"Circular spaghetti": What the tangled AI supply chain means for regulation

Last week, tech executives signed a voluntary agreement at the White House aimed at ‘self-policing’ the AI industry. But as calls for AI development to slow down get louder, there seems to be disagreement among industry leaders about what that ought to look like.

David G. Widder, an assistant professor at the School of Information at the University of Texas at Austin, spoke with “Marketplace” host Kai Ryssdal about how mixed incentives in the AI supply chain could complicate efforts to self-regulate.

The following is an edited transcript of their conversation.

Kai Ryssdal: Would you do me a favor — just so we all have the same ground truth here — and give us like a 30-second primer, maybe 45 if you need to, on the artificial intelligence supply chain right now?

David G. Widder: Yeah, everyone's talking about AI, and the AI supply chain is basically a way of thinking past that hype into the specific people building different parts of the of AI. We have Nvidia designing chips that they pass to TSMC to build, and we have Google, Microsoft, and Amazon buying those chips, and then we have model labs building models using the cloud compute — and I can go on.

Ryssdal: Well, let's just I'll short-circuit the going on by just pointing out that it is somewhat circular, this supply chain — everybody's sort of in bed with everybody else.

Widder: Yeah, it is. I mean, we have Nvidia investing in their customers, and then they buy more Nvidia chips. We have Google, Microsoft, and Amazon running startup incubator programs, and then they have to use their cloud compute. There's OpenAI underwriting finance for different data center projects, and that comes together in very convoluted deals. It's a mess. It's a circular spaghetti.

Ryssdal: Now yeah, frame that whole circular spaghetti thing for me in the context of what is now in the news about AI. This drive to possibly slow down, regulate in some way. How does the supply chain handle that?

Widder: It doesn't — and that's kind of the problem. Remember way back when Nike got in trouble for using child labor in the supply chain, and their answer was something like, "Well, we told our contractor who makes the shoes not to use child labor — that's not my problem?” And people are like, "Nah, I don't, I don't buy that.” And so we have a similar dynamic going on where there are known harms about AI, as well as these like, "Uh-oh, it might end the world scenarios,” and we have this collective action problem where it has been framed such that everyone would have to act together in order to avert these harms.

Ryssdal: Setting aside for a second the fact that Congress doesn't know what day it is, let alone can't regulate something as fast-moving and as complicated as artificial intelligence — it is worth pointing out here that there are hundreds of billions of dollars in sunk costs already. There are trillions more on the way, and so this collective action problem is complicated by the very realities of that circular supply chain that we talked about up at the top of this interview.

Widder: Yeah, let's go back a little bit. I think there's an important way of realizing how different supply chain actors are allies when it comes to certain concerns and competitors when it comes to others. So, Nvidia, Google, Microsoft, Amazon, TSMC — they all have an interest in AI being more broadly used. Nvidia would like to remain the dominant chip designer, but Google, Microsoft, and Amazon make up [about] two thirds of the cloud compute market, and therefore a large proportion of Nvidia's market share. Meanwhile, Google doesn't like being dependent on on Nvidia for its chips, so they're building Tensor Processing Units (TPUs), and other ways of trying to make themselves less dependent on Nvidia chips.

Ryssdal: Right. So, fundamentally, a lot of these players in the supply chain, it's ‘I love you, I hate you, I love you, I hate you’ — right?

Widder: Yeah, exactly.

Ryssdal: So at the end of the day, then, as a guy who is immersed in this — certainly more than the average layperson — are you optimistic or pessimistic about our ability to get a hold of what AI is doing?

Widder: I try and start by looking at what the American public thinks. And if you look at the polling, the average American — unless they make more than $200,000 a year or are personally helping build AI — is not optimistic that AI will benefit them. I think that is a good set of starting points for increasing regulation. I want to make sure that regulation isn't just banning data centers, but goes beyond that to actually engaging with the circularity we're talking about, the systemic power of some of these companies. I think there are more finely-tuned ways of doing it, I guess.

Ryssdal: You sound — if you don't mind my saying it — a tad bit dispirited.

Widder: There's a joke among some professors I've talked to lately that some people study things they love, and some people study things they hate. Both are super motivating. And I — you can't see my hands right now — but the way I explain my position to folks is, ‘AI, thumbs up. AI good. Concentrated corporate power — bad.’ But right now, AI depends on concentrated corporate power. I hope to help imagine a way out of that.