To The Moon

For those of us looking for signs of an AI bubble popping, there are no signs. Earnings coming out of chip, energy, and infrastructure stocks are incredible right now. As trillions of dollars are poured into the AI buildout, there seems to be no end to the exponential growth for the companies involved.

It’s not just $NVIDIA(NVDA)$ that’s seeing growth and guiding for 🚀 demand. It’s $Intel(INTC)$ $Advanced Micro Devices(AMD)$ $Bloom Energy Corp(BE)$ $Cerebras Systems(CBRS)$ to name a few.

Anyone who can make a chip or build some energy is going to the moon.

And that exponential growth has me worried.

In a world where demand for all things AI is growing exponentially, and supply is growing exponentially, a supply/demand imbalance can flip from undersupplied to oversupplied in a heartbeat.

And as they plan the next trillion dollars of spending, the companies involved are assuming they know what demand and pricing will look like years in the future.

Newsflash: They don’t!

If they did, Anthropic wouldn’t be scrambling to get compute to serve a business that’s 10x’d in size in a few months.

For now, the industry doesn’t have enough compute to meet demand, and they’re rushing to fill the demand until someone else does. But when multiple exponential growth curves are involved, the imbalance can change quickly. And when it does, pricing power evaporates, margins collapse, growth ends, and stock prices plunge.

If you want to know why I’m not participating in the AI hype, outside of owning $Alphabet(GOOG)$ , this is it. For better, or worse.

Let me explain.

The Exponentials in AI Supply

The AI buildout is being driven by exponential growth on multiple fronts. The three I’m going to focus on are:

• CapEx

• Efficiency

• # of Players

When we look at the tokens produced, which is the ultimate supply of the AI buildout, these factors are multiplied together.

CapEx

The first exponential that’s easy enough to see is capex growth.

Alphabet $GOOG ( ▼ 1.07% ) is a perfect example of one of the companies increasing spending. In 2026, the company expects to spend between $180 and $190 billion on capex, doubling last year's spending.

And Alphabet is hardly the only one. Here’s a look at the capex growth for the hyperscalers and big neoclouds.

What’s hard to convey with these numbers is how much of this spending is translating to increased compute. The building, power plant, roads, cooling, etc., all need to be built, and the money needs to be spent before a single CPU or GPU can be turned on.

As much as capex has been the topic of the market, there’s a lag between capex and exponential growth in token production. So, even if spending growth stops tomorrow, token production could keep growing exponentially for several quarters, or more.

In the example below, I’m going to assume capex will about double each year, which is what we’ve seen recently. It sounds crazy, but this is all crazy.

Output Efficiency

Efficiency in computing is hard to talk about because it’s not a single number. But we know systems are getting more efficient, and so are models.

Chips are also getting more efficient. For example, NVIDIA’s Blackwell is 5x more efficient than Hopper for inference.

Models are doing more with the same chips.

Chips are being built for specific models and use cases.

For the purposes of this discussion, I’m going to assume the token efficiency of systems goes up 2x each year. To be honest, that may understate the efficiency gains in AI today.

Number of Players

In 2023, the four hyperscalers accounted for nearly all of the capex in AI.

This year, we’re going to see Tesla spend $25 billion, Coreweave will spend up to $35 billion, and Nebius will spend up to $25 billion.

Just based on the growth in competition recently, I’m going to assume the number of players will also double each year.

As you can see above, it’s not just a few companies growing their spend and improving efficiency. Everyone is doing both of those things, and new players are coming into the field. Even NVIDIA $NVDA ( ▼ 1.9% ) is spending billions to both fund customers and build its own AI compute capacity.

Everyone, it seems, wants in on the AI buildout game and is throwing as much cash into it as they can.

All of this is very rational. I want to stress this. If Coreweave’s stock is going up like a rocket and backlog is exploding, it makes sense that Iren and Microsoft, and everyone else, want to capture that same growth. Managers aren’t paid to sit on their hands and let competition eat their lunch.

They need to react! They need to spend!

And this is why we see super-exponential growth in token supply.

It’s also exactly how bubbles form. Everyone behaves rationally to build a collective bubble…

Exponential Supply vs Exponential Demand

Stick with me as I walk through how supply/demand can go from excess demand to excess supply very quickly. And I promise these numbers are realistic!

The first chart shows exponential (2x) growth in performance, capex, and the number of competitors. I’ve started at a different starting point for each, so the lines don’t overlap. I know in reality these factors aren’t all equal, but for simplicity, this is how I built the example. The y-axis is just a proxy for tokens of output.

Remember, the exponential growth for each of these components are multiplied together to get the total token output, resulting in a super exponential growth curve of 8x per year in tokens supplied.

Exponentials this large are so hard to grasp that it almost looks like the supply is at zero until 2027. But that’s how impactful an 8x growth rate is. I’ll get to why this is a realistic growth rate in a moment.

Now, let’s compare that to demand growth. Below, the supply of tokens starts at 200 units in 2024, and demand is 300 units. The industry is undersupplied by 50%!

You can see that by the start of 2026, the market is still undersupplied if demand is growing 7x per year. But we’re already oversupplied if demand only grows 6x or less per year.

Demand is so hard to project because we don’t know what AI will be used for next month, much less next year. Growth in 2026 has been driven by AI coding that’s gotten significantly better and harnesses that allow a single engineer to run many agents at once. That drives an explosion in token demand, and keep in mind these early adoptors have gone from 0% adoption to ~100% adoption of AI in coding in a flash. That same adoption rate and exponential usage rate aren’t likely to repeat in other industries.

But I digress.

