Yeah. What does that have to do with home setups? No one is putting an H200 or L40 in their homelab.

Who the fuck buys a consumer GPU for AI?

Plenty. Consumer GPU + CPU offloading is a pretty common way to run MoEs these days, and not everyone will drop $40K just to run Deepseek in CUDA instead of hitting an API or something.

I can (just barely) run GLM-4.5 on a single 3090 desktop.

Most people are buying Nvidia because that's what's commonly recommended on reviews. "Want to use AI? Buy Nvidia! Want the latest DX12+ support? Buy Nvidia! Want to develop videogames or encode video? Buy Nvidia! Want to upgrade to Windows 11? Buy Nvidia!" Nonstop Nvidia adverts everywhere, with tampered benchmarks and whatnot. Other brands' selling points aren't well known and the general notion is that if it's not Nvidia it sucks.

Seems like a weird strategy to not compete out of fear of success.

Basically, consumer VRAM is dirt cheap, not too far from DDR5 in $/gigabyte. And high VRAM (especially 48GB+) cards are in high demand.

But Nvidia charges through the nose for the privilege of adding more VRAM to cards. See this, which is almost the same silicon as the 5090: https://www.amazon.com/Blackwell-Professional-Workstation-Simulation-Engineering/dp/B0F7Y644FQ

When the bill of materials is really only like $100-$200 more, at most. Nvidia can get away with this because everyone is clamoring for their top end cards

AMD, meanwhile, is kind of a laughing stock in the prosumer GPU space. No one's buying them for CAD. No one's buying them for compute, for sure... And yet they do the same thing as Nvidia: https://www.amazon.com/AMD-Professional-Workstation-Rendering-DisplaPortTM/dp/B0C5DK4R3G/

In other words, with a phone call to their OEMs like Asus and such, Lisa Su could lift the VRAM restrictions from their cards and say 'you're allowed to sell as much VRAM on a 7900 or 9000 series as you can make fit." They could pull the rug out from under Nvidia and charge a $100-$200 markup instead of a $3000-$7000 one.

...Yet they don't.

It makes no sense. They're maintaining an anticompetitive VRAM 'cartel' with Nvidia instead of trying to compete.

Intel has more of an excuse here, as they literally don't manufacture a GPU that can take more than 24GB VRAM, but AMD literally has none I can think of.

Kind of my point, these were high end and still usable by 95% of people. Everyone is chasing 1% gains for twice the price. I have an new RTX via work equipment for rendering, I play games on the side but that RTX dosnt really make the gameplay that much better. It looks great with the shine on metal, or water reflections, but when totally immersed in game play that stuff is wasted