Someone scrolling a used hardware listing stops on a card that should not exist.
An RTX 2080 Ti with 22GB of memory.
The card shipped with 11.
The buyer pulled the 1GB memory chips and fitted 2GB ones, then sent the board to PCGamesHardware for testing.
In ComfyUI, generating images with Krea-2 Turbo and Qwen 3-VL 4B, the 22GB card returned a result in 315 seconds.
The untouched 11GB version took 540.
A 16GB RTX 5070 Ti, straight off a shelf, finished the same job in 85 seconds.
That gap is the useful part.
Doubling memory on old silicon helped, because the models stopped running out of room.
It still could not touch a modern card, because the memory, the tensor cores and the architecture were several generations behind.
Modded boards like this sell for under $650 on eBay.
Anyone trying to get a model running before committing a quarter of a budget knows that pressure.
You can buy capacity cheap.
You can also buy time, and time is what puts a working demo in front of a customer.
Renting compute sidesteps the trade.
RTX 3090 capacity on BHK Cloud runs at $0.15/hr, with 24GB of memory and no soldering iron involved.
Buy Now Pay Later, zero upfront, payment after one month.
Pricing is at https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w45&utm_content=evening What is the longest run you have waited out because a new card was not in the budget?
Originally posted on linkedin_personal