My old Latin teacher would love my quoting the language at times, so to make him (posthumously) happy, Quis in periculo est? Who is at risk, roughly translated, and of course I’m talking about AI. Wall Street, as I’ve noted, vacillates between condemning it and betting on it, and sometimes both at the same time. We hear from companies like Cisco that it’s transforming networking, but at the same time enterprises and some other vendors report that AI infrastructure is largely confined to hyperscalers. This has to shake out, and that raises my (in Latin) question.
My enterprise contacts’ comments show there are two basic models for an AI business case. The first is the business analytics model and the second is the worker process model. The first model is data-driven, closely coupled to company core databases. The second is directed at assisting workers in specific tasks, and is coupled more to the task and the required results than it is to data. There’s obviously a bit of slop in the separation of these models, which we’ll get to.
Business analytics business cases depend on the results of data analysis of data that is normally subject to governance, which means that the same constraints that are applied to cloud-hosting application components that use the data would be applied to cloud-hosted AI using the data. Nearly all the AI self-hosting interest evolves from this model, but it doesn’t get a lot of publicity because most vendors have found it time-consuming to develop this type of AI application. IBM has been the only vendor to really leverage this model, in fact.
The worker process model directs AI at the production of something, the completion of a task. This is often referred to as a “copilot” approach, because in it AI acts as a kind of partner to a worker, helping with things that are “routine” (things the worker is too high-valued to be doing), presentation-related (writing, image production, slides, etc.), or specialized (review, analysis). You can see from these points that “data” is often used in the completion of the tasks, which is the largest source of overlap between the two model definitions. However, most of the AI missions enterprises assign to this model do not expose governance requirements, and so do not favor self-hosting of AI.
You can also divide AI by the delivery method, which enterprises say either capitalized infrastructure or as-a-service. The latter means “expensed”, which means that formal AI project control is normally not applied to the project at all; many line managers, even at the lowest level, can approve as-a-service AI. However, as noted above, the business analytics model of AI is constrained in as-a-service delivery by the governance limits set for any data it depends on. In some cases, enterprises say they have “interdicted” departmental AI applications that have shared governed data inappropriately in as-a-service implementations, but most enterprises say that it’s likely that some violations of data security policy have managed to sneak past enforcement, just as that’s happened in SaaS-cloud services.
You may wonder how this relates to the who-is-at-risk question, so let’s hone in on that now. The question for AI is the extent to which the combined business cases that can be made for the two models can create spending sufficient to fuel continued AI advance, both overall and for each of the relevant market segments/players. The range of potential business cases, their credibility, their relationship to the players in the AI space, and the extent to which AI influences company profits all combine to create the risk profile we’re looking for.
Enterprise information I’ve collected suggest that the as-a-service elements of AI opportunity cannot, by themselves, sustain the current levels of AI investment. Data on this sort of thing is very rough, but what I hear suggests to me that current AI spending does not quite offer a reasonable ROI on investment, so the as-a-service providers (the Big Three in particular; Amazon, Google, and Microsoft) are investing as much on hope as on logic. However, there is an indication that the economic uncertainty associated with the Iran war may be delaying capital projects, which has benefited cloud spending and covered these players with Wall Street.
That doesn’t seem as likely in 2027. Enterprises are divided on whether there’s any chance that as-a-service AI will actually prove out, but almost 60% suggest that they’ll control the cost of AI next year than expand it. That’s highly speculative, of course, but it’s an indication that risk is rising for the Big Three. Of that group, Microsoft is most trusted by enterprises and its office suite provides a natural on-ramp for advanced AI features. Google is next in trust, and its specialized research and presentation tools are also valued. Amazon has the greatest risk of this group, but Oracle and most of all, Meta, face a much greater risk in 2027. So do both OpenAI and Anthropic. There is little they can really do to offset any Wall Street skepticism, and so they are vulnerable to any bad news in the AI sector.
Of the chip vendors, it’s pretty clear that Nvidia faces the greatest risk, since their financials rely on massive infrastructure spending that isn’t likely to be sustained if the Street loses enthusiasm for AI, as it had in the spring of this year, and doesn’t regain it. The volatility of the sector is a good indication that the Street is balancing its love for bubbles with its fear of a tragic shift in market sentiment. Broadcom and AMD have some exposure here, but not as much.
What would save the whole AI space is a set of real business cases. I’ve blogged regularly on my view that only self-hosted applications in general, and real-time process control in particular, can really change the game for AI, and in fact for the tech space overall. I won’t repeat the arguments here, but smaller models and chips are the most likely beneficiaries of this new and as-yet-unvalidated wave, so the AI giants already named can’t expect much from it.
It’s also theoretically possible that enterprises would accept cloud-hosting of AI for applications involving their business-critical data, but right now enterprises tell me that they are even more wary of AI governance issues than cloud governance issues. “Could my AI provider be training his models, and my competitors, on my practices? No thanks!” one said.
This adds up to AI being a high-risk space, all the more so because it’s not replete with accurate financial and technical stories online. Educated buyers aren’t usually what sellers want, but in the case of AI we may need education to get things on their optimum track.
