Big Blue’s New “WebSpheres” To Surf The Container And AI Waves

July 24, 2023 Timothy Prickett Morgan

Without question, there is a lot of money that can be made on AI training and inference, as is aptly shown by the financial results of Nvidia. That company, which is the darling of Wall Street these days, has gone from datacenter wannabe to datacenter giant in the course of a decade and a half. It has been fascinating to watch, but it is as much the result of lucky timing as hard work, as Jensen Huang, the company’s co-founder and chief executive officer, admitted from the beginning.

Nvidia has been working on parallel computing architectures for accelerating HPC workloads since 2006 and commercialized them first in 2008 and got a toehold on a relatively small HPC market by 2010. And that is when AI research blossomed and convolutional neural networks started beating humans at image and speech recognition and large language models for translating between languages or data formats (text to speech, speech to text, image to text, and so on) were being created.

Nvidia has cornered the market for AI training systems, with AMD winning the big exascale supercomputing system deals that might have otherwise gone to IBM and Nvidia working in concert, and it will also get a foothold in commercial AI training. It remains to be seen what Intel’s fledgling GPU compute business will do, but a market likes three options so we think Intel will get a piece of the action eventually.

After years of losing money in HPC, IBM had a choice: Make a GPU or let Nvidia have nearly all of the money in an exascale or pre-exascale system without any of the risk of being primary contractor. And so, three years ago IBM walked away and then rejiggered its Power10 and z16 processors, used in system stacks it fully controls, have their own embedded AI engines, which are perfectly suited to run many kinds of AI inference and even some AI training at a modest scale – exactly the kind of scale that enterprises using pre-trained models will require, as it turns out.

As was made clear by Arvind Krishna, IBM’s chief executive officer, in a call going over the financial results for the quarter ended in June, Big Blue is going to capitalize on AI through its consulting and software arms and will get whatever AI inference action that comes naturally to its Power Systems and System z platforms. But it is pinning its hopes on two platforms – OpenShift and Watsonx – to catch the AI wave.

OpenShift, of course, is the Kubernetes container management platform created by Red Hat that still, by the way, runs atop the OpenStack cloud controller that Red Hat still supports and sells. (You need something to provision the infrastructure underneath Kubernetes, and most of the time OpenShift customers choose OpenStack although it can run atop VMware’s vSphere or the cloud substrates at the major clouds.) Red Hat was acquired in 2019, when OpenShift was still on the flat part of the hockey stick curve, and Krishna says that OpenShift supper licenses are now driving a $1.1 billion annualized run rate. Red Hat has an annualized run rate of $6.5 billion as the second quarter ended, so that is about 17 percent of revenues, which is a nice piece of business. It is not quite WebSphere, mind you, which is a commercial grade Apache Web server married to a Java application server and which lots of customers still use today and which still drives a respectable revenue stream.

And now, IBM has a WebSphere-like stack for containers and clouds, thanks to Red Hat, and now it has a WebSphere-like stack for AI training and inference, thanks to so many research and consulting engagements with large enterprises who are trying to sort out this AI opportunity. That AI stack is called Watsonx, and it looks to be a much better productization than the original Watson QA stack from a decade ago.

As we have previously reported, IBM launched its Watsonx platform in May, and last week at The Next Platform, when the Watsonx software became generally available, I did a deep dive with Sriram Raghavan, vice president of AI strategy and roadmaps at IBM Research, to get a sense of the inner workings of the Watsonx stack. I am not going to repeat that work here, but you should read it if you really want to understand how much better this approach is than the Watson QA system.

The point that Krishna wanted to make is that Watsonx is at the same point as OpenShift and consulting for OpenShift contracts was at back in 2019. Here is the way the consulting business, which IBM is very much focused on, has shaken out for OpenShift.

“We began with this journey in 2019, and our book of business was – to be precise and to round it out – zero,” Krishna said with a laugh. “In the first year, we signed about $1 billion of business. And at this point, we have inception to-date signed $9 billion with an annual run rate of $2 billion in consulting. I would tell you and expect that we’ll play out AI in a similar way. I hesitate to call it anything else until we get six months or so down the road, in which case then we’ll have more knowledge.”

On the software side, IBM is making money from OpenShift, too, as we pointed out above, and it will do the same by offering commercial support for its models and its add-on software to create the Watsonx AI platform.

“Now on the software side, we are very, very excited by the initial reaction to the Watsonx platform,” said Krishna. “The number of projects we have going on, the client interest, it really is something which we are very, very pleased by. So, what’s the model to think of it on? As I think about how OpenShift, which was a Red Hat product, came in 2019, and it grew literally doubled each year for the first four years. And at this point, the revenue is about ten times of what it was and it came in. And right now, we have quantified it at $1.1 billion on our annualized run rate basis. So that gives you a sense of the excitement we have around these projects, these technologies, and what it could do for us as you begin to go forward.”

So, yes, Nvidia can sell the GPUs. But IBM has a commercial-grade AI training and inference platform, a commercial-grade Kubernetes platform, one of the largest midrange and enterprise customer bases in the world, and a huge army of consultants to get their feet into their datacenter doors. And that means Big Blue will get its share of AI revenue at midrange and enterprise accounts. Maybe more than its share – and without having to take on Nvidia creating and selling GPUs, which would be an enormous undertaking. Just ask Intel and AMD how hard that is. . . .

This Issue Sponsored By

Table of Contents

Content archive

Recent Posts

Subscribe

Pages

Search