I watched Jensen Huang's GTC keynote Tuesday afternoon with a bag of chips in one hand and my phone propped against a coffee mug, which is about as much ceremony as a two-hour hardware announcement deserves. Nvidia held its big developer conference this week in San Jose, and somewhere between the tenth mention of "physical AI" and a slide with more acronyms than I could keep up with, a small orange robot dog walked out on stage and just hung out with him.
That's the part that stuck with me, not the chips.
The droid is a BDX unit, built with Disney's Imagineers, running on a physics engine called Newton that Nvidia developed with Disney Research and Google DeepMind. It beeped, it looked up at Jensen, it did the head-tilt thing where you can tell an animator spent real hours getting the timing right. Everyone in the room audibly went "aww," on a livestream, at a chip conference. I've watched a lot of keynotes since this blog started back in 2011 and I don't think I've heard an actual "aww" at one before.
The real news, if you strip out the droid and the leather jacket, is that Nvidia said its next couple years of GPU roadmap out loud instead of letting it leak the usual way. Blackwell Ultra is shipping this year, then a new architecture named Vera Rubin (after the astronomer, which is a nice touch) lands in 2026, and something called Rubin Ultra comes after that. Naming your chips after a scientist who spent her career proving dark matter exists by watching galaxies not behave the way they should is either very on-brand or a little too cute. I can't decide which.
What actually made me sit up, though, was the DGX Spark. It's a personal AI computer, small enough to sit on a desk, built around a Grace Blackwell superchip, and Nvidia's pricing it around $3,000 for developers who want to run and fine-tune fairly large models locally instead of renting cloud GPU time by the hour. There's a bigger sibling too, the DGX Station, for people who need real workstation power and presumably a real budget to match.
Three grand isn't nothing. It's also, weirdly, not that much for what it claims to do, and that's what's been rattling around my head since Tuesday. I remember buying my first "real" desktop back in 2004 for less than that and thinking I'd basically bought a spaceship. Now the equivalent spaceship money buys a box that can run locally what would've sounded like science fiction five years ago. I don't need one. I don't have a project that requires one. I still spent twenty minutes on the product page like I was about to click buy, which I'll admit is a personality flaw at this point, not a hardware decision.
I'm a little tired of "AI supercomputer for your desk" as a category, honestly, mostly because I've seen three or four versions of this pitch over the last couple years and none of them have shown up used on eBay yet at a price a hobbyist would actually pay. Ask me again in eighteen months, I guess.
The keynote ran long, like they always do, and there was a stretch in the middle about robotics foundation models and something called GR00T that I mostly tuned out for because I was making a second coffee. That's on me, not on Nvidia. Two hours is a lot to ask of anyone's attention span on a Tuesday, robot dog or not.
If you want the full rundown of every announcement, there are already forty other sites that wrote that post, probably with better screenshots than I managed to grab off a laggy stream. I just wanted to write down that the best five seconds of a multi-billion-dollar hardware roadmap was a fake dog looking up at a guy in a leather jacket, because in a year I won't remember the die shrink numbers, but I'll probably still remember the head-tilt.