The Shovel Salesman Has a Point (Even If He's Still Selling Shovels)

There’s a clip doing the rounds where a senior figure at Nvidia makes the case that AGI is not coming, that OpenAI and Anthropic are building walled gardens the same way AOL and Prodigy built closed internets in the nineties, and that the future belongs to every business running its own customised open source model. He says it with the calm confidence of someone who has thought about this for a long time.

The comments, predictably, split between “shovel salesman says there’s gold in the mountain” and “well, he’s not wrong though.”

Both things are true. That’s the annoying part.

Nvidia makes more money if ten thousand businesses each buy a rack of GPUs than if three hyperscalers buy everything and everyone else just rents tokens. Of course he wants to grow the addressable market. The fact that his commercial interest aligns with a particular vision of the future doesn’t automatically make the vision wrong. Oil companies have a commercial interest in you driving to work, and yet: roads exist, cars work, commuting is real.

The AOL comparison is the part worth sitting with. There was a moment in the mid-nineties when AOL and CompuServe genuinely seemed like they might be the internet for most people. Closed, curated, clean. Then the open web ate them. Not because open was morally better, but because it was more composable. You could build weird things on top of it. Weird things turned out to be valuable.

I don’t know if that exact story repeats with AI. The open web analogy only goes so far; models require serious compute to run, and “every business has its own model” sounds great until you price it out. Someone in the comments noted that 96GB Blackwell cards are already down to twelve thousand dollars, which was said as a sign of progress. Twelve thousand dollars. I bought my car for less.

Still. I’ve watched open source quietly eat industries that seemed impregnable. R replaced a lot of expensive stats software. QGIS took a real bite out of ArcGIS. Linux runs most of the infrastructure I’ve worked on for fifteen years, while the year of Linux on the desktop became the longest-running joke in tech. The joke has a point, but so does the underlying observation: open source tends to win in the infrastructure layer, eventually, in ways that are hard to predict from the outside.

The AGI claim is the other thread worth pulling. Someone in the comments made the 1909 aviation comparison: dismissing AGI now is like saying in 1909 that commercial flight will never work. It’s a reasonable rebuttal. It’s also the kind of rebuttal that sounds good and proves nothing. You can make the same argument about cold fusion. Timing and mechanism matter, and “it seemed impossible before and then wasn’t” is not a theory of how or when.

I genuinely don’t know whether AGI is coming in ten years or fifty or never. I’m skeptical of people who are certain in either direction. What I notice is that the people most confident it’s imminent are often the ones who need you to believe that to justify their valuation, and the people most confident it’s impossible are often the ones selling you the alternative.

The version of the future I find most plausible, for whatever that’s worth, is messier than either camp describes. Some businesses will run local models. Most won’t, because it’s hard and they have other problems. The hyperscalers will remain enormous. Open weights models will get genuinely good and will matter more than the closed-model boosters want to admit. Nvidia will continue to make extraordinary amounts of money regardless of which scenario wins, which is perhaps the most reliable prediction of all.

The bit that actually interests me is the distributed training problem. A few people in the thread mentioned it. Training at scale currently requires centralised infrastructure almost by definition. If that gets solved, the whole landscape shifts. It’s technically hard in ways I only partially understand, but the people working on it seem serious.

That’s the thing I’d watch, more than whatever a senior Nvidia executive says on a podcast.