IBM前CEO解析AI资本开支回报周期与历史规律

Bloomberg Audio Studios podcasts, radio news. This is Wall Street Week. I'm David Weston bringing you stories of capitalism. US Canada trade talks come to the brink. Former Canadian foreign minister Christiana Freeland tells us what is at stake. And there's been a lot of speculation about what AI will mean for jobs. Apollo's Torston Slock brings us the first study of its kind on what's actually happened since Chat GPT came on the scene.

Plus, it supports $1.4 trillion of US GDP and 16 million jobs. But the Colorado River is running out of water. We look at government plans to ration what's left. But we start with the hundreds of billions of dollars being raised to invest in artificial intelligence, including $500 billion in one deal done by Nvidia. Sam Palmisano served as CEO of IBM, where he was responsible for capital investment in new technologies and the need to make sure they made money for the company.

So Sam, there is a lot of money being invested in AI, whether it's chips or whether it's data centers, including the Nvidia announced deal of $500 billion. Give us a perspective of a CEO of a big tech company. How do you decide how much to invest in a new technology how we would have thought about as you know I had one of these big tech companies. We would have thought about what is where do we see the future state? And you go back to Watson with the 360 because that was going to be from unit record to computation caught a computer right in my particular case it was AI Watson Jeopardy you know right so how do you jump ahead of the next generation because the in investment cycles in the R&D are going to be several several years so you have to start no different than quantum is another example of that that's true for all tech companies I'd say it's true for many companies that are relying upon any form of technology could be energy and those sorts of things as well So we would always look out in time uh and that could be from the R&D cycle it' be like a 7 to 10 years on the models 3 to five for short-term we'd call that operational investment but for the longer term things and then you and then basically you start out with goals and milestones and objectives because you don't have numbers and then as you get closer you actually build your return equations as your business model dictates.

So if your IRRa is 15% or 14% that would be the hurdle rate the guys would have to come over before you launched.

7 to 10 years is a long time in any business.

Yes.

But right now it seems like 7 to 10 months in AI is a long time product cycle time. Right.

Exactly. How do you how do you project out uh returns with AI that's changing so fast?

Well, the software is changing fast and the models, right? the learning models, the frontier models, that's what's changing really fast. Data centers aren't changing fast. I mean, we'll the estimates are three to five years before this stuff comes online. That's not unrealistic. Semiconductors to get it online probably seven. Nuclear at least 10. So that's the energy requirement for a lot of these things. So these things are long cycle times uh before you actually get the capacity in place.

Now I mean the challenge with this in the interim obviously is is people the large hyperscalers the big guys out there will probably uh have enough uh requirement that they'll they'll in many ways control the market I mean because they'll take the make they'll make the big bets in the short term they'll get the data centers they'll get the energy other guys that are trying to compete or or even to participate in this market are going to I think be squeezed out

we have all these announcements is this money actually going going to be invested. Is it going to happen?

Well, right now they're memorandums of understanding. Uh people claim that there's there's commitments and there's teeth in these agreements. I have not read them. I have read asked people who theoretically should know and they tend not to comment on the actual specifics. So I'm just going to leave it at that. I I can't imagine if this thing adjusts say economically. Suppose for example just adoption slows. It doesn't have to be an economic downturn.

Just a simple thing your adoption curve you assumed would be 7 to 8 years let's make it 10 to 12 that slows you're not going to build out the capacity as quickly that alone you know could cause I think a change in their plans and then if they're firm commitments it's one thing which I doubt as they become more variable commitments over time someone again is going to have to deal with the shortages I mean I either the bond holders I mean someone committed to put a shovel in the ground to build a data center.

Someone agreed to add energy capacity to support this aggressive case that we see today. At some point, if things do slow, that would have to adjust. The way things are are today, we tend to look at a company based on return on investing capital.

Yes.

And we compare companies depending on their return. Uh for example, I took a look and IBM is running about 10% return invested capital. Microsoft, you know, Google Alphabet is like 25%.

Correct.

Um, take the 10% number on 500 billion dollars. That's a lot of incremental income that you have to generate every year. Every year, and put that on a base of I say IBM today, 60 or 70 billion. I was 100 billion when I was there, 105, whatever. Put on that base of some level of our margins were a little higher than they are today. But put on the base that you already have a stream of several billion. you got to add an additional several billion in an annual basis.

And I used to say that at IBM for us to grow at seven or eight% just on the top line, we'd have to be create a Fortune50 company every 14 months. Now, in enterprise computing, that's tougher. Commercial consumer in tech is probably an easier way to do that. You get a hit hot phone or whatever it happens to be. But if you're selling to banks or telos or governments, the odds of them taking up their expenditures at that rate, say you could create a Fortune50 company, it's pretty tough to do.

If you look back in history, some of which you lived at IBM, but also going back to the industrial revolution, railways, something like that. Are there patterns with these sort of technological revolutions that we see?

Yeah, there are actually. There's a there's a scholar at Cambridge, her name is Carla Perez, and she's done a lot of work uh going back to 1771 for the industrial revolution, then the steam engine, and then you know, oil and gas, and then micro electronics and compute, and now she would probably run her sixth generation, but the pattern's the same uh over time. It starts with she calls the installation cycle uh we would call maybe the industrial I mean infrastructure buildout today.

That's the base that that gets put in place and it takes several years to do that obviously to get it to scale. Uh and then normally what has happened over this history here is that it has corrected for multiple reasons. It could be an economic reason. It could be policy. I mean governments do things that sometimes don't always stimulate growth. So there could be multiple reasons why but in every situation it adjusted.

Now her argument is that's good for capitalism because it resets the cost base and then what happens this build she calls it the deployment cycle then people build up and then it goes to maturity with big societal impact that's pattern has repeated itself since the 18th century will it repeat itself again today people argue it won't um I'd argue back in the dotcom bubble they said it won't there and as well I mean housing crisis it was never going to go down I mean as you know we've all lived these things unfortunately circumstances do change.

It's good for capitalism to reset the cost base. It's not necessarily good for investors who invested in that. That means somebody is taking a haircu