Hi everybody. I'm Nicolola Tangen, the CEO of the Norwegian Sovereign Wealth Fund, and today I'm in particularly good company because I'm with Arvind Krishna in New York. Arvind is the Chairman and CEO of IBM, one of the most iconic technology companies in the world. Arvind has been with IBM for over 35 years and became the CEO in 2020, having since orchestrated one of the most striking turnarounds in big tech. When he took over, IBM had been declining for years, and today it is growing faster than it's done for a long time. So Arvind, warm welcome.
Thank you, Nicolai. It's always good to talk to you.
Absolutely. Now, a lot of people still think that IBM is a kind of company from another era, and you have changed that. So today, what does IBM do?
IBM is largely a hybrid cloud and AI software company. We made the transition to that; it's almost half our total revenue. We have another third that is in consulting, and we try to help our clients transform for the current era of digital and AI. And then we have about 20% that is hardware. I realize many people think that we are largely a hardware company, but that is just a fifth of the company—a very important piece, but a very small piece.
When you took over as CEO, IBM had been declining for some time. What was your diagnosis?
I always like to sit back and think: what are your strengths and what are your weaknesses? So as I talk to our own team and as I talk to clients, it comes out that we were trusted, but we were considered to be part of the past, not necessarily the future. So my diagnosis was: you have to do things that are relevant for people's future. They're not always the biggest revenue in a month or in three months or in one year, but they become the big revenue over the next many years. So we began to say, what are we good at, and then can we double down on helping people transition towards a hybrid cloud. We were strong believers that sovereignty would remain important for many, many years to come. And so you double down on the portfolio that helps them do those things. And back in 2019, I was convinced that AI would be big. It took another three years for the world to wake up to that.
What made you so convinced at the time?
Data is going to overwhelm you, and value is going to be derived from data. What can unlock the value from that much data? The only technology we knew was AI.
You made some big acquisitions, and a very large one, Red Hat, which has been tremendously successful. Just tell us about that.
So in 2017, I came to the conclusion that public cloud is very important, but that IBM becoming a big investor in public cloud is probably not good for us. It was a pretty economic conclusion. When you are that far behind, you would have to spend multiple billions—5 to 10 billion a year—to try and catch up. And if you think at the end of five years you're still going to be number five, that doesn't seem like a worthwhile investment. So instead, I wanted to partner with all the big cloud providers. And Red Hat—there were only two or three companies I thought that could help give us a portfolio that makes us a great partner for all of them, that helps their business and helps our own. Hence, Red Hat.
What is the analysis that goes into such a decision? You got to see because in some areas you can say I can compete because I can actually build something, and at the end of that time I can carve out our own niche or your own market share.
I felt that the others were spending so much that they would remain ahead. So that is not purely financial analysis; that's also analysis based on looking at the technology trends and at the strength of their technical teams in addition to the pure numbers. Then you say, okay, if it's going to take a lot of capital investment, then is that capital investment going to have its ROI or not? That is perhaps much more analytic than not. And then is there an alternate capital investment that pays off better? I felt that if we put our capital into doing software M&A, that's probably a better return for IBM. Maybe I'm lucky, maybe I'm smart—that's what the last six years have shown.
Yeah, well, I suspect you have been both. Well, we need—I mean, it was a very smart decision, and you've been proven right. Now, you also spun off the IT services business, which was like a huge part of your number of people, right? A third of the workforce. What was the thinking behind that?
Look, so the negative always is that these are clients that are intertwined with everything else we do, so trying to separate it is painful. Let's acknowledge that for the employees and for the clients. I was clear—I mean, as you began, you said IBM was declining. I was clear that revenue growth is essential. If you say and conclude revenue growth is essential, then something which is itself declining at 5% is something that should not be part of it; otherwise, your target for growth becomes that much harder. That means everything else would have to grow at 10, not at 5. So the revenue decline—I felt that that was an area that it didn't quite fit us well now, though it did very well 30 years ago. I wanted to have a company that is based on innovation, that is based on high margins, and that can grow. That is an area where people look for stability, not innovation. It's an area where it is not going to grow because I think it is fundamentally deflationary, and it is not going to be high margin because by its nature it's going to be decent but not high margin. So that made it a decision to say such a company is probably much better served by being by itself and can unlock more value for its investors over time as opposed to being part of us.
The latest you've done is Confluent. What does that add to the business, you think?
