Jensen Is Out. Here's What Changes for Nvidia and AI

Nate B. Jones (0:00)

Jensen Huang has resigned from Nvidia. He's out as CEO, president and a member of the board. Collette Kress is taking over as interim CEO while keeping her job as CFO, and Nvidia has appointed Harvey Jones as its first ever board chair. I think the biggest consequence here is what happens to the next deal for Nvidia. Nvidia must defend its position without the person who tied together its technology, its customer relationships, its brand and its political influence. And has to do that just as customers are getting more and more credible alternatives to its chips. That affects who makes money from AI, but also how much computing gets built and what you pay to use it. Now the board says it expects operations, customer commitments and the product roadmap to continue without interruption. Of course, they're going to say that. Thousands of people still have the expertise to do the work. That's certainly true. And customers desperately need the chips. Also true. But I want you to imagine for a second, you're building an AI cloud and about to commit to years of Nvidia equipment and deals. You need delivery dates, technical support, and someone who can resolve a problem when your deployment falls behind. Which customer gets that scarce supply of chips? Which new system gets priority? What terms make a giant deal doable? Jensen spent decades building the authority and relationships to make those calls, and he was unquestioned in his ability to do so correctly.

The company now has to replace that coordination and prioritization function while searching for a permanent leader. And that gives a bunch of competing suppliers openings. And one of the biggest obstacles to taking that supplier seriously is getting smaller. That is the cost of moving your software. Developers have spent years building around CUDA. Now CUDA stands for Compute Unified Device Architecture. Say that five times fast. It is the way that programmers are able to tell Nvidia chips what to do with AI. That's the long and the short of it. It is considered to be one of the best motes in the industry. Moving that code to another company's processors takes a lot of engineering work. Even attractive hardware can lose the sale if switching takes too long or costs too much. Anthropic's IPO Roadshow here gives us a useful example of how this story is starting to change. The company says its Mythos model got Claude running on AMD chips, a different chipset in about three days with virtually no human intervention. Now, the qualification here is that this was prompt processing for one model on AMD's particular MI355X chips at a useful level performance. That is not the same thing. Prompt processing is just a part of what the model reads when you send it. It's not the same thing as getting the full model running at scale. It's certainly not a demonstration that every part of Cloud could move to another chipset in a weekend. That would be the incorrect thing to infer here. That being said, it shows us that that moat may be starting to erode. Anthropic says Cloud wrote or adapted nearly all the low-level software routines involved, which are typically very tedious. It compared that with earlier efforts that took dozens of engineers months of work.

But also, Anthropic has been investing for a long time in multiple chip platforms, and so they are also already thinking about and already have expertise in this direction. The company is unusually well prepared to attempt this. So the thing that I want to take away here is not everybody can write, you know, a way around CUDA in a weekend. That would be wrong. What matters here is the possibility that AI can do more of the engineering that has made switching chipsets expensive. So I want to think about it not so much as the CUDA moat is gone. The CUDA moat is absolutely still there, will be there next year. But I think that there is a little bit of soft risk that's emerging in that space. And it's something that we may see more of as people become more fluent at using these frontier models for these complex software engineering tasks. Go back to that cloud. Imagine that you're investing in a new cloud that you want to build. It's a private cloud. You don't need AMD to be better at every task to make that cloud work. If some of your work can run economically on AMD chips, you might divide your next order between suppliers. Nvidia might keep you as a customer, but they might have to negotiate harder next year to keep your business. That's how demand for AI can keep growing, while Nvidia maybe captures less of the share of the profit than it has for the first few years of the AI revolution.