Wall Street’s good at finding reasons to stay bullish.
Note: Being good at coming up with excuses isn’t the same thing as conjuring good excuses. Not all excuses are plausible.
When US equities shuddered late last month on news that a Chinese AI startup might’ve created a viable ChatGPT competitor for a few lousy million, and without access to the best chips, Wall Street strategists quickly found a silver lining.
Cheaper AI, the story goes, will invariably be a net positive. If AI is indeed the key to unlocking heretofore unimaginable productivity gains, then the more accessible the technology, the better.
Think of it this way: The personal computer wouldn’t be much use if PCs were $250,000 — maybe they’re miracle machines, but if nobody could afford them, that’d be irrelevant.
I actually don’t think the comparison works. Anybody and everybody already has access to AI tools, and they aren’t expensive. Admittedly, I’m not using them at any sort of scale and I’m certainly not paying for any enterprise-level offerings, but if all you want to do is draw bulls and bears fighting, OpenAI will take care of that for you, and it’ll only cost about $20/month.
And yet, according to US government data, almost no one’s using AI in their businesses currently, at least in aggregate. The figure on the right, below, suggests just 6% of US companies were actively participating in Jensen Huang’s “new industrial revolution” as of last year.
The figure on the left is a kind of CliffsNotes version of the DeepSeek story — it gives you some perspective on the company’s claims vis-à-vis performance and cost.
Plainly, some sectors and industries will be slower to adopt AI than others by the very nature of their businesses, but it’s fair to suggest the hyper-scalers need more in the way of economy-wide buy-in than what’s illustrated on the right, above, to justify enormous AI outlays.
If the only way to get that buy-in is for costs to come down dramatically, the hyper-scalers are staring at a Catch-22: They need economy-wide adoption to justify massive AI expenditures, but that adoption requires the advent of cheaper AI, which would itself call into question the relative wisdom of the very same expenditures.
Either way, it seems increasingly likely that the hyper-scalers will face a reckoning. They overspent on two assumptions: 1) AI adoption will be rapid and widespread, and 2) AI innovation’s everywhere and always a function of money spent. Both of those assumptions look shaky to me, or anyway shakier than they did six months ago.
If you need to extract a bull case, Goldman’s got you covered. “Some clients believe the cost efficiency of the DeepSeek model bodes poorly for the likely returns of US tech investment spending [which] in turn could jeopardize the exceptionalism of US growth,” the bank’s David Kostin said, before quickly noting that the bank’s economists believe “the DeepSeek development represents a clear positive for overall US economic growth [as] increased AI adoption [may] lead to productivity gains [which] should also be a positive for US corporate earnings.”
That leaves unanswered the question of hyper-scaler capex. Was it worth it? And will future massive outlays tipped by the mega-caps this reporting season be worth it? My guess, actually, is that it was worth it, and will continue to be worth it. I don’t think DeepSeek is what DeepSeek claims to be, not entirely, and I think most of the cash America’s tech champions are pouring into AI will prove to be money well spent.
However, between now and whenever hindsight allows us to pass judgement definitively, I’d expect a few (and maybe many) more “DeepSeek” moments, and although it’s not my “base case,” I wouldn’t be shocked if, 10 or 15 years from now, a running joke among “older” market participants asks: “Remember Nvidia?”
Read this very topical opinion piece this morning on MW. It reminded me of a prescient comment here by John Taylor which he posted before the DeepSeek announcement where he contended that the US tech companies are the only ones on earth totally focused on rolling out every larger models. His comment was salient because each new LLM model iteration was already producing fewer and fewer gains per dollar wasted, oh I meant spent.
https://www.marketwatch.com/story/deepseek-could-sink-big-techs-ai-growth-plans-and-why-shouldnt-it-1f6a887b
But the sunk costs at the hyperscalers are too enormous to allow them to admit they were misguided. So they focus their energy on questioning the validity of DeepSeek instead of answering why they didn’t come out with something similar.
Of course, we Americans firmly that big is better in all things. Think of your neighbor who just bought a black Dodge Ram 2500 Rebel to use for his weekly dump runs and monthly trips to Home Depot. Or the fascination with the 44 magnum handgun thanks to Dirty Harry. Hell, it’s the BIG Mac, not the Mac, that competes with the Whopper.
So why shouldn’t we lavish endless money buildking the largest LLMs on the planet? BIG is better!
My software company is using AI a lot, and it certainly makes the competent folks much more productive. It’s making my life easier every day in new ways. Though measuring productivity in aggregate for the industry is basically impossible imo
Hard to extrrapolate the impact of AI on wider productivity given that the US tech sector, which includes software, computing, data storage, and more, only accounted for around 8.9% of the US GDP in 2023. Some estimate that software accounted for about 25% of that. Not sure how much of that came from software coding and such as opposed to marketing and ongoing support.
Your point about expense is valid, for a single consumer who occasionally leverages the tooling. For example, I played around with using AI to help author a novel. It would write a few chapters and then claim it needed time to think, when I asked how long it needed, it said unspecific hours. Obviously AI doesn’t need hours to do what I was asking but this is a way of implementing throttling at the consumer level.
At scale and, depending on which model, the expense increases dramatically. If you expect a specific class of service at scale you have to pay an additional up front cost called PTU to ensure a given level of throughput. Again, depending on usage, AI at scale, has the potential to become a FinOps nightmare for firms.
From that perspective, I tend to agree with the assertion that lower cost foundation models are good for business. Everyone thinks they need the latest and greatest OpenAI model but when the expense starts piling up, executives will start to question that. If you can accomplish the same task on say Amazon Titan, why wouldn’t you?
I have two quick questions. Do you intend to put your name on that novel? Will you get royalties?
AI may be neat and all but most of it was trained using data owned by one or more of us. AI is mostly valuable because it has been created with our sweat and tears and never paid for. My HIPPA data is used to train health care AI machines and programs. I paid my doctor to create and record that data and my insurance company took it, sold to others and most of it I can’t even access. If I don’t sign an order to allow that process, the provider won’t allow my physician to serve me. (I know, I tried it). That should be illegal. This goes on everywhere. Credit card data is routinely sold. When lived in Iowa the state sold all my personal info to anyone who offered to pay for, while I had to pay them to take from me.
AI is based on pure statistics. All statistics are subject to measurable and predictable errors. There are two types of statistical errors and at least one must always be non-zero. When one goes down, the other must rise. Once AI is released to do a task for us it is automatically subject to its inherent error rates, the levels of which you will never be allowed to know (classified secret). We tend to think something that is programed into AI its going to be great, no errors, blah, blah. Not at all true. Many commonly used new AI degrades over time (statistics behaves that way). Think about it before to go bed tonite. Do any of you want to be operated on by someone using a surgical robot operated by AI you know nothing about? I have been so treated and the machine made a mistake I was fortunate a human was able to eventually rectify. What about life-changing policies arising in similar “black boxes?”