Nvidia's GTC keynote happened today in San Jose, and I watched about ninety minutes of it before I had to go make dinner. Jensen Huang did his usual thing: no slide deck in the boring sense, just walking around the SAP Center stage in that black leather jacket he's worn to basically every one of these things since I started paying attention, holding up chips like they're newborns. This time it was Blackwell Ultra (the GB300, shipping in the back half of this year, apparently) and a first look at the next architecture after that, which they're calling Vera Rubin. Named after the astronomer who found the evidence for dark matter. Somebody in Nvidia's naming department has a sense of humor, or at least a physics degree.
I'll be honest, most of the keynote was aimed at people running data centers, not people like me who just want to play Cyberpunk at 4K without my apartment turning into a sauna. Huang spent a big chunk of time on "AI factories" and something called Jetson Thor for robotics, and there was a whole section on physical AI that I mostly tuned out for because I was refilling my coffee.
But here's the thing that actually got me typing tonight instead of just doomscrolling: this is the same company that launched the RTX 5090 back in January at a $1,999 MSRP and then, as far as I can tell, made approximately eleven of them. I've had a browser tab open on Best Buy's stock page more mornings than I want to admit, and I've watched that card sell out in under a minute on the two occasions it actually showed up. Street price on the secondary market has been running $2,800 to $3,500, sometimes worse. My brother-in-law paid $3,200 for one off a reseller last month and I did not have the heart to tell him what I actually think about that.
And it's not just supply. There's also that whole missing-ROPs mess with some early 5090 and 5070 Ti units, where a manufacturing defect meant a chunk of cards shipped with fewer render output units than they were supposed to have, quietly underperforming without anyone noticing until reviewers started digging and comparing benchmark numbers. Nvidia's been running a card-swap program for it since. Not exactly a confidence builder when the company's also up on stage the same month promising you a chip that helps run entire AI data centers.
I don't think Huang's lying about any of it, to be clear. Blackwell Ultra is real, the roadmap is real, and Nvidia's stock has mostly recovered from that scare back in late January when the DeepSeek news knocked something like $600 billion off their market cap in a single day. People panicked that a Chinese lab had trained a competitive model for a fraction of the usual compute, decided that meant Nvidia was doomed, and then mostly calmed back down within a couple weeks once it became clear everyone was still buying GPUs by the truckload anyway. Enterprise demand for this stuff isn't slowing down even a little, from what I can tell.
It's just a weird split screen to live through as an actual consumer. On one channel: trillion-dollar company, chip named after a Nobel-adjacent astronomer, robots, "AI factories." On the other channel: me, refreshing a cart page for a graphics card that costs more than my first car did, that may or may not have all its render output units intact, competing against bots and resellers for a product the company itself set the price on in the first place. I don't need Nvidia to feel bad for me specifically, that's not really the point. I do think it's fair to say the gap between the keynote version of this industry and the actually-buying-a-GPU-in-2025 version of it has gotten a little silly.
Also, side note, does anyone else find it a little much that these keynotes are two-plus hours long now with no intermission? I love the guy's enthusiasm, genuinely, but I had to pause it twice just to stretch.
Anyway. Dinner's cold. If anyone actually manages to snag a 5090 at MSRP, tell me your secret, because at this point I'm starting to believe it's a myth, like a UFO sighting or a landlord who fixes things promptly.