What’s in store for the cloud? That may be a hard question to answer because there’s a pretty significant gap between what’s currently happening in the cloud and what we think/hear is happening. Still, I’ve gotten comments from 167 enterprises in 2H24 that could shed some light on things. Some are direct views, and others (which I’ll identify) are interpretations I’m offering from what I’ve heard.
First, all 167 of these enterprises were cloud users in 2024, and I think the great majority were also in 2023. No non-users expressed plans to adopt the cloud in 2025, no current users planned to turn their back on it either. Of the group 37 companies said they expected to increase their cloud usage next year, 35 said they’d be looking to reduce it (repatriation), and the rest had no comment one way or the other.
Of the group, 48 said they had multiple cloud providers, but I think the actual number is perhaps more like 65; remember I’m getting unsolicited comments not asking/surveying, so there’s always a chance that an enterprise simply doesn’t say anything about a topic. However, only 5 of this group indicated that they actively used multiple providers to support the same applications; the great majority of multiple cloud users assign applications to a single cloud. Of the 48, 29 said that one of their cloud providers was used for SaaS, and of that group 21 said that the SaaS services were managed directly by a line organization, not by IT.
Interestingly, four of the five “true multi-cloud” users said their alternative provider was not one of the cloud Big Three; IBM and Oracle were both cited. In addition 14 enterprises said they were “considering” or “evaluating” the use of multiple providers, and in that group, all said they would look at both the other Big Three and outside that group, again mostly at Oracle or IBM.
Of the 37 companies who expected to increase cloud usage in 2025, all thought the majority of their increase would come from current cloud applications. Nine expected “some” contribution from new applications, but only two thought that new applications would add significantly to their cloud usage. Of the 35 who expected reduced cloud usage, the majority expected this to result from either usage management of current applications or redesign of the applications to reduce cloud usage. Only two said they planned to remove an application from the cloud completely. Interestingly, four said that their reduced usage would come from dropping a second cloud provider.
Overall, it is expected that cloud usage and spending will rise, since net of the gains and losses appears positive, but cloud growth isn’t expected to be as much in 2025 as it was this year, which in turn was less than it had been the prior year. Like any form of computing, cloud computing has a kind of natural level that, as it’s approached, limits incremental growth.
Note here that enterprises separate “cloud” and AI, except where they acquire AI as a cloud service, something only 14 of the companies said they’d done. That doesn’t mean there’s a lack of interest; the notion of getting a secure, meaning sovereignty-guaranteed, form of GPU as a service is popular with a full third of enterprises, mostly for use in training self-hosted AI models.
GPUaaS is complicated, say enterprises, for four reasons. First, it’s a balance between data sovereignty and cost, two things that enterprises need at the same time. How do you then balance them? Second, more and more enterprises see AI hosting as a kind of “revenge of the cloud provider” thing, meaning that they believe that just as they’ve gotten onto cloud-provider manipulation for normal applications, those providers introduce cloud AI to keep feathering their nests. Third, there is relatively little enterprise interest in GPUaaS except during model training, and during that period enterprises would tolerate a higher cost than they would for enduring AI hosting. Finally, fitting GPUaaS training to self-hosting of AI means understanding what the latter demands, and enterprises see AI as a kind of ultimate moving-target example.
Sovereignty is a contractual guarantee of data security, according to enterprises, and that means it can only be provided by a credible source with deep pockets. There are a lot of comments out there on the potential opportunity GPUaaS presents telcos, and there’s some basis for these given that telcos are near the top of the list of trusted guarantors of sovereignty. The problem is that enterprises realize the linkage between AI models and techniques, training, and GPUaaS is critical, and telcos aren’t seen as having any understanding of AI at all, and presenting no credible career path for AI professionals. One CIO told me “You know who’d be an even worse employment choice for an AI expert than an enterprise like us? A telco.”
For the second point, enterprise IT sees the cloud-AI play, directed at personal productivity augmentation of “Office” tools, as largely wasted money. One reason is that their own experience with “copilot” tools in coding has been increasingly disappointing. A year ago, five enterprises out of six said they believed AI assistance in coding would be highly valuable, but within six month that had dropped to less than half, and in the last quarter of this year, one in four continued to have a positive view.
The remaining issues reflect the increased dominance of analytics, what might now be called “agent-oriented” AI. This AI application set depends on company-private data, which means that it’s the same sort of data that enterprises have already decided they can’t cede to cloud storage. It has to be kept in-house, which means that the AI model has to be hosted in the data center. But training demands (say enterprises with experience) five to ten times the AI resources that running a trained model would require, and there’d be little or nothing to do with those extra GPUs when training was completed, which would make self-hosting prohibitively expensive.
Then, of course, there’s the problem of deciding how to self-host in the first place. Enterprises see AI as a moving target, and are reluctant to aggressively pursue it when it seems a new “better” model, technique, or both comes along every week. Enterprise uncertainties on AI are magnified by the concern that an early AI decision will prove sub-optimal when something better comes along. On the average, enterprises say that an AI approach is obsolete in about four months.
This enterprise “wait-till-things-settle” reaction to the pace of change in AI is impacting the cloud for 2025, I believe. AI is expected to revolutionize business, yet right now it’s changing too fast to adopt with any sense of securing full value from it. While this is the case, do you adopt something else in its place? That’s not how to deal with a revolutionary technology. AI is holding the cloud, and IT in general, hostage right now. I think that’s likely to be true through the first half of 2025, but by 2H25 we may finally understand how to move forward with AI, and where AI isn’t the answer. We’ll also know where AI needs augmentation, in particular how it links up with IoT and digital twins to generate data and establish context in real-world activities. Watch 2H25, then, for signs of where the cloud may be heading in the long run.
