It used to be that getting more video memory meant buying a new graphics card. That's the opening premise of Linus Tech Tips' new video, in which Luke Lafreniere travels to Michigan to meet Greg of GPU Lab and upgrades an RTX 4090 to 48GB of VRAM. It's a great piece of hardware theater, and the question it raises is more interesting than the soldering: why would anyone go to this much trouble in 2026?
The short answer is AI, and that's what makes this more than a stunt. The project isn't really about gaming. As Luke frames it at the start, the goal is running local AI models on your own machine rather than renting someone else's. Two cards, each going from 24GB to 48GB, is the plan.
How the upgrade actually works
This isn't a matter of clipping on extra chips. The GPU and memory come off the original card and move to a different board, a custom circuit board from China, so the process is closer to a transplant than an upgrade.
Greg explains the layout at 3:21: the 48GB configuration mirrors the RTX 3090, with memory on both sides of the board. From there the video walks through the work in order. Memory goes on the back of the new board under a microscope, then through a rework station. The GPU core and memory are pulled from the donor cards at 8:55. Both get cleaned and reballed with fresh solder from 13:51, then the core is lined up with a prism camera and soldered down at 20:06.
The payoff comes at 24:04, when the card boots and passes memtest_vulkan on Windows. That's a memory test, not a measure of how the card performs at any particular job, and the video isn't a benchmark comparison. It shows the card works. It doesn't show what the extra memory is worth to you.
Why this matters beyond one card
Here's my read, and it's analysis rather than anything the video states outright. A niche repair skill is being pulled toward the mainstream because the economics of memory have changed. Luke's argument in the opening minute is that enterprise AI demand has pushed hardware beyond ordinary buyers, leaving enthusiasts to find scrappy workarounds. Luke also makes the case that the industry would like people to forget what the "personal" in personal computer means.
That fits a bigger pattern on the supply side. Xarmo reported this week that AMD now calls itself "data center first" and expects PC shipments to fall. Greg's own workflow carries the same mark: the memory modules used here are harvested and reused, and sourcing them comes up at 5:49. When a good route to more VRAM runs through a specialist's bench, that tells you who the market is currently serving.
It's worth being clear about the limits, too. This is skilled work on expensive cards, not a weekend project, and the video shows a blower cooler, a long bake and a lot of specialist equipment. If you game on a 4090, nothing here says you need it. The audience is people running local models who've run out of memory.
What viewers made of it
The comments were warm, and mostly about the format. Many viewers praised seeing the whole process instead of a quick talking-head summary, and were glad to see Luke presenting. Greg's explanations got plenty of credit, and several viewers guessed the shop is about to get busier. A recurring gripe was that this route can look cheaper than buying a card with 48GB in the first place, which is viewer opinion rather than a price check.
So what does a 48GB RTX 4090 say? That when the industry's priorities move elsewhere, some of the demand doesn't vanish. It just turns up at a workbench in Michigan.
From the creators
Watch the video
Reporting notes
