Even as the popularity of AI grows, and growth enthusiasts on Wall Street whet their risk appetites on AI startups and market hopefuls, we’re getting increasingly clear warnings about the AI business case. Some are coming from Wall Street itself, some from thoughtful analysts, and even some from enterprises whose spending on AI would almost surely have to drive any real ROI in the future. What are people saying, thinking, about AI’s business case?
In the recent quarter, Wall Street pointed out that both Microsoft and Oracle talked about significant capex increases associated with AI. Generative AI is a major resource hog, in training and also in operation. Literally hundreds of researchers are launching initiatives that aim to make it more consumable, but so far the majority of generative AI users are using it only where it’s free. It doesn’t take much of a business case to justify something that has zero cost.
Microsoft, according to the Street, invested billions in OpenAI (the ChatGPT company) and spent over a billion to create the supercluster system that runs OpenAI’s various versions and models. They continue to buy NVIDIA GPUs but literally can’t get enough, and Microsoft’s capex explosion is a sharp contrast to the roughly flat capex from Amazon and Google. For all of this, Microsoft’s proposed application of AI technology seems focused on simply enhancing the features, and presumed stickiness, of things like Windows and Office 365. Google, who the Street estimates is spending between three and four billion a year on AI research, seems to have even hazier notions of how it’s going to recover the cost and earn any respectable return.
Of course, generative AI isn’t the only use of GPUs, but ironically it may be starving other more readily exploited AI technologies for resources. Of the roughly three-dozen enterprises who tell me they’re trying to get capacity for cloud-hosted AI applications, only three claim to have been able to fulfill their needs. Even these three admit that they’ve not yet proven out the technology.
What may be even more interesting is the rapid shift I’m seeing in the attitude of what I’ll call “non-aligned AI pundits”, people with expertise who aren’t actively trying to promote an AI startup or working for an AI company. In February, this group was totally enthusiastic about AI potential, with almost all of them forecasting that within six months, the business case for AI would be clear. In July, the majority were saying that it would “take a year or more” for the technology to pay back overall for users. They still believed that cloud providers, chip vendors, and software players in AI would see profits sooner.
And the cloud providers? The comments from the cloud providers on their own AI return on investment have stayed cautious, mostly because none of them have any idea of how long their customers will take to build a business case for the AI services they’re consuming eagerly. My non-aligned pundits think that by the beginning of 2024, cloud providers will be pulling back on the pace of AI GPU deployments and pushing for other GPU applications. A couple of people told me that they’ve already seen signs of that in two cloud providers, though they don’t think it will be visible to Wall Street until Q4.
Why all this is happening could perhaps be likened to the way that a forecast of significant snowfall leads to a run on snow blowers, shovels, and deicing compound. Yes, most people think that the news plays up bad weather stories because more people watch. I knew the news director of a local network affiliate, and he told me that they always forecast the upper end of the plausible snowfall because the ratings went up, and if another station out-forecast them, they’d engage in what he called (and contributed the phrase to my own vocabulary forever) a “bulls**t bidding war”. “I see your foot and a half and raise you two inches plus some ice!” To put it more charitably, nobody really knows how big AI could become because it’s changing so fast, and so nobody wants to take the chance that it’s the true Next Big Thing that will make or break a host of different companies.
We do seem to be making some real progress with AI, finding strategies to reduce the number of “hallucinations” or AI-created errors, reduce the resources needed to train the models, reduce the number of parameters needed to produce credible results. But we’re also generating risks of lawsuits arising from “stealing” data for training, problems with the bias created by the fact that the most common source of data, the Internet, contains every possible truth and every possible lie, and the risk that exposure to free AI will forever contaminate the business model. People today, brought up on the thought that the Internet was “free” now expect almost everything associated with it to be offered forever at that (zero) price. Why would AI be different?
I don’t think we can say confidently that AI will make a broad-scale, massive, business case. We can’t really say with much conviction that it will do any better than ATM or 5G’s network slicing. The former was overtaken by events (the Internet and plummeting cost of bandwidth) and the latter was a solution in search of a valuable problem. AI could end up falling prey to both.
With regard to events, I think it’s important to note that the AI buzz we’re seeing is really generative AI buzz, and in truth generative AI may be way more hype than reality. It has enormous resource costs associated with training and using it, and while there are literally announcements every day that claim to add features or reduce burdens, fundamentals for generative AI haven’t changed much at all. It’s gotten a lot of attention because it’s the first AI model to be deployed in a way that lets the broad population interact with it. And, of course, because the ability to generate texts or poems or code or paintings makes it look sort-of-human, so it’s spawned the “will-it-replace-us” or “is-it-coming-for-us” discussions. All these are interesting, entertaining, but they’re not applications that generate real revenue.
Which is what leads to the second point. You need a business case for anything to drive massive market changes. There has to be buyers willing to pay and sellers able to profit. Right now, we’re in the venture stage of generative AI, the stage where everyone is excited except the accountants, who are increasingly frightened. The fact is that there are many valuable AI applications, but those applications use something different from the generative AI models we see, use, and hear about. Many are really machine learning, a few add some neural network processing, and a very small number are based on a less resource-intensive form of deep learning or large language modeling, but trained and used on specialized and even user-specific data. This is going to be big, I think, but it’s not generative AI.
And, friends, I think that the AI pundits themselves believe all of what I’m saying. They recognize that the future of AI is very different from the kind of AI we’re playing with now, but if they can ride that wave of public interest to get venture funding or other capital, they’ll do that in the hope that what they develop will end up realizing some of the real benefits, benefits we’re now hiding in the fog of PR.
