About this rig
Eight EVGA GeForce RTX 3090 cards on an open mining frame, photographed with the LEDs on. That is 192GB of VRAM in one tower, which at 4 bits per weight holds roughly a 300B parameter model before context. Each card carries 24GB at 936GB/s, and the frame spaces them for airflow instead of enclosing them in a case.
The post that surfaced it is a joke about subscription AI: dad says no to a monthly plan because we have AI at home. The punchline is a wall of 3090s bought for mining and now doing inference.
What the photo does not show: the CPU, the system RAM, the power supplies or how the cards are linked. Treat the host as unknown until the builder posts it. The card count is what the image supports: eight illuminated RTX 3090 labels, top to bottom, with the bottom card cut off by the frame edge.
Why this rig matters for the site’s own math: the 3090 is the card the value tier already recommends, at about a third of a current 5090’s street price for the same 24GB per card. Eight of them at used prices is still cheaper than one current flagship plus the memory to match, and the aggregate capacity is the part you cannot buy on a single consumer card.
Components
Models this rig can run
Computed from 192GB of pooled memory, keeping 8 percent for the OS, runtime and context. Throughput is a ceiling from memory bandwidth, not a measured number.
Setup config
Continue.dev / config
8x EVGA GeForce RTX 3090 24GB on an open mining frame. Host CPU, system RAM and PSU not visible in the photo. Aggregate VRAM 192GB. Per-card bandwidth 936GB/s. Interconnect unknown.