So I think that back to why can you get more value from data? AI is one thing, but then you've got to be able to expose the data to all of AI. Confluent is the best infrastructure for beginning to move data and for exposing data for everything else. Also, there are so many companies that struggle with having real-time data. Confluent with the Kafka backbone is the best in the world at taking data and making it real-time for all the purposes that you might want to use it for. So those two things together make it a wonderful acquisition. I believe time is going to show how great it is for us.
It's for sure a great company and great product. But how, to which extent, do you integrate these acquisitions, or you know, your view on letting them do their own things and operate on their own?
The only one that we have not fully integrated yet—and I'll emphasize the yet—is Red Hat. In all the rest, I always look for three things. It should be a great capability. So to the point you're making on integration or not, you've got to make sure that even if you integrate, the engineering team who's building that great capability still has freedom to build it because they are the ones who really should be in charge of what they are building. We'll try to bring them more of how they build. We'll bring them more AI, we'll bring them more tools, we'll bring them more global capability, but they should remain in charge of that. On the go-to-market side, I really firmly believe in full integration because in order to unlock the value of many of these companies, we can take it to more clients. We generally have a presence in far more countries and geographies than anybody we acquire. But to do that, you got to integrate; otherwise, you've got to grow a unique capability everywhere else. So we believe that bringing our digital capability, our geography footprint, our complete international global go-to-market capability is a big plus for all of these entities. The third part—to operate, you also need to worry about compliance, about contracts, about recruiting, about HR, about payroll, about taxes, about cash management. There is no value to me to have those as unintegrated functions. So my love would be to do it on day minus one, but to be fair, it takes a few weeks or months to get it done.
You said you haven't integrated it yet. What does yet mean?
I believe that we are right now in the process, even in Red Hat, that what I call the third bucket—around HR, legal, contracts, cash management, treasury, all that—we did over the last two years. So that's done. Now, Red Hat definitely had more scale, so they were not going to get an advantage by being integrated with IBM, let's say, in the UK or Germany, unlike smaller ones like Confluent. So we let the go-to-market be independent. I think that we are now discovering that even on the go-to-market, there are areas around digital that are outside the top few markets where integration was going to help more than be a detractor. Engineering in Red Hat, I actually believe, given the open-source nature of Red Hat, is going to have to be its own function. So engineering I will likely not integrate because working at their scale of open source, maybe that is one where IBM learns from them, and things that are open source should belong much more in the Red Hat methodology than ours.
What's the best thing you've done since you became CEO?
Make the culture much more willing to take risk. I think that we had become a very risk-averse culture. That is not the thing I would have told you if you'd asked me on the first week. But after a couple of years of observing and doing, I think that making the culture much more willing to take risk is the biggest thing I've done.
It becomes a question of you have to first ask yourself that if I don't like the culture taking risk, why is that? You got to begin to understand that. And I think these lessons are drawn from whether it's biology or whether it's history. You said we were declining. When a culture begins to decline, the focus becomes inward. And I think it's a natural—I don't actually call it malicious. Human beings are very good at saying, how do I survive? If the culture is in decline, then people begin to say, I survive by not raising my head, by not looking like an outlier. And so it becomes the thing which becomes self-reinforcing, not necessarily by design. So then you have to say, how do you unlock that capability to take risk? You've got to hold up examples. You've got to tell people, 'I want you to take risk.' I tell people, 'Don't give me your 90% confidence. Give me your 50% confidence.' And then you lean into it, recognizing they're probably not going to meet the timelines of the quality you want at 50. So you build a bit of a buffer, but urging them to go down that path is a big unlock then in total productivity and in how delighted clients feel.
Do you think also there is a factor of, you know, a fact that when you decline, a lot of the risk-takers leave the company, and so you are in a way stuck with the most risk-averse people?
That is definitely a big piece of it. But then if you can unlock risk-taking even amongst those that are left—of course, you got to get new people. I'll completely acknowledge that. I think a 10 to 15% refreshment rate per year is a great one. But you can actually unlock even that, because my point being, if the ones are risk-averse, it's a learned behavior as opposed to inherent in them. Then they can unlearn it. If it's truly inherent, then that's different.
So that's the best thing you've done. What's the worst thing you've done?
I think I've been slow so far in terms of client expansion. I think that we are very good at dealing with large clients, and many of our people, and probably myself, turn that B2B into thinking that you're B to large B. You need to be good for B2B to everybody, and I think that those are things that we are yet to unlock.
