Epochal change is everywhere.
Have you noticed that? There’s a bull market in tectonic shifts.
That’s tongue-in-cheek, but now that I think about it, it makes intuitive sense. Watershed moments tend to beget more paradigmatic change.
That’s a kind of paradox in that definitionally speaking, watersheds are singular events. At the same time — i.e., despite being in some sense oxymoronic — it’s almost a truism.
Put as a question: What would it even mean for “epochal change” to happen in a vacuum? When a critical levee breaks, there are downstream consequences every bit as dramatic, and in many cases even more so, than the initial event.
Nagging ROI questions notwithstanding, it’s safe to say AI’s an epoch. (There’s a machine learning joke in there. Don’t miss it.) This is not a drill. Maybe it’s not the wheel or the internal combustion engine or the PC, but AI’s not the metaverse.
And while it’s too early to speculate on the full spectrum of downstream consequences, we’ve already witnessed one knock-on event of historic import: An overnight transformation among the most important companies in the history of capitalism from lean, mean cash machines, to debt-laden, asset-heavy businesses.
I realize this is well-worn territory by now, but it’s kinda hard to resist the temptation to harp on it given the rapidity and magnitude of the shift.
The figure above, from Goldman, gives you some context. Among the hyper-scalers, revenue-to-assets exploded from 55% to nearly 80% in the half-dozen years leading up to the launch of ChatGPT. Then, everything changed.
“The physical footprint required to support AI workloads has led to a steady decline in asset turnover, which analysts expect will continue in 2027,” Goldman’s Ben Snider wrote, adding that “the capex boom boosting semiconductor earnings today will weigh on hyper-scaler ROE going forward as depreciation expenses grow.”
Note from the chart that depreciation and amortization as a share of sales is expected to double by 2028 from pre-ChatGPT levels.
You’re reminded that this year’s capex plans call for more than $750 billion in outlays across the hyper-scalers. That figure, Snider remarked, is equivalent to 100% of free cash flow. Hence the need to ramp up borrowing and, more recently, raise equity to avoid taking on too much debt.
The figure on the left, below, shows you the inflection in hyper-scaler leverage beginning, again, with the launch of ChatGPT.
As Snider went on to write, “share counts have begun to rise” too, as illustrated on the right.
The prospect that between AI IPOs and hyper-scaler follow-ons, net equity supply may turn positive next year is a new bogeyman for market participants.
If the overall supply of US public equity does indeed stop contracting as a result of the hyper-scaler arms race, that’ll be yet another example of how watershed moments tend to multiple and propagate.
“If the overall supply of US public equity does indeed stop contracting as a result of the hyper-scaler arms race, that’ll be yet another example of how watershed moments tend to multiple and propagate.”
Great point. And don’t forget the impact on the most sacred bull market propellant – share buybacks. Buybacks are not limited to mega-techs, so folks like JL and others here are probably running screens to reveal the largest non-tech companies buying back their own shares. Those shares should rightfully be given a premium.
Buybacks are part and parcel of that float equation. So when I say “as a result of the hyper-scaler arms race,” that’s largely what I mean.
But as discussed here last week, the interesting thing about this is that it’s now hitting from both sides. Hyper-scaler capex is siphoning cash that might otherwise go to buybacks (i.e., reduced demand), and then when they try to preserve some of that cash by funding capex with debt, they end up uncomfortable with the leverage and resort to equity raises (i.e., increased supply).
From the Weekly: “The irony’s almost too much to bear: ‘We have to borrow to fund some of this spending so there’s money left over for buybacks. Sh-t, now we’ve borrowed too much. We need to think about a secondary.'”
Markets don’t care…onwards and upwards it would seem.
How long is the useful life of the AI assets in the depreciation schedule? Is it accelerated for a tax break, or realistic? Does anybody know? It’s non-cash, right?
I’m from the period before AI, and I’m telling you the golden age is the 1920’s.
Modified- but the intent of the line remains. 🙂
Trump promised to bring back steel and make it great again, and he finally delivers on a promise. Except it turns out he’s brought steelmaking, albeit just their financial profiles, to the MAG 7/10/20 — heavy, lumpy CapEx, debt, dilution.
