Getting hold of a graphics card just for AI development is expensive in 2026. But one modder has used a novel method: repurposing an old Nvidia V100 enterprise GPU as a large language model (LLM) accelerator. As VideoCardz reports, the eight-year-old graphics chip needed a specialized adapter to work properly, but when it was up and running, it proved more capable than an RTX 3060 and RX 7800 XT.
Not bad for a card that should have been destined for the scrap heap.
Put together by the YouTube channel Hardware Haven, this particular mod began with a used SXM2 form factor Nvidia V100 for $100. Considering the card has no display outputs and doesn’t have a viable interface for a modern PC, getting it working wasn’t easy. But apparently someone has made an SXM2-to-PCIe adapter, and with that and a cooling mod, the V100 was more than capable of running LLMs locally.
The adapter cost another $100, and the fan and local taxed another $35, but for under $250, Hardware Haven made a very serviceable AI accelerator.
In the Ollama LLM Benchmark, this bootstrapped V100 with 16GB of HBM2 was able to put out 130 tokens per second, outpacing an RX 7800 XT. In another test using Gemma 4 E4B, it managed 108 tokens per second, compared to the RTX 3060’s mere 76 tokens per second. It needed nearly 200W of power to do that, but when power limited to just 100W, it achieved similar results: 95 tokens on the same test using Gemma 4.
None of this is efficient, as V100 GPUs aren’t easy to come by, and most people are unlikely to manage the mod themselves. But it does raise hopes for the mountains of GPUs that have been bought by AI companies. Some of those firms will fold in the years to come, and there are likely to be thousands of old AI accelerator GPUs that will be worth modding into something useful.
