
While the hyperscalers and cloud builders provide the best indicator of what it takes to create state of the art GenAI models and the infrastructure to train them as well as to put them into production for practical use through an API interface, perhaps IBM is one of the best leading indicators for how GenAI will slowly be adopted by the enterprises of the world within their own organizations.
In short: Slowly, but surely. Just like relational databases on mainframes and proprietary minicomputers many decades ago, which made Big Blue what it was back then and what it still is today. The interesting bit for us is that IBM is the one company that has been around for the two major corporate computing revolutions that have caught the imagination of the world, one six decades ago and the other starting a decade ago and absolutely exploding two years ago with the launch of the GPT-4 model and its API interface from OpenAI.
IBM is still a systems company, but it is one that has created AI models and hardware and software platforms for companies to deploy those models on the same systems that they use to run their transaction processing. Like us, IBM believes that most of the AI inference and some of the small-scale AI model tuning can and should be done on the CPUs that are running the back office applications, and it has enhanced the homegrown processors used in its Power Systems servers and System z mainframes, which are deployed at more than 125,000 customers worldwide.
Because Arvind Krishna, the company’s chief executive officer for the past several years, is in charge as Red Hat and soon HashiCorp are giving Big Blue a play in modern computing and it has deep expertise in HPC and more traditional statistical AI that can be leveraged for this GenAI boom, a growing and reasonably profitable IBM is setting new records for market capitalization in the wake of reporting its financial results for the fourth quarter of 2024. As we go to press, IBM’s valuation by Wall Street is north of $211 billion, which is nothing compared to the behemoth that Nvidia has become, but it is up by a factor of two compared to two years ago. If this were normal times with a normal IBM, the company would do a one-to-three stock split and try to run it back up to above $100 a share in the coming quarters.
IBM has new Power11 and System z17 servers coming this year, which have integrated matrix math units for running localized AI routines on those platforms instead of offloading them to expensive GPU accelerators from Nvidia and AMD.
IBM has not talked about the revenues it has generated to date from AI, but it does talk about its cumulative bookings for hardware, software, services, and consulting related to GenAI. Here is a table that outlines how this has grown since IBM first started providing data back in Q3 2023:
In the December quarter, IBM said that it had more than $5 billion in cumulative GenAI bookings, which is a factor of 18.2X greater than what it was back in Q3 2023. (Our table has rounded these numbers to the closest billion dollars because IBM is not being more precise.) Of this, about $4 billion of the booking is for consulting services and $1 billion is for various software – Watsonx, RHEL.AI, OpenShift.AI, and other code – that IBM sells for creating and running AI models. Thus far, IBM has not given out an allocation of its system hardware revenue dedicated to AI workloads, but it might once that starts taking off. Then again, it may not. IBM is generally pretty secretive about its hardware numbers.
While that GenAI software number is pretty small, it will eventually equal and probably surpass the consulting engagements once customers move from prototyping GenAI to rolling it into production – provided that customers using Power and z machinery for their backend applications decided to deploy their AI inference models on that iron. IBM has its own AI accelerator, called Spyre, which we talked about last August, that can be added to Power and z systems that need more AI acceleration than the IBM CPUs offer.
This is still very early days for the deployment of GenAI in the enterprise, but we know one thing for sure: If IBM makes GenAI easy and integrated with Power and z iron, Big Blue has a good chance of turning GenAI enthusiasm into money – very likely billions of dollars per year, and maybe tens of billions.
In the meantime, we have to count the systems business that IBM actually has, which is the foundation of the systems business it is evolving into.
In the December quarter, IBM posted $17.55 billion in revenues, up 1 percent year on year, with gross profits up 1.7 percent to $10.44 billion, but pre-tax income down 11.9 percent to $3.31 billion and net income down 11.3 percent to $2.92 billion. IBM is at the end of the Power10 and System z16 product cycles, which are nearly four years in at this point, and as you might expect, revenues are slowing down.
The fourth quarter is historically a good one for IBM, and that is because IT departments have to spend their budgets before each calendar year ends; the second quarter is usually a pretty good one, and the first and second quarters are alright but nothing too exciting unless a product introduction doesn’t end up where it was supposed to be.
So the fact that IBM’s Infrastructure group had a 7.6 percent decline from the year ago period to $4.26 billion but was up 40 percent sequentially from a pretty weak third quarter was no surprise to us. Of that, just a tad under $3 billion was for servers and storage and $1.28 billion was for tech support on the hardware. Mainframe revenues were off 12 percent year on year, and Power Systems plus storage was flat compared to Q4 2023. These figures, by the way, include the fairly substantial monthly licensing fees for proprietary operating systems on IBM’s platforms.
Including base middleware and transaction processing software, as well as financing, we think the core systems business at IBM represented $8 billion in revenues; this figure does not include databases and other higher level application software. About $4.28 billion of that was pre-tax income, which means the other parts of Big Blue are losing money. Red Hat drove $2.07 billion in sales in Q4 2024, up 16 percent, and If you think that Red Hat’s systems business – meaning not databases, application development tools, and such – is 70 percent of that total, then add another $1.45 billion to the IBM “real” systems business to push it up to $9.46 billion in the quarter, actually up 2.1 percent year on year and up 26.6 percent sequentially.
This is why IBM is still a player in systems, and why we think it can build respectable GenAI revenue streams in hardware and software, not just in the consulting services that represent its current play.
It is hard to overstate how important the Red Hat acquisition has been to the transformation of Big Blue, and as we pointed out last quarter, that $34 billion investment has more than paid for itself, and if you take into account market cap and the capability to tell a new story about itself, then the Red Hat deal paid for itself a long time ago.
By adding matrix math capabilities to both the Power and z CPUs, IBM can offer inference processing within the security perimeter of its systems and have customers just keep adding Power and z cores to their compute complexes. This strategy has worked for the past two and a half decades on the mainframe with Linux partitions and new workloads that would not have been run on a mainframe were it not for the fact that IBM realized it needed native Linux on these behemoths in the late 1990s and did something about it to make Linux run – and run well – on its iron.
Just like Linux represents more than half of the installed processing capacity on System z mainframes today – how much more, IBM has not said – and probably less than half of the revenues, the day might not be too distant when System z and Power machines will use a mix of on-CPU accelerators and Spyre accelerators that look to systems software like they are inside these systems to drive something close to half of the revenue stream for IBM’s systems.
This is IBM’s plan, as we said three years ago in s story called For IBM, AI Inference Is The Most Important HPC, and we can’t think of a better one for Big Blue. Chasing after Nvidia – or even AMD or the zillion AI startups – in the AI accelerator business directly would be folly. But making this all transparent and easy for its own customers makes very good sense.
Given how big a large language model must be for the benefits to outweigh the hallucinations, creating a CPU that is GPU-like allows comparatively bigger main memory to be used to run the model.
IBM is clearly solidifying its heavy metal iron workhose of the enterprise posture here with the Spinal Tap Power11 CPU, a bit like a steampunk locomotive in a way, staying on its inference rails as some other chief Crazy Horse (of Paris? Texas?) trains AI brains on GPUs, risking losing its sails in the process, Moulin Rouge-style ( https://www.foxnews.com/world/sails-iconic-moulin-rouge-windmill-paris-collapse-ground-lost-his-soul ).
No such emotional distraction here, just steady GenAI consulting, software, SerDes-attached memory, and Spyre Gyra accelerated jazz fusion ( https://www.nextplatform.com/2020/09/03/the-memory-area-network-at-the-heart-of-ibms-power10/ ). There’s something to be said for the non-folly of a temperate fox, wearing a red hat (I think)! 8^b