The Unexpected Investment Case for AI Safety

Tighter AI safety requirements could reshape the pace of AI investment. Ariana Salvatore and Michael Zezas dig into why the spending may shift toward more compute, not less.

Read more insights from Morgan Stanley.

----- Transcript -----

Ariana Salvatore: Welcome to Thoughts on the Market. I'm Ariana Salvatore, Head of U.S. Public Policy Research at Morgan Stanley.

Michael Zezas: And I'm Michael Zezas, Deputy Global Head of Research at Morgan Stanley.

Ariana Salvatore: Today, we'll be talking about AI safety and regulation.

It's Wednesday, September 23rd, at 10am in New York.

We put out a note last week on AI frontier capability gain and the associated safety risks.

Those have been in focus in recent weeks, and as a result, we've gotten a number of questions about the path forward for government regulation.

So today, Mike and I are going to get into some of the newest developments, where we think things are headed, and how the midterms could shape that path.

Michael Zezas: Yeah, and this is pretty important because the concern is that if AI safety scrutiny increases, it's going to slow everything down. You might have less CapEx, fewer model releases, and there's all sorts of downstream effects for the pace of U.S. growth and investment strategy in equities and throughout the AI investment theme.

But Ariana, you and the team landed in a bit of a different place and are arguing that a bigger focus on AI safety could end up being a tailwind to compute spend rather than a brake on it. Can you break that down for us?

Ariana Salvatore: Sure. So, the way we see this playing out, is there are five potential states of the world. Some include industry self-policing; some include the prospects for heavier government intervention. Across all of them, as you mentioned, we actually think this is a pretty big tailwind to compute spend and CapEx more broadly.

That's because as the labs integrate greater safety monitoring infrastructure, we think that spend is only going to accelerate, especially as LLM capabilities increases at a nonlinear rate. Similarly, on the regulation front, we think there are a few things that prevent something like a large comprehensive AI regulation bill from coming to fruition.

We think there's really three, kind of, key obstacles to something like that happening.

The first is the politics. So, the president himself has said he's against some sort of large-scale regulation. The second is the procedure. So mechanically speaking, there would need to be a legislative vehicle for this sort of thing to ride on. That's hard to see emerging in the very near term. And the third is precedent.

So, historical precedent here tells you that usually regulation is catalyzed by some sort of high salience event. That's why our framework for government reaction here hinges on two components: incident salience, as I just mentioned, and instrument availability. Instrument availability basically reflects the extent to which the government already has a tool that it can pull in this direction.

So, that's how we think about it going forward. That doesn't mean all policy action is off the table, but that supports our expectation for higher CapEx, higher compute spend over the coming years.

Michael Zezas: Right. So, the idea is that the spending continues and the things that would otherwise limit that spending, you don't see as real plausible policy options at the moment. And can you break this down a little bit more? Because I know there's a lot of different proposals floating around Washington, D.C. from policymakers right now.

What are you paying attention to?

Ariana Salvatore: We don't expect an overarching AI regulatory authority in the near term. Now, importantly, we also don't expect sweeping open weight model regulation. The reason for that is threefold. First of all, we think the U.S. is keen on maintaining this managed stability relationship with China.

We've written about the expectations around the U.S.-China summit. That's kind of a delicate balance that we think is likely to persist. So, overly restricting open weights models might throw a little bit of a wrench into that equilibrium that we see. So that's the first reason.

The second reason is diffusion. We think the U.S. administration wants to see the proliferation of open weights models. We know that companies are using some sort of hybrid of open and closed weight. So, to the extent that, you know, banning these models would slow adoption, we don't think that's in the interest of the administration.

And the third reason is purely mechanical. It's really hard to enforce these sorts of restrictions. Once a model weight is published online, it can be really hard to clamp down exactly who and where it's going to.

Obviously, companies can download them, customize them, et cetera. So, the enforcement picture here is also really challenging. That being said, we do think that the executive can continue to lean in and, sort of, make some incremental adjustments or changes on the regulatory front. But we think it's likely less severe than some of the proposals you're seeing in Congress right now. Things like the Kill Switch Act, for example, which basically mandate that companies can maintain an ability to shut down models at a moment's notice, right? If a certain threshold is crossed.

So, that's something that we see as less likely to come to fruition. But again, setting safety standards, guardrails, all of that from the administration we think is possible in the near term.

