00:00As AI infrastructure grows, some of the biggest questions around the trade are only getting harder.
00:04How much AI will companies consume? Do cheaper open models change the economics?
00:09And even with high demand, is there enough labor and power to even run data centers?
00:14Michelle Weaver, Morgan Stanley, U.S. dramatic research strategist and executive director,
00:18has the job of passing these questions on a daily basis.
00:21We're going to talk about the AI economy.
00:24And for me, at the end of the day, it is just simply trying to measure demand as it is
00:29now,
00:29against supply in the future states of both.
00:32Where are you netting out at the moment?
00:35Currently, we are still very much undersupplied.
00:38We are still seeing compute be a very constrained resource.
00:43And I think what you really have to look to is the AI adopters.
00:47What is the everyday business consuming AI seeing?
00:50Are they starting to see ROI?
00:51Are they starting to see benefits from AI adoption?
00:54And the answer there is yes.
00:56We're still very early on the S-curve for AI adoption.
01:00But our tech team's CIO survey shows that by the end of this year, the majority of companies
01:04will have at least one AI project live and in the field.
01:08And we've seen that when we look at S&P 500 companies.
01:1125% of those companies are now able to quantify the benefits that they're getting from AI adoption.
01:17That's up from 14% a year ago.
01:19So companies are incrementally seeing positive signs around what they're getting out of AI adoption
01:24and voicing that to investors.
01:26So you've gone straight to this through the lens of basically corporate America,
01:29its attitude towards AI and the action it is or isn't taken.
01:33And that's measured on token spend, the debate around token maxing.
01:38Yes.
01:38What is the current in aggregate attitude towards token spend?
01:43So we're certainly seeing a decline in the culture of token maxing or trying to consume
01:49as much compute as possible.
01:51So I think you have to separate what's going on with the engineer and with the power users
01:55versus the more typical user.
01:57So power users, yes, had been consuming a ton of tokens.
02:01And enterprises are now thinking about how can we rationalize that?
02:04How can we make sure our employees are using an appropriate budget of tokens?
02:08But the typical user, the median user, spends less than $11 per month on tokens.
02:14So for the vast majority of people, token spend is still very much under control.
02:18The big story this week has been NVIDIA and the $500 billion it is channeling through those
02:24six Wall Street firms, right?
02:26The idea is it greases the wheels.
02:28It takes away the excuse of not being able to finance data center.
02:31But it seems like capital just isn't enough, right?
02:34There are labor bottlenecks.
02:36There are other supply chain bottlenecks.
02:38How severe are they?
02:40Yeah, still significant bottlenecks to getting data centers built.
02:45And that does make the value of compute even more strong.
02:48But I would say the number one bottleneck is labor.
02:51There is a huge shortage of electricians and specialized labor to actually put up these
02:56data centers.
02:57And then the second biggest bottleneck is power.
02:59We estimate around a 40 gigawatt shortfall in power needed through 2028.
03:04When we start thinking of some of these innovative time-to-power solutions like Bitcoin site
03:08conversion, innovative fuel cells and turbine projects, you still have around a 10% to 20%
03:14shortfall in that power needed.
03:16So still a huge bottleneck there as well.
03:19We have the midterms fast approaching and the NIMBY movement, anti-data center movement.
03:25Seems very real.
03:26Have you modeled for that in your forecast for the market and growth in industry?
03:32Yeah, so I would say they fall roughly into three different buckets of concerns.
03:36And there are ways for these concerns to be addressed.
03:39There's just a lot of nuance here.
03:41So first concern, power bills.
03:43Communities are very concerned that if a big data center comes to town, starts consuming a
03:47bunch of power, ultimately that could end up on consumer power bills.
03:51And that really speaks to how worried consumers are about affordability.
03:54But being off the grid helps to solve for this.
03:56You're not touching the grid.
03:57You can't possibly impact power bills there.
03:59Second bucket, environmental concerns, water and air.
04:03This really depends on the type of cooling you're used.
04:06Closed loop cooling, much less water intensive than evaporative cooling.
04:10And then the last bucket is local quality of life concerns.
