It's my pleasure to introduce Rob Thomas, chief commercial officer at IBM. We are the reality guys. So, I know you guys were hearing some really cool frontier tech stuff.
That robot pancake looked pretty good, right?
Yeah, that did look good. [laughter] So, Rob runs all software at IBM and all of sales GTM. He reminded me over dinner that there are 20,000 folks in the GTM team, and just has an amazing wealth of experience on how does this AI that's being built and talked about at this conference, how is that going to actually translate into the enterprise? You spend a lot of time talking to, you know, CIOs and heads of technology and obviously heads of countries in your role. Rob, it's...
Like I'm in the epicenter of doom here in the valley.
Supposed to be the land of optimism at all.
The land... Have you ever seen this Louis C.K. skit where he's like everything's amazing, but everybody's miserable? I feel like that's kind of where we are right now.
I don't know. I feel like it. That's the vibe I think we feel across the country is like... why is that?
Yeah. So, well, the question I want to ask to start with is that what is it about enterprise AI that the Valley is getting wrong? As in, when you go and travel middle of America, the rest of the world, what is it that we are getting wrong when we think about enterprise AI?
I think scaring the customers is a bad idea. So I do think we should stop that because I think obviously we should have the right safeguards, but scaring people is not good. It's not going to be good for business for anybody.
And I think there actually is a huge reason for optimism. You look at the 250-year history of America, for every kind of technological revolution, until this expense gets to be about 25% of GDP, it can keep running. Everything kind of falters at about 25% GDP. It's true of steam, electricity, internet. And if you take that as a rule of thumb, we still have like six trillion to go.
So, let's not scare people when we're only a trillion in. We got another six to go. Let's not scare people yet because the companies that we work with, I mean, I'd say the typical IBM client is Fortune 2000. They really want to do this, and it depends on the level you're talking. It's interesting, there's like a barbell. Like in the C-suite, they are dying to do something on AI, and at the developer level, they're dying to do something in AI, and everybody in between is like scared to death. That's the feeling I get, and so we need to overcome that.
So when you walk into a boardroom and you talk to like six execs, you have to make a call as a leader of IBM which side of the divide they're on, right? So what are the symptoms? Because a lot of the people in this room are actually heads of large IT shops. What are the symptoms that you perceive that says, okay, on a scale of 1 to 10, this team's at a four versus an eight? And what does it tell you? And could you talk about some sort of steps that folks can take to get what we call in the Valley AI build right? Like how do you actually start to make the elephant dance? Because it's very clear from this room that these companies are moving and AI is coming one way or the other and one has to get ready for it.
I'll give you two real examples. We've done a lot of work at Deutsche Telecom. And they're trying to manage a network. This is not super exciting to most people, but it's like existential for them. Can we manage a network? Will it perform? Can we prevent vulnerabilities from occurring? Can we ensure resilience? And they had a team of a couple thousand that were basically managing the network. We went in there and said we can do this a lot better with AI. We can massively increase your resiliency. We can also prevent you from ever really having to hire people to do this again because we can do a lot of automation around what you're doing. And when you tell that kind of story to the head of network, they love that.
The developers love it because for those of you that have managed the network, it's not like super fun work. And so they're like, 'Hey, if we can offload a lot of the stuff we don't like to do...' So again, I think you read the room and you can find the barbell shape almost everywhere. So then it's how do you convince, in this case it's the operations person in the middle, that you can do this without creating unnecessary risk and it will be more resilient than it was to start. And we found you can work through that process, but you kind of got to find those levers that are going to get people's attention.
And if you were being direct and honest with your team and just do a self-assessment on IBM, I'm walking off script right now, you told me that I could... where are you internally on that scale of 1 to 10?
I'll give you an example, then maybe you can give me the rating.
So our ambition four years ago was we said we want to become the most productive company in the world. We were not at the time, just to be clear. We were spending 12% of revenue on G&A. Far, far from top decile.
You fast forward to now, we've said publicly we've taken six billion out of our back office. So we've gone from at best mid-decile to the top decile, and this is all through using AI to automate repetitive tasks from HR to supply chain to procurement to how we manage products, manage product life cycle. So like the $6 billion is real, and you can see it in our income statement and in our balance sheet because we've increased our R&D by two billion that didn't come from nowhere.
Our cash flow has gone from 12 a year to 16. And so the six billion is real. And so I think that's the measure. Now, is everything leading edge and AI? No. Some of it is like let's just be way more productive in our processes and our ambitions, but there's a lot of AI that's behind doing this. So, where is that on a scale of 1 to 10?
Okay. [laughter] Just ribbing. Yeah.
It's hard to tell, and I'll actually ask the CIOs in the room to load up your questions for Rob because he's got a wealth of knowledge here. But what I find when I'm talking to CIOs is that there's central IT that obviously has directionality and saying, okay, we're going to do XYZ to get R&D going faster and better or G&A, whatever the systems are, and then there are these rogue projects, let's call it that, going on in departments. Do you try to balance? Is it like everybody can go as fast as they need to in AI? Is that the central thinking at IBM, or is there some sort of central command and control on where you're pushing these things? And I guess the question behind my question is, have you figured out a formula that you are now taking to your customers and saying this is one way to be successful in the AI journey?
I think the top-down piece is you have to have a clear ambition for why you are doing this. And I'm actually pretty amazed that I think most companies don't have that yet. They're just saying, 'We want to do AI. We heard we should be doing this.' Again, our ambition was we want to take six billion out of the back office so we can put more into R&D.
And when did you put that stake in the ground?
22. So, right after... well, actually...
Arvind took over in 2020.
So, 2022. And I think when you have that ambition, then at that point you want to set people loose to say, 'How can we achieve that? Give me ideas.' Because we're not going to come up with all the ideas top-down. But I'm afraid if you don't have a clear ambition, I'm not sure the rogue behavior is going to lead to anything other than a massive increase in token cost.
Which could be good, but it may not be good.
Right.
So, Rob, switching to the stack, right? So obviously there's models, there's infrastructure, there's applications, and the question is where is IBM putting its bets? That's sort of one question. And the other part of this is you partnered with both Anthropic and OpenAI. How do you think of those two guys vis-a-vis open source, where we've talked about that before? How's that going to change? So maybe you can answer both those questions.
So where are we investing? We've invested a lot in vLLM to build basically an open inference layer, and a lot of people have adopted it, which has been good to see. So at the core infrastructure layer, we've done a lot of investment because our belief is we're going to be running a lot more models, primarily open weight, and you need an infrastructure that's performant enough and efficient enough to handle that. So we've done a lot there. Not so much at the application layer, to be... depends on how you define an application, but a decent bit in the middle on governance and data governance, but not so much at the app layer because it's a little hard to figure out, at least for us at the moment, which applications will actually survive or thrive through this.
So those have been the main areas. On your second question, we want to be multi-model. We've invested a lot in optimizing, a little bit of what some of your companies have done, optimizing models. So in our SDLC tool, which is largely focused on proprietary languages like RPG, COBOL, we're doing ABA, we're doing a bunch of things. Anthropic gets about 40% of the queries. The other 60% go to Mistral or Granite. Doing some experimentation with GLM. We haven't brought that into production yet. But we're basically investing in the model optimizer because we're not sure if you allow a developer to choose, they're always going to choose the most expensive model just because, you know, they're lazy. I mean that affectionately. That doesn't really work at scale. So we've done a lot on... we want to build an... you know, back in your database life, you need an optimizer that's making the best decision for the query. I think there's a big spot in that for models.