And how will you unlock that?
A lot of focus on it. You got to say the longer tail is not going to necessarily want to buy everything that we sell. We tend to have a habit right now of everybody should buy everything. So you got to then get more focused in what you do there. Those people also are not making large decisions. They're not viewing IBM as their partner. They're typically purchasing a capability. So you got to be really good at saying, okay, if all you want to buy is that capability, I'm going to give it to you at a great price, at a great quality. So you've got to begin to do those things, but those are changing some of the plumbing of the company in terms of how you sell, how you price, how you go to market—all of that.
Moving to AI, which part of AI is a bubble?
I think that some of the infrastructure buildout is probably a bit ahead of what the world can tolerate for the next few years. The way I would phrase it is—because I have been accused sometimes by saying that it's not a bubble, but I believe that some will disappoint, many will thrive. But not all will thrive is the way I would phrase it.
So when you say the infrastructure buildout is a bit ahead, what does that mean?
Look, by the math that I have done, about a gigawatt of power—you can debate—costs you 60 to 80 billion dollars worth of semiconductors to go populate it. So if you look at people have committed over 100 gigawatts of AI data center buildout, that points to 6 to 8 trillion worth of buildout. If you say that that's got a 5 to 7-year payback, you are going to need an extra 1 to 2 trillion a year of revenue, because inside that 1 to 2, even if it's high margin, that high margin would be 20 to 30%. So that much incremental revenue I don't believe is there, and so that's why I think it's a bet ahead. I also believe on a second one that many of the largest models are going to be a commodity. Commodities can have a lot of value, but there is low switching cost usually between commodities. If there is low switching cost, that means you can have a margin, but it's not going to be a margin with a massive moat around it. So those two make me believe that perhaps there aren't going to be a half dozen to a dozen companies who can build the largest models and survive. Maybe two or three. And that then tells you the second side of it is how much can be the total capital expense that goes into the data centers. If you had said it was half as much as today, I would have said that completely makes sense. I mean, that aligns. But when it's double of that, then maybe some of those are not going to be able to get a great return.
So who are going to be the losers?
That is very hard to predict. I mean, like having gone through a few technology cycles, it's pretty different who's going to be the winners. Generally from AI, I think like, look, some of them that already have a very large consumer business—that means you have a natural distribution advantage on the consumer side. On the enterprise side, I think it's wide open to decide who's going to win. I don't think that that is predetermined. On the consumer side, I think history has shown us if you have distribution and if the distribution is aligned to AI, there's a pretty good chance you will be one of the winners.
Are you surprised that there is such a tech overhang? That we are so slow in utilizing the full capability of this technology?
No, there's a human time scale always. Technology can move at its rate and pace, but people take time because you get into the questions always of is there a risk? Am I going to lose something? Always you get both sides of those voices. Whether you go back to the 1700s industrial revolution or whether you look at the last one, which was the internet era, or you look right now at the AI one, you get those voices that makes some people run and embrace. It makes many kind of watch cautiously on the sidelines, and it makes some dislike and hate it. So as you look across those three, each one, I'll just observe, has gone faster. So if I look at computers and semiconductors from the 50s to the 70s, probably took 20 years to get fully embraced. If you look at PCs, that was circa 1980, probably took 10 years to get fully embraced. If you look at internet, that probably took five years, '95 to 2000. So this one, but it's still years. It's not yet months.
So how long will this take?
I think we are right now—I used to say, if I do a baseball analogy, and our global audience may not always love baseball and know baseball, but innings—it has nine innings. I used to say AI was in its first innings. I would say maybe it's in its second innings now. So if we are in the second innings, it probably is going to take three, four years to get enough to embrace it, but probably not—definitely not 10, but it's not already done.
But when you talk—you talked about these other technology shifts, how would you compare the AI advancements compared to those shifts?
I think it's bigger than mobile and it's bigger than cloud. It's probably in the same category as internet if you go back to 1995. I use '95 because that's—
You are much more—I mean, many clever people say it's like 10 times the internet, but you don't think it's that.
