Shade at the Parade

It’s July 4th in Chappell Hill, Texas. The “town” of Chappell Hill is more of a spot on a map than a town. And yet there is community. They throng along the sides of the street each year for the independence day parade.

Our family, huddled together and joined by thousands — a mix of full time residents, weekenders and Houstonians looking for nostalgia — watch as the parade begins. The first in the parade is Jane Smith’s Sebring carrying the unofficial mayor along with a few business owners. Jane’s Sebring will come back shuttling more of the crowd for their turn waving at the sweat drenched attendees.

Between sightings of Jane’s Sebring, attendees get brief periods of shade cast by massive combines, tractors, and commercial agricultural equipment. Chappell Hill, though just an hour from Houston, is rural Texas. You can’t escape the centuries old economic roots.

Around the 4th — thinking back on years of watching the independence day procession — the similarity with the pending AI bubble struck me.

Imagine new milling capacity was announced in Washington County. Word gets around, the demand will be huge for all grains traditionally harvested in the area — much larger than ever before; several orders of magnitude of additional demand. Everyone would build.

Gentlemen ‘ranchers’ would convert latent fields into tillable acreage. New barns, silos, access roads and equipment would be put in service. Nobody is waiting to see the first harvest, they’re depending upon the demand given the expectation of the new mill.

The independence day parade would be longer. Shiny enormous new combines and harvesters would provide hours of constant shade as proud farmers rumbled them by a cooler crowd. The attendees might even cheer — growth is in the air.

This is roughly where we are in the AI cycle right now. Companies building the compute capacity are dedicating nearly a trillion dollars this year — more than twice last year’s record blasting pace. But the harvest is pro-forma, expected but not delivered. The demand they are building for may well show up. But the buildings and infrastructure are going up first.

Those outside the industry feel justified — they’ve used the technology and can attest to its revolutionary potential. Beyond that, they have watched the investment pour in and seen the promised returns. Just like the independence day parade attendees, the investment provides shade from the heat. It feels electric, needed and important and there’s a little something in it for the viewer.

But the part those participants don’t notice is the amount of the investment that is short-lived. The barn, silos and roads will last decades, the tractors and prepared acreage will wear out in a few years. About sixty cents of every dollar spent on AI infrastructure goes into the computer chips, and those chips wear out the way a laptop wears out. And if you haven’t noticed, with the new demands on CPU and memory, they wear out a lot faster than before.

On the company’s books, though, the chips are being consumed slowly. Bought with debt and put in service quick; the asset shows up on their balance sheet. Reasonable for sure. However, there is healthy debate on the actual life of these chips. It bleeds onto the P&L through depreciation over six years — or more.

None of this is necessarily dishonest. But, like all bubbles, a market full of honest participants can paint an altogether dishonest picture of the future. This cost is being spread out over time and the pressure to elongate the useful life of these investments is extreme. If all the CAPEX stopped cold, the wear-and-tear on their books has to roughly quadruple from about $178B per year to about $660B per year against combined total profits of $511B. That’s not a forecast or a warning; that’s second grade arithmetic and the equation ends in a number below zero.

Then there’s a deeper question — who is paying for the barns, tractors and fields? It used to be the farmer’s money — retained earnings from decades of hard work. But that’s not enough now to get ready for the demand. Instead, and increasingly more troublesome, it is borrowed. The share funded by debt went from one dollar in eleven two years ago to one in three today.

A lot of that lending isn’t coming from banks — they know better. It comes from investment funds, which get their money from insurance companies and pension plans, which get their from ordinary people’s retirement accounts. There’s nothing sinister going on here; that’s how investment money works. And in a market driven economy, it flows to the area of greatest perceived need and impact. But, that’s also how bubbles work — when the outside observer willingly cheers on the flow despite the growing risks.

The strangest part of this dynamic that is somewhat new is the circularity — so called. Retained earnings, private investment and debt aren’t enough. The company selling the chips has also invested billions into the companies buying the chips, and has promised to buy back capacity those companies can’t sell, and has now organized a $500B fund to help other people lend against the chips it manufactures.

Picture the tractor dealers lending the farmers their down payments, then co-signing on their loans, then guaranteeing to buy any grain they can’t sell. Every individual step is defensible. Together they mean that if one farmer fails, it isn’t one farmer’s problem — the dealer is on the hook too, and so is everyone who lent on the dealer’s good name.

Meanwhile some of the cost has already left the tech industry entirely. These buildings need more than chips. They need enormous amounts of electricity. The wires and power plants to serve them are paid by everyone on the grid. In the mid-Atlantic, the regional grid’s own watchdog calculated that data center demand added about $9B to one year’s power cost — a 174 percent increase on what it otherwise would have been. Your neighbor who has never heard of any of this is paying for a piece of it on their monthly utility bill.

