141GB)

DGX GH200 · Grace Hopper

DGX GH200 (Grace+H100, 256x GH200 superchips, 144TB memory)

DGX GB200 NVL72 · Grace Blackwell

DGX GB200 NVL72 (72x B200 GPU + 36x Grace, liquid-cooled rack-scale)

Components

GPU boards · NVLink/NVSwitch · power supplies · cooling loop (GB200) · networking modules

Generation status of DGX line

DGX maintenance is the absolute highest-value lever in the TPM market. NVIDIA factory warranty for DGX typically runs 1-3 years, then standard or premium support becomes due — proportional to hardware value: typically 15-25 percent per year. For platforms with very high hardware value such as DGX H100 or DGX GB200 NVL72, maintenance thus becomes one of the largest recurring cost items in the data center. TPM reduces this to 30-50 percent — for an 8-box DGX cluster, these savings apply to every box — a substantial annual lever.

We service DGX platforms with OEM components from our own warehouse and certified refurbishing sources — DGX hardware is proprietary (NVLink topology, NVSwitch backplane, NVIDIA-specific power delivery, plus liquid cooling loop from GB200), component availability is structurally weaker than for standard servers. We continuously build the component pool for current DGX generations, with focus on GPU board replacement, NVSwitch modules and power supplies. Our engineers are specifically trained for DGX architecture — DGX maintenance is not comparable to standard x86 server service. CUDA, AI Enterprise, Base Command stay license-free active under TPM (software subscriptions run independently of hardware maintenance with NVIDIA).

The full OEM-vs-TPM matrix (spare parts, SLA tiers, language, sub-outsourcing, compliance docs, reporting) is the same for every Nvidia line — see the Nvidia battle card.

BAIT/MaRisk documentation package, DORA Art. 28 contract template, ISO 27001 and GDPR-compliant media handling apply identically to every Nvidia line — details on the Nvidia battle card.

Same four steps for every Nvidia line: inventory → fixed-price quote in 48 h → contract (BAIT/MaRisk/DORA hooks) → hand-over with spare-parts provisioning. Details on the Nvidia battle card.

Which DGX models do you service?

Current DGX family: DGX A100 (Ampere, 320GB and 640GB variants), DGX H100 (Hopper, 8x H100 SXM5 80GB), DGX H200 (Hopper refresh, 8x H200 141GB), DGX GH200 (Grace Hopper, 256x GH200 superchips with 144TB memory) and DGX GB200 NVL72 (Grace Blackwell, 72x B200 GPU plus 36x Grace, liquid-cooled rack-scale). Including all GPU boards, NVLink/NVSwitch modules, power supplies, cooling loop (GB200) and networking modules. For older DGX-1/-2 we have a separate spoke (DGX Legacy).

What does TPM cost for DGX compared to NVIDIA support?

30 to 60 percent savings. Specifically: for DGX A100 640GB with 24×7×4, TechCare is typically roughly 60% below NVIDIA premium support, as it is for DGX H100 and DGX GH200; DGX H200 similar. For the DGX GB200 NVL72 it is roughly 55–60% below NVIDIA premium. For an 8-box DGX H100 cluster, TechCare is also roughly 60% below the NVIDIA price — year after year.

Do CUDA, AI Enterprise and Base Command Manager continue to work without an NVIDIA contract?

Yes. CUDA toolkit, GPU drivers and all DGX-OS functions continue to work license-free on the hardware — all AI workloads (PyTorch, TensorFlow, Triton Inference, NeMo) stay functional. NVIDIA AI Enterprise as software subscription runs separately from hardware maintenance — customers actively using the subscription keep it with NVIDIA. Base Command Manager (BCM, cluster orchestration) is subscription-based and independent of maintenance contract. CUDA updates and newer GPU driver versions are freely available from NVIDIA — TPM customers can download them without restriction. Firmware updates on NVSwitch/BMC require an active NVIDIA contract but are usually uncritical with stable AI workloads.

How is component availability for DGX hardware?

Structurally weaker than for standard servers, but addressed with targeted component pool buildup. DGX hardware is proprietary — NVLink topology, NVSwitch backplane, NVIDIA-specific power delivery boards. We stock replacement GPU boards, NVSwitch modules, power supplies and networking modules for DGX A100/H100/H200. For very new generations (GH200, GB200), pool depth is limited — there we recommend hybrid setups: TPM for standard components (power supplies, fans, storage) plus selective NVIDIA subscription for GPU board coverage. At contract conclusion we create a risk assessment per DGX model — for critical workloads we recommend on-site spare components (negotiable in contract).

Which SLA levels do you recommend for DGX?

DGX in productive AI training workloads: 24×7×4 with German-speaking onsite engineer is standard. AI training jobs are often multi-day — outage of a DGX box mid-training means loss of training progress and possibly data loss. For AI inference production workloads (customer-facing LLMs, computer vision pipelines), 24×7×4 is mandatory due to direct service impact. For DGX in test/dev environments or as backup compute, 5×9 NBD can be economical. DGX GB200 NVL72 as liquid-cooled rack-scale system: 24×7×4 mandatory due to cooling loop complexity — cooling failure has high impact.

When is the natural entry point for TPM with DGX?

Factory warranty expiration. NVIDIA factory warranty for DGX typically runs 1-3 years depending on contract. Specifically: DGX A100 fleets from 2020-2022 have been out of factory warranty since 2023+ and are the natural TPM entry point. DGX H100 from 2022-2023 will run out of factory warranty from 2025-2026. DGX H200/GH200/GB200 are current — TPM is to be planned 1-3 years ahead here. We recommend: inventory check 6 months before warranty expiration, TPM contract from day 1 after warranty end — no gap, no NVIDIA standard support extension at full price as bridge.