Cycles Update: AI Waning and Auto Waxing
I thought I'd have an update on the cycles, and why I think one is much better positioned than the other.
After last week’s note, you can tell I’m decidedly more bearish on the economy and in general. It’s not just me. Remember that uncertainty is a synonym for volatility in the economy. And as we’ve seen bearishness, we’ve witnessed increasing volatility. I expect this to continue.
Moreover, I must make a pretty simple observation about cycles, especially the AI-driven ones. I wrote in capital cycles and AI what we were missing for a traditional bubble, and we’ve started to see some of these happen—specifically the announcement impact and a few large IPOs.
I am not ready to call it all over, but like everyone, I am watching the economic uncertainty. I want to use a historical analogy to explain how I could see the AI trade being “over.” That's why history doesn’t match, but it often rhymes.
Double Ordering and AI
In semiconductors specifically, I see a pervasive pattern over and over when it comes to cycles. Some new demand or supply shock happens, and then a shortage occurs. In the recent past, we can point to automotive semiconductors during 2020 as a prime example.
The initial supply shock was COVID-driven, and then the quick rebound led to almost no available car capacity because newer generations of cars required more semiconductors. Automotive companies started to take note - and then began to double or triple orders to match supply to the strong demand they were seeing (as well as lack of inventory).
At first, this is a healthy order flow. Then, double ordering. Then, orders were so large that their suppliers asked their customers, “Surely you cannot be serious.” So suppliers created the fad of NCNRs (non-cancelable and non-refundable) orders so lagging edge foundries would be sure that the added capacity would have demand. The entire time automotive inventory was in shortage, as most cars missed a “golden screw” that would often dictate total supply availability by the availability of one part.
And then the cycle happens. It almost always occurs as supply ramps, and as supply reacts to demand, the shortage becomes a glut. All those double orders and supply build-ups happen, and it just takes a slight kink in demand from macroeconomic weakness or somewhere else, and all that built-up inventory becomes a glut as the golden screw is released. In the case of the last cycle, inflation and over-ordering were involved.
I like the phrasing of Ganesh Moorty, the outgoing CEO of Microchip. It’s funny because they messed up the cycle the most last round, so this hits a bit harder.
And just as when we were on the upside, it looked like it was impossible to know how could we ever satisfy the demand. On the down cycle, you get the sense that how can we ever turn this thing around. But it does happen. It is a cyclical industry, and it will happen.
And what I want to highlight today is that we are at the" “how could we ever satisfy demand phase” in AI. It rhymes heavily. This time the golden screw is not an 8-bit MCU but a CoWoS supply from TSMC.
Cycles often end during supply ramps. Today’s obvious supply ramp is the GB200, and the problem is it’s clear that the rest of the year’s demand is accounted for, but as supply ramps meaningfully, will we see softness on the other side? That is almost literally how every cycle ends, and often, the stocks reflect that reality before the actual financials turn. A perfect example was how TSM revenue skyrocketed during 2022, but shares experienced a 50% drawdown. It wasn’t until 2023 that revenue started to decline, and then shares recovered into drawdowns.
My concern is that the same dynamic could be in play. This time, the shortage is in GB200 racks, and as supply ramps continue, the market looks forward to the following year. And I believe it will be hard to grow revenue if there is a recession. That recession case is being priced quickly into capital markets.
So now, let’s turn to where the actual gas (capital) is coming from for the AI capital cycle build, as that’s also an important consideration.
Sources of Incremental Capital Requires Fundraising
In my view, only a few large investors are currently investing heavily in AI. The first and safest are the hyperscalers.
Doing some very simplistic math, let’s assume ~250 billion in chip spending in 2025 and 200 billion in total data center spending happens in ~2025. That’s 100% of cash from operations from the big four hyperscalers. So, the AI trade cannot rely on this bucket alone.
Sovereign nations are the second interesting bucket and are becoming a more significant part of the pie. France just announced a ~112 billion dollar deal over multiple years. Ironically, this is not as much spending as a single company does, but it is a lovely second bucket. I think this has a chance to be less economically sensitive, but time will tell. There’s minimal spending here, and this could change quickly.
Last but not least are venture capital and private equity. Venture capital tends to invest in what will drive AI chip demand, while Private Equity invests in what supports datacenter spending. I will argue that this is a cyclical bucket. We do not have a long history of Venture capital, but one comment is that “dry powder” is evident everywhere. I’m a bit more uncertain than that. Since Venture capital, in particular, is very young, it has been increasing secularly. However, in 2009, there was a constraint when liquidity went down.
It’s more pronounced in Private equity.
Liquidity in markets freezes during downturns. Think of the economy as 1000s of little bits of extension of credit or capital slowing down simultaneously. This is the velocity of money, which slows down doubly for long-term investments.
But that’s another good observation. As economies decelerate, rates will go down, so liquidity should rise. This is how economic cycles come out of bottoms, but this time, the rate decreases to spur investments when capital is scarce. It’s a kickstart to encourage spending.
It’s all a cycle from spending to the economy itself. And today I wanted to write what a pre-mortem of the AI cycle could look like. But I want to be clear - I do not know the outcome. I want to repeat this because I cannot emphasize this enough: I do not see the result of the macroeconomy. Does the economy soften, have rate decreases, and then reaccelerate? Does re-shoring mean a new, higher leg for the AI trade? Maybe.
No one knows. Remember, macroeconomic opinions are like assholes in that we all have them, and they all stink. But while the entire world stares and obsesses over minor details in the biggest capital cycle so far of our lives, a much smaller cycle has turned.
AI isn’t the only cycle around. While all eyes (and I mean every spare eyeball) in the semiconductor world are on AI, I am starting to see positive cyclical indicators elsewhere. I’ll discuss this behind the paywall.