Is the cloud lagging, perhaps fatally lagging, in the AI race? If so, what does that say about both the cloud and AI? There’s been a rethinking of the notion of “the universal cloud” going on for some time, and it’s clear that cloud computing has been a center of hype. Obviously AI is the same, perhaps even more so, and these questions are important to anyone planning, using, or selling either of these technologies. The Financial Times is a pretty reliable financial-markets source, and we’ll start with its comments to see how they do at what’s really a question of technology planning and adoption.
The article is saying that software companies and the cloud are not receiving an even share of AI benefits. But what, if anything, is the relationship between the cloud and AI? Is it just that both technologies are over-hyped and therefore not as well understood as they should be? Is that why the current trend to rethink cloud plans is projected to impact AI? The title of the FT piece, “The cloud over cloud companies”, suggests that somehow AI isn’t helping the cloud, and yet generative AI is dominated by cloud-hosted services.
What FT seems to have done in the piece is best understood from this paragraph: “If generative AI really does represent the next great sales opportunity for the tech industry, then software companies ought to be among the biggest winners. After all, most AI is likely to show up as enhanced features in the business software that companies rely on in their daily operations.” The problem here, I think, is a kind of conflation, at two levels.
First, there’s the level of just what we mean by AI. Generative AI is the topic of the first sentence, and AI in general is the topic of the second. Generative AI is an application of large-language-model AI, which in turn is a subset of neural networks and machine learning forms of AI. All of this, all of AI, is software. Yes, it runs on special silicon, but all software runs on some kind of silicon. Most of the biggest winners in generative AI, like OpenAI and Microsoft, are software companies.
The second conflation is the conflation of business software and the cloud. The mission-critical software that the paragraph is talking about doesn’t reside in the cloud except, the front-end portion. Companies have been reluctant to move core applications to the cloud for reasons of security, and also because of the cost.
Wait, you may be thinking, how can the cloud not be cheaper? The basic problem with the cloud, as enterprises have been telling me from the first, is that it’s more expensive than traditional data center hosting. That’s not the same as saying it’s always more expensive for every mission, only that if you have something running in the data center, it’s almost certain that moving it to the cloud would be more expensive. Where the cloud wins, and can win decisively, is where an application has exceptionally variable workloads to process, and so sizing resources for the peak workload is likely to result in a significant waste of resources, and money.
If we look at these conflations, we see that the big question is how AI could be used to advance enterprises’ core operations, and what that would require of AI software. Generative AI in the form we’ve been seeing, and using, it is aimed at enhancing aspects of personal productivity. For the great majority of users, it’s a writing tool. AI as a core business tool would require a more planning-and-evaluation mission, something more like business analytics. IBM has been the one company that’s made that point consistently to its accounts, and so is a leader in that sort of AI (see my piece in Network World). Almost two-thirds of CIOs who report success with AI in improving their business operations said IBM was their AI partner.
Companies, in my view and based on what they’ve told me, are not finding AI to be a broad business revolution (see another of my Network World pieces). One big reason is that the basic generative, cloud-hosted, AI tools aren’t useful in that mission, and unless some vendor promotes a reasonable approach to AI analytics, the user is adrift in a sea of generative AI hype.
There are a number of issues that appear to be limiting software vendor interest in pushing a realistic analytics-focused AI strategy. If data security is an issue, then self-hosting the AI model is a requirement, and that involves a significant hardware investment. Vendors are always titchy about pushing a technology step that involves a large investment in something they don’t make. Another problem is that generative AI adoption bypasses traditional tech sales processes because it’s delivered as a service, with no project to need approval and no capex. A broader business mission for AI means both.
Training the AI is another major issue, both in terms of data and data security and in terms of resources. Public AI models of the sort most people use regularly are broadly trained on web data, and as more and more web data is AI created, there’s a growing risk, pointed out in this thoughtful piece, of models suffering a drink-your-own-bathwater risk. In any event, broad training may not be helpful for business analytic applications. Training most LLMs will take more resources than using the model, which means that more time is likely to be needed, and there’s a question of how to harmonize the need to train an analytics-focused model and keep it up to date; you need to train it on a data set, but provide current data to analyze.
Self-hosted AI has that same bathwater risk, perhaps more so since enterprises aren’t saying they pay much attention to how AI-created data is used and stored. Business operations works off history, so to the extent that AI tools create history in the form of regular analysis, they’re creating a risk that an AI hallucination, even a small one, will contaminate the future analysis because it’s fed back onto the AI analysis. It takes a fair amount of expertise to design an AI analytics flow to avoid this problem, a level rare in enterprises and in most AI advocates as well.
All this complexity has, I believe, focused vendors more on erecting AI billboards than on actually promoting business AI. We have a lot of vendor interest in self-hosted AI, but so far enterprises aren’t getting much sales attention on the topic, except for the big IBM accounts. I think that may be due to the fact that there’s still a broad set of options for how one self-hosts AI, how it’s integrated into business-analytics missions, and how it’s monetized both for user projects and vendor profits.
I think the FT story is really about the problem with how generative AI has captured the focus of AI business use without realizing the real value, but I don’t think it’s the fault of AI but of the software vendors. Hype is addictive and too many are simply raising the AI flag rather than doing the work of creating a practical way to use AI to improve business operations. The problem with that is that it leaves the buyer without a real business case, and so leaves the vendor without the revenue benefit of AI.