Data- and compute-as-a-service may not turn out to be as profitable as current expenditures might lead you to believe. Breakthroughs in steelmaking technology and scale tend to be generational, whereas for the hyperscalers, it’s built in to the guts of what they are installing now, which will be obsolete in 2-3 years, perhaps before it is even fully utilized. What if this arms race isn’t a one-off launch, and is merely presaging a world where, sure, you spent $100 billion last year on some impressive technology., for THEN? But the new NVIDIA GPUs or Seagate drives or whatever change the competitive economics and you are just chasing your more efficient and larger hot-rolling mill?
The agile, capital-light and frictionless expansion model of the MAG 7/10/20 is being put to the test by these massive investments. Cet ready for monthly reports on data center and AI model capacity utilization, not to mention more boom-bust markets which the “as-a-service” subscription model thought it was making obsolete. We just invented a couple new commodities — data and compute. No reason for them not to be immune from mispricing due to “sentiment” and “expectations.”
Furious One, you have just written some great stuff. I liked your concise explanation of the capex deduction issue. You make a strong point when you remind us that the accelerated depreciation write-off is especially compelling when you have record profits.
“We just invented a couple new commodities — data and compute. No reason for them not to be immune from mispricing due to “sentiment” and “expectations.””
I wonder. At first glance that makes sense, but second thought is that the larger datacenter operators as well as the major takers don’t have a lot of incentive to voluntarily share that data?
Oligopolists may be less averse to sharing data than you’d think. The rub is often what can be considered collusive or anticompetitive in terms of the FTC and other regulators. There are different rules of thumb, but a widely used one is that if one company has a 60% share and any two companies have a combined 80% share, then publishing the aggregate number can be problematic, but I admit that maybe old school now.
The parallel I drew between asset-heavy steelmaking and formerly asset-light data and compute is already showing up in the latter’s free cash flow as H has emphasized repeatedly. I hinted at another parallel — the fact that it is difficult to profitably make steel without high capacity utilization. But it’s also difficult if you don’t have modern (up to date) capital. It remains to be seen how long it takes for the data and compute companies to earn their paybacks and how long their current giant investments will remain in service before new generations of equipment and components force new rounds of lumpy expenditures for faster, smaller, less energy, etc. The risk of having outdated and inefficient capital seems far greater for data and compute than for steelmaking — for sure re the time frames, except for true technological breakthroughs in the latter (blast furnaces vs electric). And while capital expenditures by steelmakers are emblematic of being “lumpy,” they pale in comparison by dollar amount to this AI boom which is unprecedented in pretty much all respects.
And finally, an aside. The MAG7/10/20 companies helped put a lot more energy into the GAAP/non-GAAP debate, in favor of the latter, when compensation expense got moved off the income statement. Even though it was a ridiculous distortion that also led to multibillion dollar pay days for many execs, you were an idiot if you didn’t invest in these same companies — they almost all killed. There was no comeuppance. They had plenty of operating cash, and salary expense saved could be magically directed towards buybacks instead, even if at market values and ultiples of the granted stock exercise. All the market cared about was the stock was at ATHs AND the company was still buying it back.
I don’t know enough other than to just generally expect things, but accelerated depreciation treatment of a generationally significant investment boom is going to make the GAAP/non-GAAP chasm wider like you’ve never believed or thought possible, especially for these critical MAG 7/10/20. As H has documented, there’s also a lot of off-balance sheet, commingled money pockets around to suggest that some are already trying to insulate themselves from potenitally stranded capital “investments/tie-ups” in data centers. To take it one step further backward from steelmaking, we could be looking at a “chicken finishing” op where Purdue fattens its chickens (or Cargill fattens its beef), at shallow-pocketed Maryland Eastern Shore independent “operators” who take a fixed amount per chicken in return for all of the environmental liability. Our AI and storage overlords seem to think maybe that model will work here — you front the costs at the outset, I’ll take the toll and the heat, and if things go tits up, it’s already mostly off-loaded to an off-balance sheet hiolding company with no real assets.