04:13A lot of concern around land use, traffic, noise, dust.
04:18Here, I think you'll start to see more brownfield sites where you're saying, OK, we're coming
04:22in.
04:22We're cleaning up this site.
04:24We're able to provide jobs to the community and provide a productive use to this piece of
04:28land.
04:28So a lot of nuance here.
04:29But I think this is going to be a very big topic as we head into midterms.
04:33Michelle, we started this conversation with you saying we're still supply constrained,
04:38demand vastly outpacing supply.
04:40So this is strange to ask, but like, how real is the risk that we get to a place of
04:45oversupply?
04:46Too much capacity gets built.
04:49I think we're still very far off of that.
04:51I don't see any signs that we're seeing supply overbuilt.
04:55I think those bottlenecks we were just talking about, the power bottlenecks, the political
04:58bottlenecks, the labor bottlenecks, that's really going to keep a check on supply for the
05:03next few years.
05:04Supply, I wouldn't start to worry about that within the next few years, just given how short
05:08we are on power through the next few years.
05:11So this is interesting.
05:12Later in the program, we're going to talk to silicon data, the benchmarking for computes.
05:17Everything I'm hearing from you is that all the factors are in place to keep compute prices
05:22high.
05:23We talked about token maxing or token spend.
05:26There's a difference between quality of token and bringing the cost of tokens down.
05:30None of that seems to be the direction of travel.
05:34So, correct.
05:35I think the big debate right now is the open model versus closed model.
05:40Who is ultimately going to process those tokens?
05:43And I think there's a different picture for when we look at the U.S. versus global.
05:47Globally, for the majority of firms, I think they're just looking for something that is good
05:51enough.
05:51They just want to be able to process the requests that they need.
05:54In the U.S., there is still a lot of concern around security.
05:58With using non-U.S. models, there is a lot of concern around just business continuity
06:03risk.
06:03If the government comes in and regulates and you've spent all this money fine-tuning an
06:07open weights model, how do you deal with that?
06:09So I think in the U.S., you likely end up in a hybrid state of the world where you
06:13have
06:13both closed and open models being used in parallel.
06:16Currently, around 63% of companies use both.
06:19So I'd expect that to continue.
06:21You know, Michelle, it's your job at Morgan Stanley, as much as is possible, to have a
06:26crystal ball to tell us what happens next.
06:28You know, what are the outcomes you're expecting here, particularly sort of on the open closed
06:33debate that we are having?
06:36That hybrid end state of the world, I think, is most likely.
06:40I think when we think about companies are already adopting AI.
06:43Yes, they're early on in that adoption curve.
06:46But adoption is sticky.
06:47A lot of this is changing employee behavior.
06:50This is changing human behavior.
06:52And so given that a lot of companies are already using closed models, employees are already
06:56building skills on these models.
06:58They're already changing their workflows.
06:59I do think there will be a pretty strong degree of stickiness there.
07:04But maybe you have smaller use cases where you really want to fine-tune a model.
07:08You want something really specialized.
07:09That would argue for an open weights model.
07:11If you have new use cases that you just want to serve it as cheaply as possible, that would
07:16argue for an open weights model.
07:17But there's a lot that's already been built on top of those closed models.
07:20And I think it will be very challenging to come and rip that out.
07:23Michelle, real quick, regulation a wild card here in the U.S.?
07:28Yes, very much, very much a wild card in that open versus closed debate.
07:33I think that is something companies are thinking about.
07:36If they were considering implementing an open weights model from overseas, what if those
07:42companies are no longer allowed to operate in the U.S.?
07:46And so I think there's a lot of hesitation to go fully into some of those open weights models,
07:50given that geopolitical concern.
07:53The other big issue, too, is like we're talking about the community pushback, the state-level
07:58moratoriums.
07:59New York State was the first to introduce a state-level moratorium that could be spreading
08:03to other states.
08:03So that's the other political wild card to watch here.
08:05The only thing I think everybody has to cover is aπ grandfather.
08:06that if I did either 멈ate to attractions for some of the real cases, it'd be a transfer
Comments