I got an Nvidia GH200 server for €7.5k on Reddit and converted it to a desktop

Deal, risk, and sourcing

• Many readers call it a “deal of the century” given that a dual-GH200/H100 system was had for ~€7.5k, far below perceived market value.
• Some assumed the hardware must have been “fell off a truck” or scrap-tier, but comments clarify it came from legit Nvidia server OEMs and was heavily used, sold “as is” to avoid support/returns.
• Paying in cash and picking it up in person raised mild “farmhouse in the forest / dirty white van” murder jokes, but others note that large cash transactions are normal in Germany (e.g., buying cars).

Form factor, cooling, and bring-up pain

• GH200 is a Grace CPU + H100 GPU on a custom module, not PCIe add-in cards, so you can’t just drop the GPUs into a normal EATX workstation. The only realistic path is to keep the whole server board and build case/cooling around it.
• Getting the system to boot required arcane driver and platform knowledge: specific datacenter ARM64 drivers and a hack to tell the driver to ignore NVLink so GPUs would initialize over PCIe.
• That hack likely disables the high-speed interconnect, but without it the GPUs were unusable.
• The physical modding (custom frame, liquid loops, fine-pitch soldering with epoxy “insurance”) impressed people as extreme but inspiring “McGyver” engineering.

Performance, economics, and use cases

• Reported throughput: ~100 tokens/s on GLM 4.5 Air (166B), and claims of running 235B–600B-class models at home.
• One commenter does a rough ROI: with heavy batching, ~1M tokens/hour and ~€0.8/hour net profit at €1/1M tokens and typical EU electricity, giving ~1-year payback—while admitting huge uncertainty in utilization and demand.
• Others argue the real value is private, unrestricted models (e.g., answering questions censored by hosted LLMs), on-prem AI for law/medical offices, and “owning the stack,” not pure resale value.

Gaming and desktop viability

• Datacenter GPUs lack outputs; the build adds a low-end Nvidia card for display. Gaming would need GPU copying, streaming, or VMs; plus the ARM CPU means emulation and DRM headaches.
• Several note datacenter cards typically underperform far-cheaper gaming GPUs for games, as drivers and hardware are tuned for compute, not graphics.

Used enterprise hardware market

• Some say this system’s “original $80k” price is misleading; modern RTX 6000/Blackwell cards give similar or better performance in standard workstations.
• Others counter that dual H100s with HBM3 bandwidth and NVLink still beat RTX 6000s on large-model throughput, and that data-center GPU resale can be surprisingly resilient.
• Multiple anecdotes celebrate hunting used servers/GPUs (P40, MI50, V100) and flipping or repurposing them, framing this build as a pinnacle example of that hobby.