Texas, which usually is the last place to slow anything down, is flagging caution. It is sitting on a 474 gigawatt request to plug in — more than five times the most electricity the state has ever used at once. They inspected how few of these requests were viable and halted the process in August pending an audit due in December.

So, what does this all add up to? Not a prediction, not a line in the sand, but a major caution flag that the market is doing what it does. Money flows into promise and if that promise shows up later than expected, then bad things happen. This is predictable and ordinary in a market economy where individual actors can choose to act against that promise. The problem, as always, is that money flows easiest to those investing into the excitement.

Reality can meet promises. In fact, reality can overshoot promises. And in the case of new paradigms, like AI, that is more than just possible. For euphoria to overshoot, market participants have to overwhelmingly agree. There is plenty of attention on this looming concern. As a result, it might already be priced in; the bubble might have a relief valve. In fact, the bubble might not be a bubble at all. It might not be enough.

Compute demand is compounding faster than capital expenses by some estimates. And while chips depreciate in usefulness, they don’t in utilization. CPUs and DRAM don’t burn out if well maintained. They don’t get slower. The risk is new improvements make them obsolete. And all of this could mean the market is right on the need. Even the excesses of the 90s paved the way for the fiber layout that powered the early 2000s.

The mill could be coming to Chappell Hill for good reason and it might not be enough. Regardless though, even if demand fails to meet the expectation, the mill, the barns, the roads, and most of the equipment will serve a different purpose. They will find a clearing price and a different crop will move through it all later. Lots of people lost money on the fiber rollout of the 90s so that others could make billions on Google a decade later.

The whole cycle can continue and can even work if revenue keeps growing fast for another six years. At roughly fifty percent revenue growth per year for the hyperscalers, measured against capital spending that is itself still growing thirty percent, the money spent gets earned back around 2032 and the barns were a good idea. At thirty-five percent — it never catches up no matter how the accounting is handled. Everything else is just a detail. The uncomfortable thing about the gap isn’t that it is large. It’s that it’s narrow enough that no one can tell you which side of it we’re on.

We aren’t immune to this cycle. We have real stakes. Our companies all sit under a dark cloud of competition from the very models that promise such potential. In an instant, behavioral norms that form the basis for good product design can change, business processes morph, and value shift to other areas on the human-to-computer value chain. The infrastructure players, in a desperate attempt to drive revenue could demand it from our portfolio. Competitors could emerge instantly from nascent departments of one-time cooperators. The risk is there.

Beyond the usual business risks looms one larger. A founder could do everything right, drive top decile growth, deliver excellent SaaS metrics, and emerge for exit to find multiples compressed and an empty hall where acquirers used to gather. Just like 2001, when a bubble pops, all risk assets take a dive.

Implications

For founders, our primary constituents, the implications are profound. The cost of compute has been dropping recently and may continue to drop as capacity comes online before demand. However, if the demand does catch up, expect to be rate limited by the major models. A big reason so many hyperscalers are investing so much is to have primary access to compute.

Imagine finding an ACV tripling agentic feature in your application. For a while everything looks great. Customers are signing up, revenue is surging, and profits are forming. You’ve also embedded your preferred LLM into your operations. It answers customer tickets, manages your pipeline, prepares your financials, joins your team slack sessions. Efficiency is raging. Quietly AI has embedded itself into everything you do. Imagine how loud an email rate-limiting your use down by 50% immediately would be.

This is not just likely; it’s a near certainty. And it doesn’t have to come from other customers with greater access. It can come from the looming energy crisis. We live in Texas and remember the winter storm Uri all too well. The grid reliability is now an input into whether your company can meet its SLA.

So, what does that mean for you? First, you must build model portability as a first step. This allows you to migrate to a different model provider or incorporate another one as a backup for when service interruption occurs. Second, if your application has customer financial risk associated with downtime, layering on reserved GPU rental can be helpful. This is expensive, but protects your service level.

For LPs, a very important constituent for us, the implications are similarly profound. The AI revolution — or rather the ‘promise of AI revolution’ has already worked itself into the price of risky assets. The S&P 500 would be down marginally this year (YTD) if you strip out the AI boon. And that’s against a 7-10% earnings lift. All risky assets have re-priced and the gains you see in your portfolio are built on this promise of continued good times. That’s a shaky position.

Lean to hard earnings priced for the long term. The inflated valuations of AI driven names and the propped up valuations that are selling into it are going to fall hard. Start steering allocations into low earnings multiples — like enterprise software priced at 8x cash flow (our fund III thesis). It’s time to start thinking about intrinsic value from a value-perspective.

Bubbles are easy to call but hard to predict. There’s little cost to calling the contrarian view and a lot to gain if you’re right. Will the shade at the parade prove necessary to meet the promised demand or will it sit idle shading empty fields? Time will tell.