No. And you are very clever, by the way. I should mention that. Not that he needs explaining, but our mutual friend Malcolm Gladwell said that you were the cleverest person he had met. I don't know whether I'm the cleverest, but I'm older than many, and I have been around a few of these. The internet to me was fundamental. If you think about it, because many people don't realize the scale of global business that happens today is because of the internet. The ability to move work in technology from country to country is because of the internet. The fact that a small producer, be it in Europe or be it in Africa or in China, can sell and be part of the supply chain for a global corporation anywhere in the world is because of the internet. If I look at that, the internet has had a massive impact. The direct revenue even to the technology providers is measured in the trillions today. The rise of all the social media companies could not have happened without the internet. The rise of cloud could not happen without the internet. If I include all those as the impact of the internet, then AI is going to be in that category.
But do you think you are biased in your assessment of AI given that you have tried before and failed?
The fact that we tried, failed, but remained committed, and I made the bet on AI in 2019 again before the advent of ChatGPT says no, I don't think we are biased or jaded, because you can imply that in your question. I actually think it's going to be very, very powerful, but it's in the same—
But tell us about—I mean, you tried with Watson, right? Just tell us about Watson.
I think that Watson was the right goal. So in 2011, we wanted to prove that AI can do and solve problems that were unimaginable at that time. People couldn't imagine that you could do natural language question and answers on kind of a gray area using AI. Winning the game Jeopardy, which had these questions in English but with all kinds of hidden puns and language that had to be interpreted, not just the black and white question, proved that AI can do those things. Then you could say, but you failed. Why did you fail? The reason is we took that technology and we said we want to construct monolithic applications, and we unfortunately picked the vertical that is the hardest: health. That was the mistake. If we had taken it and said, let's use it to help corporations get better, let's use it to let their customers get better customer care, let's use it to digest all the documents that enterprises have, I think we would have been five years ahead of where we are today on AI. But it would have been a long journey because the technology then was still what I call bespoke. What I call bespoke is yes, AI worked, but it worked when you had one use case with one corpus of data. Today's AI actually can do many use cases. So that's a big advantage. Two, you don't have to relearn if some of the data changes. You can add the new data and add to it. Those two things make it much more industrial scale today than the one from 10 years ago. And that is why we didn't succeed then. But I do believe that we will succeed this time around.
Are you basically building the infrastructure layer that the AI systems need?
We are not. We rent the semiconductors from other people as appropriate. So you can imagine all the big cloud players—I won't name them on this call—but we tend to use many of them. We tend to not build frontier models. Those we will get from those. The point I had made earlier, I believe that there will be low switching cost. So this will be based much more on what are the T's and C's, how can you have a business relationship with our providers. We will build small domain models when necessary, but not because we necessarily want to, but because very few people are building the smaller models that can then augment the bigger models. So we will actually—we have a belief that most people are going to be multi-model down the road, and so we want to help enable that for our clients.
Yeah. And you got like Granite and so on, right?
We have open-weight models called the Granite family. But in there, I'll note we have not built a single model that's over 100 billion parameters, and the biggest ones now are in the trillions. So we don't want to go there because we believe those are good enough. We don't have an advantage in coming there. But we will build many, because smaller models are going to be much more power-effective to operate or much more cost-effective to operate, and where people care deeply about where is the data, they could operate them on-premise or at the edge, which is an advantage for some workloads and some clients.
The Chinese seem to be agreeing with your thinking.
I think that the Chinese are very focused on enterprise deployment, and they're very focused on sovereignty. We can ask ourselves how many people should care about those same two attributes.
So if you just look at how you work with AI differently from the likes of Microsoft, Google, Amazon, what would you say it is?
So many of them today are very focused on having a consumer approach to AI, which is I want to build a single thing which hopefully more and more of the billions of people in the world can begin to use. That's not us at all. We want to build great AI that our clients—is it Nestlé? Is it Elevance? Is it Pepsi? Is it Bank of America?—can they use this? And that's a very different goal. So I am not ever going to build one thing that everybody's going to use. We are very much focused on building that which our clients need, and which is probably not the consumer side of their business. It is the enterprise side. Can I help them on procurement? Can I help them on accounts payable? Can I help them on how they are leveraging data for their own business decisions? Can we help them on all those things is where they will tend to use us for artificial intelligence.
When you see things like the latest Anthropic model, which is so powerful they can't even release it, what kind of reflections do you have?
