AI workloads are compute-heavy and fully utilized, unlike traditional cloud jobs, making it cheaper to run large models locally than in shared cloud environments. A 70-billion-parameter model now runs on a few-thousand-dollar machine. This shift could reshape how companies price AI services and budget capital through 2027.
I was going to say, though, I certainly hear the security concern. I've worked with law firms and other regulated industries that have those kinds of concerns, which is one of the reasons they often still have significant on-prem facilities, even as they recognize the cost savings of going to the cloud. But I am wondering whether we're going to see more move back on-prem, because part of the premise of cloud was the utility and the economies of scale associated with working and sharing these larger server environments, you could drive down your cost that way. But I think AI makes that happen less because the premise was previously you were largely I/O blocked. You were waiting for databases, you're waiting for file writes, whatever. But that's not the case with AI. AI is compute blocked, and more than that, it's memory blocked. You sharing that compute resource with someone else, particularly if you're going to keep on using it, is the question becomes like, are you going to be using, you know, a very high percentage of the compute associated with the resources that you are putting to work? And because the answer largely was not before, right, the actual CPU utilization was way down here, you start storing and like most things were just about routing data to a database or delivering it back to the front end. The shared infrastructure made a lot more sense and sort of that leads you to go cloud. But as you have more and more use for compute, if your computer is running hot all the time, the utility of going to the cloud goes down significantly. I'm wondering whether we're going to see a lot of institutions, even smaller institutions, you know, taking on more compute that's going to run locally, if, if only for the reason that, you know, it's cheaper. Like, it is remarkably cheap and powerful to be taking a few thousand dollar computer and be running a local model on it, particularly a powerful local model. I'm not talking about one of those 2 billion parameter models that you can run in the background on your laptop. I'm talking about 70 billion parameter DeepSeq model that can do quite a good job to coarse point, can have 100,000 token window, something that's on par with last year's OpenAI, and be completely private and use it for a greater degree of safety. When we talk about everything as code, I wonder we're also going to be saying everything is tokens, right? And everything— and the question's like, and tokens, and tokens are a unit of compute, right? Every token is a, is a cycle. So the— so if that, that means not everything is good, everything is code, everything's tokens, everything is compute, that, that can significantly change, I think, the economics of, of what makes sense for businesses to be putting into place and how they're going to be using it. And that, that can be changing, you know, just even over the course of the next 12 months, but probably having significant impacts on capital budgets heading into 2026 and 2027.
