A Quietly Circular Agreement Regarding Silicon

There is an established convention in elementary commerce that when a person wishes to buy a loaf of bread, they provide the baker with money, and the baker provides them with the bread. Should the customer lack the necessary funds, they are generally expected to go without toast until their personal finances improve. At no point in this transaction does the baker offer to co-sign a bank loan on behalf of the customer so that the customer might borrow the requisite two pounds to purchase the loaf, while agreeing that should the customer default, the baker will repay the bank out of his own pocket.

Yet in the lofty and rarefied heights of high-performance computing, this whimsical financial arrangement has recently transitioned from a farcical hypothetical into standard enterprise practice.

It was reported this week that a major semiconductor manufacturer is considering standing as a financial guarantor for loan facilities raised by data centre developers. The capital from these loans, which naturally run into the hundreds of millions of dollars, is intended to buy the very graphics processing units manufactured by the guarantor itself. In essence, the vendor is offering to vouch for the borrower so that the bank will lend money to the borrower, which the borrower immediately hands back to the vendor in exchange for crates of silicon chips.

It is a mechanism of breathtaking elegance, resembling nothing so much as a gentleman handing his nephew a ten-pound note for his birthday on the strict condition that the nephew spends it on taking his uncle out for lunch.

To the uninitiated observer, this circular financial choreography might appear somewhat perilous. If the data centre operator fails to generate sufficient revenue from its algorithmic workload—perhaps because the software running on the chips spent three weeks attempting to write mediocre poetry rather than revolutionising global logistics—the bank will turn to the guarantor for payment. The vendor will then find itself in the peculiar position of paying the bank for chips that it already manufactured and delivered to someone else's warehouse.

However, modern corporate finance operates on a plane of abstraction where such worldly anxieties are treated as mere bookkeeping details.

From the perspective of the semiconductor firm, the arrangement is undeniably tidy. Wall Street analysts demand continuous, quarterly revenue growth. Data centre operators demand thousands of liquid-cooled circuit boards. Banks demand a firm promise that someone with an enormous balance sheet will step in if the entire enterprise turns out to be a temporary hallucination. By placing its own signature on the bottom of the credit agreement, the manufacturer satisfies all three demands simultaneously, creating an unbroken circuit of mutual reassurance.

I was once acquainted with a clerk in a small provincial insurance firm who attempted a similar strategy with his personal wardrobe. Having taken a fancy to a tailored tweed coat he could ill afford, he persuaded the tailor to accept a promissory note backed by a third-party guarantee. When asked who the guarantor was, the clerk triumphantly produced a letter signed by himself in his capacity as Assistant Claims Assessor for the local parish. The tailor, impressed by the official letterhead, handed over the coat without further inquiry. The arrangement collapsed three months later when the clerk was forced to review his own claim for non-payment, but for twelve brief weeks, he was undeniably the best-dressed man in the office.

The broader implications of this new credit paradigm are fascinating to contemplate. If hardware vendors continue to assume the credit risk of their own customer base, the traditional distinctions between supplier, customer, and central bank will inevitably blur into a warm, harmonious haze. We may soon enter an era where buying a computer graphics card requires an interview with the manufacturer's credit committee, while applying for a mortgage requires demonstrating that one owns at least three server racks in a facility near Wheeling, West Virginia.

There is also the intriguing question of what happens to the collateral. Should a data centre project default, the guarantor-supplier would presumably repossess its own chips. Having repossessed them, it would then be obliged to sell them to another customer, perhaps co-signing another loan for the new buyer to facilitate the transaction. Taken to its logical conclusion, a single batch of high-end processors could conceivably be sold, repossessed, guaranteed, and resold five or six times, generating billions of dollars in recorded top-line revenue while sitting entirely undisturbed in the same climate-controlled building.

Underwriters at major credit rating agencies are reportedly spending considerable time staring out of rain-streaked windows, attempting to decide whether a loan guaranteed by the seller of the asset being financed constitutes a AAA investment or a elaborate parlour game. The consensus appears to be that so long as the fans in the server room continue to spin at maximum velocity, the distinction is largely academic.

In the meantime, the financial engineering underpinning the artificial intelligence boom continues to reach new heights of artistic expression. We have progressed from paying for software to renting software, and from renting software to borrowing money from the software maker to rent the server that runs it. The entire edifice rests upon an unshakable faith that somewhere down the line, an algorithm will do something so spectacularly useful that it pays for all the loans, all the guarantees, and all the electricity consumed along the way.

Until that glorious day arrives, one can only admire the sheer chutzpah of the arrangement. It takes a remarkable degree of confidence to look at a multi-million-dollar credit exposure and conclude that the best person to guarantee it is the merchant standing behind the counter.