Selling Shovels to Yourself

The oldest piece of investing folklore is that in a gold rush, the safe money sells shovels. You do not bet on any one prospector striking it rich. You sell the tools every prospector needs, and you get paid whether or not the gold is ever found. It is good advice, and the AI industry has built something that looks like it and is not.

In the AI build-out, the shovel-seller is also funding the prospectors, and the prospectors are spending the funding on shovels, and a large share of the revenue that justifies the shovel-seller’s valuation is money it routed to its own customers in the first place. The shovels are real. The mine may or may not exist. What has been engineered is a loop in which capital circulates between a small number of firms, gets booked as revenue at each pass, and is presented to the market as organic demand. This is the financial structure of the boom, and it is worth understanding before the next pension statement arrives.

The shape of the loop

Strip the names out and the pattern is simple. A chipmaker sits at the centre, selling the scarce hardware that AI requires. It takes some of its profits and invests them into the model labs and the specialized cloud providers that are its largest customers. Those customers use the investment, plus money raised elsewhere, to buy the chipmaker’s hardware. The hardware purchase is booked as revenue by the chipmaker, which lifts its valuation, which gives it more capacity to invest in its customers, which generates more hardware purchases.

A parallel loop runs through the hyperscalers. A cloud platform invests billions in a model lab. The investment is structured so that much of it is spent back on the cloud platform’s own compute. The lab records the compute as its operating cost and the platform records it as revenue, and the same dollars appear on two sets of books doing two different jobs. The platform tells its shareholders that AI is driving cloud growth. The growth is, in meaningful part, its own money making a round trip with a press release attached.

None of the individual transactions is fraudulent. Each one is a real contract for real goods. The problem is not any single deal. The problem is that the aggregate is being read as a signal of end demand when a large part of it is internal circulation. The market is pricing the loop as though every dollar of revenue came from an outside customer who wanted the product, when a substantial share came from the seller’s own balance sheet on a return trip.

Vendor financing has a track record

This is not a novel financial innovation. It is vendor financing, and it has a history, and the history ends the same way each time.

In the late 1990s, the telecom-equipment makers financed the upstart carriers that bought their gear. The equipment makers booked the sales, the carriers built networks against demand that had been forecast rather than observed, and the whole structure looked like a boom until the forecast demand failed to show up. When it did not, the carriers could not service the debt, the orders stopped, and the equipment makers discovered that a meaningful slice of their revenue had been their own credit coming back to them. The losses were not confined to the firms that made the bet. They were spread across every pension and index fund that held the inflated equities on the way up.

The mechanism rhymes precisely with the present. Demand that is forecast rather than observed. Revenue that is partly the seller’s own capital recirculated. A valuation that prices the loop as organic growth. The tell, then as now, is that the capital expenditure is racing ahead of any independently measurable revenue from end users who are not themselves inside the loop. The build-out is justified by the build-out.

The capex is the bet

The defenders of the boom say the spending proves the conviction. Firms do not pour tens of billions into data centres unless they believe the demand is coming. This is true and it is also exactly what was said in 1999. Conviction is not demand. A capital-expenditure commitment is a bet that demand will arrive at a scale and a price that justifies the asset, and the asset in question depreciates fast. A graphics processor is not a railway. It does not earn for forty years. It is a fast-aging industrial good with a short useful life, and the economics only work if it is run near capacity, billed at a healthy margin, against paying customers, for the few years before it is obsolete.

So the real question is not whether the firms believe. It is whether there exists, outside the loop, enough end demand at a high enough price to amortize a hardware build-out of this size before the hardware ages out. That number is knowable in principle and is being asserted rather than demonstrated. The revenue that would prove it, the revenue from customers who are not also suppliers or investees, is the number that gets the least airtime, because it is the number least flattering to the story.

Whose pension is in the loop

For a Canadian reader this is not a spectator sport, because the Canadian retirement system is a large, sophisticated, global investor, and large sophisticated global investors are exactly who ends up holding infrastructure bets at the top of a cycle.

The major Canadian pension managers and the large Canadian asset managers have real, disclosed exposure to the AI build-out. They hold the chipmaker and the hyperscalers through public-equity portfolios that track concentrated indices, which means the loop is over-represented in the holdings whether or not anyone chose it. They invest directly in data-centre infrastructure, which is precisely the asset whose value depends on the demand forecast holding. The exposure is not reckless on its face. It is the ordinary consequence of being a giant investor in a market where a handful of names have come to dominate the index and the infrastructure pipeline at the same time.

That is the part that should focus the mind. The risk is not that one fund made one bad bet. The risk is structural. When the loop is large enough to move the index, every passive holder is in it, which means every Canadian with a public pension or a market-tracking retirement account is in it, whether or not they ever formed a view on AI. The losses from a vendor-financed bubble are not borne by the firms that built it. They are socialized across the savers who held the inflated assets, and the Canadian saver holds them through institutions that are supposed to be the adults in the room.

The argument

Selling shovels is safe only when the prospectors are spending their own money in search of gold that someone, somewhere, actually wants. The AI build-out has bent that logic into a circle. The shovel-seller funds the diggers, the diggers buy shovels, the cloud platform funds the lab and bills it for the compute, and each pass of the same capital is booked as revenue and sold to the market as demand. The transactions are real. The demand they are taken to prove is, in large part, the industry buying from itself.

Vendor financing has done this before, and it has ended the same way every time, with a capital build-out that outran the end demand and a bill that landed on the broad public that held the equities on the way up. The Canadian version of that public is the pension system, which is deep in the loop by the ordinary mechanics of indexing and infrastructure investing.

The discipline this calls for is unglamorous. Ask for the end-user revenue, the revenue from customers who are not also suppliers or investees, and treat everything else as circulation rather than growth. A boom that cannot show you that number is selling you the shovel and the loop and the story all at once, and arranging for your retirement to be the buyer of last resort.