I have two. At a very deep computer science level, this is not new. Let me observe: really clever people with extremely sophisticated tools have been able to find vulnerabilities and have been able to exploit them for decades. Yeah, I remember when I was in graduate school, this is 35 years ago, we had the Morris worm that came out of a really clever kid at Carnegie Mellon. Okay. But if AI can inherently use some of those capabilities and you can harness or harvest it into that, that means you're now letting a person with a high school education able to exploit these models to do what those really well-trained, really clever people did. That opens up the speed aperture, unfortunately. So these models may well be able to exploit things in seconds that used to take really clever people months to get done. So the intensity and the speed is what we have to worry about. It's not the fact that it exists. It has always existed. But that means that you have to be able to have layered defenses, and you need to pick your right partners to help protect the enterprise, because I actually think that some of the smaller vendors may not have the capability to defend against these models.
Do you think it'll be a requirement to pre-release models to some kind of regulator or oversight function to make sure it's not too powerful before it's released to the public?
I think that that sounds good on paper. So in practice, how do you do this globally? Because you can say that here. So do you stop the Koreans from doing it? You stop the Chinese from doing it? How do you get to being able to control these things globally?
I am skeptical that the technology can be regulated. I believe the use cases can be, because it's embedded in all the races, and it's physical goods that are tangible or possible to regulate because you can put border conditions, because there is a weight to them. You can control how much and where. A digital good that can cross a boundary, I think, is really hard to regulate. If you think about some countries in the world that some or the others don't always love the regime, and that side tries to control internet access, you tell me how effectively they can or cannot do it.
Is software going to die, you think? I mean, your stock was down 13% when Claude Code was released.
No, I'll give you my reason for explaining why. So we didn't talk about demographics, but if the number of people in the world is going to decrease, then seat-based software by definition has a smaller market 10 or 20 years down the road. I think investors are quite smart at recognizing long-term trends. So I think one thing is how well will seat-based software do. Let me fully acknowledge to you, I think AI and agents will replace some of the front end of software. But if you replace the front end, then by definition it has less total value. So you combine fewer seats, less value for some things. That said, the system of record, the database that contains the business function, the business logic, that is still important. But I do think that the value for some of the software where the front end was the prime value, that is going to decrease. And then to give full credit to the investors, they're saying, look, I can't decide today who falls into that camp, who are the few who might benefit. And then let's acknowledge maybe half are not going to have a help or a hurt. If I can't determine that, I'll take the sector down, and then over time that'll determine itself based on the numbers.
that you print. So I think that I could see that a fourth of people where the value was largely the front end could actually have a long-term impact.
That would kind of be where I would finish in this. So I would look at you and say no I actually think that we were hit in a way that was unfair but look the market depending on who you talk to is down from anywhere to 40 to 60% in software for many companies and we are down about 25. So in some sense we are taking a hit but I think that means there are enough people who also think wait a moment you may not get that full hit and so that's kind of where we are.
Talking about expected hit, your mainframe business is thriving right? And a lot of people didn't think it would.
So the thing which cloud core thought that it is going to replace is the thing that is thriving the most. That's interesting isn't it?
Yeah. So why is mainframe still thriving?
Because the mainframe workloads tend to be workloads in critical industries. They're doing workloads like retail banking. They're doing workloads like credit card transaction authorizations. They're doing workloads like airline reservations. They're doing workloads around warranty and maintenance. It's systems of record. It's places where that full six nines of availability. I want to make sure that the transaction is not corrupt. I need to make sure I can get the batch workload done in 20 minutes at 5:00. Those workloads are only increasing.
But should those workloads have been moved to cloud already?
If you want to pay three times as much, right? So economics would dictate that that's the more expensive answer.
Did you think when you became CEO that mainframe would continue to grow?
I was 100% convinced of it.
So the 10 years before I became mainframe was declining. It's surprising, isn't it, that in the last six years mainframe has grown every single year.
It's incredible what one person can do.
It's not one person. It is unlocking back to risk taking. It's unlocking the enterprise. Many people there believed it could and so you had to unlock them and give them the environment which allowed them to succeed and thrive. We have built AI into the mainframe that was not being done a decade ago. We built more and more capability. We built and encouraged our partners to build more software. Actually we built our own COBOL conversion code leveraging large language models three years ago. So it's harnessing innovation and putting it into the platform that then allows it to thrive.
And you put AI on the platform before ChatGPT existed.
That is correct. We put it on the platform in 2021. Then we put more in 2024 and then we put even more this year.
And tell us about this new version, the Z7.
