If you have an interest in literary classics, you’ve probably heard of the “Tale of Two Cities”. I suggest that to recap and discuss AI this week, we need to launch a “Tale of Two Sitters.”
Every successful technology needs a product and a kind of retail bridge, so it needs at least a key player in both spaces. This week we have evidence that the key player in both these areas, for AI, may have taken a seat, and what they’re sitting on (philosophically, not anatomically) is critical in projecting how AI might evolve.
On the producer side, we have Nvidia. They beat on revenue and EPS, and issued good guidance, and their stock took a hit on the earnings day, but this morning it’s up considerably pre-market. The movement is most likely to be related to hedge fund activity (short anything you can short to drive out a few retail investors, then cover at a profit and maybe even buy in more), but some of my Street friends tell me that nothing in the earnings report has lessened their concern.
Nvidia is sitting on “hyperscaler AI”. Their strategy is less to find real AI missions that could drive growth than to finance the plays they’ve been winning on. This strategy seems validated if you look at the fact that the three AI hyperscale giants all had better results, but as I noted at the time, these results were more likely related to the exploitation of cloud agility to respond to current global economic tension than to AI. In any event, companies who finance the purchase of their products, whether it’s direct or through indirect deals (Nvidia has done both recently) are always multiplying their risks, and risks there are.
At some point, hyperscaler giants will have to either find a good ROI for their AI spending or stop spending. The challenge in that lies in the fact that any new technology is always marketed first to the low-apple buyers, and that means that growth is increasingly challenging. Yes, Google and Microsoft have succeeded in promoting a personal-assist model of AI to the majority (my model says over 60%) of candidate consumers and workers. But those were the easy ones, and even if all the rest are addressed, the pool of incremental revenue and rate of revenue growth will be less than they’ve enjoyed so far.
The problem is that hyperscaler AI is really SaaS in a brown paper bag, which means that it falls prey to the problems of high cost (like what has been driving “repatriation” of some cloud apps) and data governance. I saw a comment on network TV regarding the use of AI in personal finance, and the speaker warned consumers not to upload their tax returns or other personal financial information. Would companies have less reason to worry about data security?
But (roughly quoting a song from “Camelot”), where in the world is there in the world a company who can resolve these challenges? IBM. From the very first, IBM was a champion of “business AI” rather than personal AI, focusing on business analytics and process control, and in their solutions focusing on self-hosting. But…IBM sat down after finding that it couldn’t drive the AI process far enough fast enough. IBM doesn’t sell what self-hosted AI would run on.
A lot of IBM accounts had major AI plans for 2026, and almost all of them have told me that their plans were delayed or even canceled. Why? Because all the PR on AI was about something different, and because IBM couldn’t really push its story to anyone who wasn’t running a mainframe and wasn’t willing to introduce another vendor to execute the AI-hosting piece of the mission. If, in fact, IBM account teams were prepared to suggest that.
Sitting is comfortable, particularly if the alternative is a risky form of musical chairs, where you leave your comfortable seat and hope to find a better one. Nvidia can’t tell the world about the problems of hyperscaler AI. IBM can’t tell the world that the solution to AI business cases is to connect to IBM for data and do all the AI lifting on something else. How do they resolve the problem, for themselves and for AI as an industry? And, of course, for investors?
Skootch your chair to a new position.
Nvidia needs to protect hyperscaler-worker-focused AI, while prepping for self-hosted business AI. One step in that, announced this month, is a kind of AI token/prompt router. You have some smarts that divides AI tasks up among models, model hosting points, hyperscalers and data center tools. You don’t put enterprises in the position of tossing hyperscaler SaaS AI out (and Nvidia revenue with it) to realize new business cases. You let governance, both in the form of cost and data security controls, be applied to jobs so stuff that hyper-AI does well isn’t pulled in-house by the need to host stuff that hyperscalers can’t.
This adds to long-standing Nvidia initiatives in process control and digital twins, and work on chips that run more contained models at a much lower cost. Those have already forced competitors like AMD to spend money and resources on advancing their own hyperscaler-suitable chips, on the theory that you can’t win a war if you’re on the defensive everywhere while your adversary is free to attack your own positions.
IBM’s skootch also came along in August, in the form of a dual-mainframe processor that would let new-model IBM Z systems run both IBM’s z/OS and ARM-compatible Linux (supplied by ARM or IBM’s LinuxONE). This new chip would let IBM sell systems that can be AI hosts, and let IBM push AI tools into Red Hat more aggressively without having to start selling open servers and alienating Red Hat OEMs. Nothing has been said about introducing AI chipsets to IBM Z mainframes, so the AI part of this is still going to have to run on something IBM doesn’t sell, unless IBM buys a source.
What this does, in effect, goes beyond opening up IBM’s options. It pulls AI into the Z and IBM software, management, and security ecosystems. That will help IBM’s chances to profit by broadening the benefits it can offer buyers.
Two skootchers can collide, of course. IBM’s processors are not in their current systems, and some of my contacts tell me that this is going to be a late-2027 move at best. That means that IBM can’t move fast to validate self-hosted AI, and also that IBM is likely to think that server competitors won’t do that either, perhaps won’t even try. So self-hosting is delayed. What does that do to Nvidia?
They hope, nothing, because I’m sure Nvidia thinks that getting the other 40% of the hyperscaler-worker-AI opportunity will cover them. In fact, I think that IBM and Nvidia’s skootches are more symbiotic than competitive…in the long term. In the short term, the big risk AI faces is the political backlash against data centers. At the least, this could raise the cost of supporting more AI missions in the cloud, and that could reduce the period when worker-centric AI can carry the AI banner. In short, we’ve got a couple of interesting developments we’ll need to watch.