Extend these demand growth rates of 5x to 7x through the end of the decade, and the market is oversupplied with tokens by 2027 in all three scenarios. And if we extend these exponential growth curves to 2030, the market would be oversupplied by 50% if growth in demand is “only” 7x per year.

I mentioned the lag between capex and token supply coming online above. Here’s how you can see the impact of that dynamic. If demand continues growing 10x per year and supply only grows 8x, of course, there’s no problem the for the AI buildout. But if 8x growth in supply is set in stone for the next two years and demand only grows 5x — which is still a CRAZY growth rate — there will be problems.

Matching exponentials this large is hard!

Maybe you’re wondering if 5-8x growth per year is feasible on either the supply or demand side. To that, I present Google’s tokens processed 7x-ing over the past year. So, my 8x number has some precedent in modern AI.

This is why this buildout is so crazy. We have super-exponential growth in the supply of tokens, and the companies spending hundreds of billions of dollars to provide more supply are trying to project what demand will look like in 12, 24, 36 months, or longer.

Get that answer wrong, and you’re either losing market share or going bankrupt! Neither is great!

Pricing & Demand Curves

You may be wondering why this matters. Won’t demand just fill the available supply like it always does in tech, a concept known as Jevons Paradox.

In economics, the Jevons paradox, or Jevons effect, is said to occur when technological improvements that increase the efficiency of a resource's use lead to a rise, rather than a fall, in total consumption of that resource.

Definition of Jevons paradox

Demand may, in fact, fill supply. But at what price?

Remember, a data center and associated power can cost tens of billions of dollars to build. Once they’re built, the economic incentive is to keep them running at 100% capacity. The upfront buildout costs are high, but the marginal price of compute is low.

These are the economic incentives of a commodity, which is how I think we need to start thinking about compute.

A 90% reduction in token prices could be devastating, particularly for companies building these assets with debt. We see this in memory every few years. There’s high demand, and supply is built. That leads to oversupply and a crash in prices, resulting in negative gross margins (operating margins are even worse).

But this also assumes tokens are price sensitive and demand is elastic (changes with price). Is demand for tokens more like memory (elastic and goes down as prices go up) or oil (inelastic because people will buy at any price)?

Logically, if we’re actually in an undersupplied market, the model makers and cloud providers should be raising prices! This would be the right move whether demand is elastic or inelastic.

And it turns out…they are!

Google’s new Gemini Flash 3.5 is 5.5x more expensive than Flash 3. OpenAI and Anthropic have also raised prices quietly.

Yes, the AI companies are getting profitable because they’re raising prices. But we are also starting to see signs that companies are starting to scrutinize the higher cost of compute.

Jensen Huang said an engineer should spend their salary in tokens. But should they spend 4x their salary in tokens?

Uber used its token budget in the first few months of the year. Microsoft is reportedly scrutinizing token spend internally. Demand has been exponential, but will that be the case if prices also go up?

Is there a “normal” demand response to token prices in the AI world?

I think the increase in prices indicates we’re heading to a more economically rational pricing structure. And that’s not the only sign of rationality.

This economic reality recently hit Elon Musk’s SpaceX. Colossus 1 — the massive Memphis data center everyone was touting as showing Musk’s prowess in building things — was recently rented out to Anthropic because it was reportedly only 11% utilized. Anthropic is paying a pretty penny, $1.25 billion per month, for the compute, but can also jump ship with 90 days notice.

Everything is up and to the right until it isn’t.

Will a sharp increase in token prices lead to a kink in exponential demand growth?

When super-exponentials are involved, no one can project how fast demand will grow. Even Elon Musk…

The Debt Risk Is Looming

You may be saying, “This is all fine because the buildout is being funded with cash flow from these enormous businesses.”

That WAS the case.

But in the past few quarters, we’ve seen everyone start to take on debt to fund the buildout. Alphabet, Microsoft, and Tesla (a different story but I included them in these charts because they’re buying more compute than many neoclouds) still have net cash on the balance sheet, but that will likely change in the next 12-18 months.

The next leg of the AI buildout will be fueled by debt. And the purchase orders are already in!

Coreweave, Nebius, and Iren are three of the smaller players who need to use debt to even try to keep up.

But $Amazon.com(AMZN)$ now also has net debt and is planning to spend $200 billion on capex this year, well over the $140 billion it had in operating cash flow last year.

Only Alphabet, and maybe Microsoft, can keep up this level of spending and hope to remain free cash flow positive. That’s one reason Alphabet is my only investment directly tied to the AI buildout and my biggest position in the Asymmetric Portfolio.

While it’s easy to see the growth of AI and semiconductor companies as a trade to follow today, we also need to understand the underlying economics behind the buildout. AI related compute is growing exponentially, and that’s a great thing for investors, until it isn’t.

For now, AI is a race.

It’s a land grab.

Who can build the fastest?

Who can sign the most deals?

Who can get the most power? The most chips?

But it won’t be that way forever.

Eventually, the exponential supply curve and the exponential demand curve will cross. We will go from supply-constrained to an abundant supply. And when supply is abundant, and there are many players involved in a commodity like tokens, we’ll see pricing collapse.

It happened with railroads.

It happened with fiber (1990s).

It happened with condos (2008).

It’ll happen again.

The question isn’t if, but when. And who will survive?

Maybe I’m wrong. But with exponentials, I get worried.

😍 Been eyeing Tiger merch but short on Tiger Coins? Now's your chance.

🎁 We’ve selected 4 high-demand items across practial, lifestyle, and learning, now with a lower redemption threshold!

Hot Merch Returns · Up to 43% Off