Michael Zezas: What about some of the pushback that would at least appear to be rising at the state and local level around construction of data centers?

Is that something that you think might materially slow the industrial build-out and the CapEx levels around AI?

Ariana Salvatore: So far, what we've seen is that AI safety risks are not the top of the priority list when it comes to data center pushback, right? So, things like environmental concerns, affordability – those tend to be the main vectors of the opposition.

That being said, we've gotten the question, right, to your point, of does this, sort of, risk focus mean that the data center backlash is likely to grow? We think that it could, but at the same time, we think this is a highly idiosyncratic issue, meaning that this is something to pay attention to on a very granular level.

Certain states and localities will be the ones to really administer these restrictions, and we think in the aggregate, hyperscalers are going to be able to continue to mitigate. We've already seen these mitigation measures employed. We're still constructive on AI CapEx this year and next, because overall, we see the build-out really becoming more of a conditional build-out.

So, that means contingent upon some of these concessions, maybe it's more expensive in certain areas. But overall, we don't think that the concerns around safety are going to derail that story.

Michael Zezas: So, then when it comes to data centers, the conditions that might be being put on their construction at the state and local level, for the most part – those building out the data centers have been willing to make those concessions, so it hasn't slowed that much. Is that fair?

Ariana Salvatore: That's right, and it really depends on where the pushback is coming from, right? So, in some cases, you're seeing communities push back on things like water usage, right? And we're seeing the hyperscalers come out and respond and say explicitly, you know, how much water they're using in some of these operations. Google is proposing a regulatory framework, so that's something that they're mitigating through that lens.

In another example, you've got local communities pushing back on just, sort of, disruptions to quality of life, and you're seeing companies like Meta announce a fund to engage more locally there.

So, it really is different. There's no one-size-fits-all solution here. But yes, I agree with you that overall, we don't think this is going to meaningfully constrain the build-out.

Michael Zezas: Got it. So, it seems like the idea here is that the secular trend around AI development is going to continue in your view. Is there any way that you think the midterm elections or the outcome around that might change your thinking?

Ariana Salvatore: So, I think the midterms will be important for sentiment, but when it comes to the actual policy path, we don't think they're the main driver, and there's two key reasons for that.

The first is obviously the president is not changing until 2029. So, the fact that President Trump still has to be involved in any capacity – if we were to see a bill emerge from Congress to us gives a little bit of clarity on what that bill could actually look like. And so ultimately, whatever comes to fruition will have to be a product of collaboration between Democrats, Republicans in Congress, and the president. So, that's a pretty much a constant.

The second reason I would say is because, as I kind of alluded to earlier, you tend to see government response when there's a high salience event. And in that case, it doesn't really matter what the government configuration is if it's reactionary.

When you think back to things like the pandemic, we saw the CARES Act. In 2008-2009, you saw the ARRA. Those are all efforts that were produced in a divided government. And so, in that vein, we basically think that you need to see some sort of event catalyze a response.

The key driver is not going to be government configuration. It's going to be the salience of that event specifically.

Michael Zezas: Okay, got it. So, the guidance to investors on the back of all of this is what?

Ariana Salvatore: So, the thematic recommendations from our team are intact, right? So, what we were talking about is basically we see these all converging towards a tailwind to CapEx and a tailwind to compute supply.

So, in that vein, we still think that you should own inference compute bottlenecks because of that excess demand relative to supply. We think that's going to persist regardless of most of the policy scenarios.

We also think own leaders in cybersecurity, as we mentioned. This should drive increased spend, especially as open weight models become much more capable. And then in the third piece, we think you should own AI adopters. AI models are already pretty capable to drive significant productivity. We think that that's just going to continue to unlock.

The last thing I would mention, we didn't really get into it in this podcast, but in this conversation we typically also talk about AI sovereignty and the U.S.-China restrictions here.

So, in that context, we would avoid negative exposure to rising U.S.-China technology transfer restrictions. We think that's an increasing probability because we do see the governments taking more of an active role, which we think drives bifurcation of the global market.

Michael Zezas: Well, Ariana, thanks for breaking it down. Appreciate talking to you today.

Ariana Salvatore: Always great speaking with you, Mike. And thank you for listening. If you enjoy the show, please leave us a review wherever you listen and share Thoughts on the Market with a friend or colleague today.