Used workstations for local AI: Dell T7810, HP Z840 and similar

A used Xeon workstation is the cheapest good base for a multi-GPU AI machine. Machines like the Dell Precision T7810 and HP Z840 were built for engineering and video work around 2014 to 2016. Today they sell for a few hundred dollars with lots of RAM, and they have what a GPU box needs: several full-length PCIe slots, big power supplies and room for heat.

Why they suit local AI

• Lots of PCIe slots. Several full-size slots, so two GPUs fit without risers in many configurations.
• Big power supplies. The higher-end options were specified for workstation graphics cards. Check the label on the one you're buying.
• Cheap RAM. They use DDR4 ECC server memory, which is plentiful used. 64 GB is common and helps when a model spills over. See how much system RAM you need.
• Built to run all day. Good airflow, solid boards, and usually quiet at idle.
• The CPU barely matters. The GPU does the work. Old Xeons are slow per core, but that's fine for running models on the graphics card.

The two builds we list

Both of these are on our builds page, with October 2026 used prices.

Best value: workstation + 2× Tesla P100, $450 to $650 total. 32 GB of fast GPU memory for 30B-class models. On our rack, two P100s run Qwen3.6-35B-A3B at 53 tok/s, measured.

Serious: workstation + 2× RTX 3090, $1,900 to $2,500 total. 48 GB runs Llama 3.3 70B at 4-bit, about 12 tok/s estimated. This needs a board with two x16 slots and enough physical space between them ($350 to $500 with 64 GB RAM), plus a 1,200 W supply, since most workstation supplies weren't designed for two 350 W gaming cards.

Similar machines

The same idea works with other workstations from that era and a little later: Dell's T7910 and later T7820 and T7920, HP's Z640 (smaller, fewer slots), and Lenovo's ThinkStation P-series towers. The newer models cost more but bring newer CPUs and memory. Check slot count, spacing and power supply rating for the exact configuration.

What to check before you buy

1. Power supply rating. These machines came with several supply options. Look up the one fitted to the unit you're buying.
2. GPU power leads. Workstation supplies often have only a few PCIe power cables, and some use proprietary connectors. Count them and plan adapters. Data-center cards like the P100 and P40 take an 8-pin EPS plug, not a PCIe one, so buy the correct adapter. A wrong cable can damage the card or supply.
3. Slot layout. Check which slots are wired for x16 and whether two thick cards physically fit. Long consumer cards can hit drive cages or RAM shrouds.
4. Second CPU. On dual-socket machines, some PCIe slots only work when the second CPU is installed. Check before you remove one to save power.
5. BIOS and boot. Update the BIOS. Some older machines won't boot with certain data-center cards until a setting for large memory mapping (often called "Above 4G decoding") is enabled, if the board offers it.
6. Noise. Fine at idle, but cards with blower shrouds add noise under load.

The limits

• PCIe 3.0. Slower than modern boards, but it mainly affects how fast models load, not generation speed. See does PCIe bandwidth matter.
• Power draw. Dual Xeons idle higher than a modern desktop. If it runs 24/7, that adds up. See power and heat of local AI.
• No Resizable BAR on many of these, which rules out Intel Arc cards.

Setting it up

Install Linux, the NVIDIA driver and Ollama, then check every card is seen:

For passive cards, sort cooling before the first long run: see cooling passive data-center cards. For the full card guide, read the Tesla P100 for local LLMs.