A Reddit user paid $35 for a rusty crypto mining rig found in a garage and got 54GB of total VRAM — roughly 65 cents per GB — from nine Nvidia P106-100 GPUs that all worked on the first try.

The buyer, posting as markpronkin under the title "54gb vram for 35$," ran Alibaba's Qwen3.6-35B-A3B model locally in Ollama at more than 30 tokens per second on the decade-old hardware. The mixture-of-experts model carries 35 billion parameters in total with 3 billion activated per token, according to Qwen's model card on Hugging Face, and requires a roughly 24GB download.

How did the deal come together?

The seller had discovered the rig in a "damp garage" years after buying the property and listed it on Avito, a Russian classifieds site. The negotiated price was 3,000 rubles, about $35. The buyer "had zero idea if it works, so did the seller," and a pre-purchase test request was refused. The price was low enough to justify the risk.

The host system is unbalanced: a weak Celeron CPU with very little system memory drives nine Pascal-generation cards. Each P106-100 carries 6GB of GDDR5 on a 192-bit bus — about 192 GB/s per card, per Gigabyte's specifications for the model. The OP calls the cards "basically gtx 1060 6gb." Nvidia introduced the mining-specific parts around mid-2017.

What are the software constraints?

Nvidia is winding down support for this generation. CUDA 13.0 removed offline compilation for GPUs below compute capability 7.5, which includes Pascal, and driver branch 580 is "the final [one] that supports GPUs before CC 7.5," according to Nvidia's developer blog. Security updates continue through October 2028.

The OP estimates the model could run faster "with mtp and some optimizations," referring to multi-token prediction.

Why does an MoE model fit the rig?

Not every model will. A dense model such as Qwen3.8-27B is unlikely to run as well, the OP notes. The rig should handle "any MoE model that fits entirely into VRAM decently well." PCIe 2.0 x1 motherboard slots add a further bottleneck.

Capacity math narrows the options:

• Qwen3.6-35B-A3B at 8-bit: 36.9GB — fits
• Llama 3.3 70B at 4-bit: 42.5GB — fits
• Llama 3.3 70B at 5-bit: just under 50GB — leaves little room for context

The 70B 4-bit variant would hit a ceiling around 4.5 tokens per second, the OP calculates, because each token must read all 42.5GB card by card at about 192 GB/s each. Larger models are probably not workable.

What does 65 cents per GB buy against modern hardware?

For comparison, an RTX 5090 with 32GB of VRAM and 1,792 GB/s of memory bandwidth launched at $1,999 in early 2025. The rig offers 1.7 times the VRAM for under 2% of that price, but each card delivers about a ninth of the bandwidth. According to the OP, the mostly compute-only P106 cards trade at roughly 50% of the price of their GTX counterparts.

Electricity is the hidden cost. One commenter estimated the nine GPUs at a combined 1,080W TDP. The OP says power is free at their location and generally inexpensive in Russia.

When another user argued similar performance was achievable on a CPU alone, the OP replied: "On a 35$ device without any optimizations?"

What comes next?

The buyer plans to move the hardware into a case supporting up to an i7-7700K and 32GB of RAM. The new chassis holds 12 GPUs, so adding three more cards would bring the total to 72GB of VRAM. The appeal, the buyer told WiredEntrepreneur, is "unlimited usage and actual privacy." In the OP's framing, more VRAM means more freedom — and the used-mining-card market keeps offering it at pennies per gigabyte.