The ancient future of finance

The promissory notes recorded on clay tablets in ancient Mesopotamia were often denominated in barley — an arrangement that was particularly convenient for farmers of barley, who had the distinct advantage of issuing claims redeemable for a product they happened to produce.

This is far preferable to the modern way of doing things. 

For example: When a ball-bearing manufacturer borrows in dollars and the price of ball bearings collapses, its income may no longer suffice to service its loans. The company goes bankrupt and investors are left with a claim on some worthless factories. Everyone loses.

If the same loan is denominated in ball bearings instead, the company can make its payments no matter what the price of its product does. Its obligations and revenue are matched.

Better still, if the lender is also a consumer of ball bearings, everyone wins. The loan simultaneously finances production and guarantees future supply.

Unfortunately, these kinds of perfectly matched relationships are rare. Which is why we have money — a universal settlement asset that facilitates trade.

It’s not perfect. One downside of money is that borrowers are often left with liabilities that no longer match their income, causing ruin.

But what if money itself was a liability-matched promissory note?

Bob Gelfond thinks it’s possible.

The quantitative trader and technologist argues that the current buildout of AI computing capacity should be financed by debt denominated in AI computing capacity — and that the debt should circulate as money.

This, he explains, would be a far safer way to finance history’s largest capex boom. 

The danger is that Big Tech has exhausted its immense cash flows building AI compute capacity and is now borrowing to build more of it. 

This raises a once-unthinkable prospect: A recent paper warns that if AI fails to deliver the expected productivity gains, technology giants like Amazon and Alphabet could face "failure to make interest payments on debt, and perhaps bankruptcy."

Yikes.

If, instead, the payments were instead denominated in compute, there would be very little chance of hyperscalers ever missing them — to the great relief of both investors and the economy.

Gelfond explains that holders of this debt would have a claim on a hyperscalers data center capacity instead of its cash flows, thus matching the companies’ obligations to the assets they’re building.

That could make hyperscaler debt something close to risk-free. Just as the US government can always print the dollars it owes its lenders, hyperscalers should always be able to produce the compute they owe.

Who would buy such debt?

Potentially everyone.

If managed correctly, a perpetual version of this debt could become the holy grail of monetary economics: a currency that never loses its purchasing power.

Left to its own devices, compute currency would naturally depreciate. For example, one unit of currency issued by Amazon might be redeemable for an hour of compute on Nvidia H100 GPUs in an AWS data center. As H100s were surpassed by newer generations of chips, that hour of compute — and the currency it backed — would become less valuable.

To incentivize people to hold the currency, Amazon would have to offset that depreciation in some way. Perhaps by paying a yield to holders, or periodically upgrading the backing to a newer generation of processors.

This would be a kind of monetary policy, like the Fed uses to manage the dollar.

But whereas the Fed aims for the dollar to lose 2% of its purchasing power each year, Amazon would likely aim to preserve the purchasing power of its currency — or increase it, even.

Unlike the US government, which simply mandates the use of money it issues, Amazon would have to persuade people to use theirs by offering better terms than its competitors.

The competition would likely be fierce.

The more widely a hyperscaler's currency is held, the less frequently it's redeemed — and the less frequently it's redeemed, the higher the issuer's margins.

In other words, issuing perpetual debt that circulates as currency would be to hyperscalers what gift cards that never get used are to Starbucks.

Hyperscalers would therefore be highly incentivized to encourage adoption by making their currencies trustworthy: supply would be carefully managed, issuance would be transparent (on a blockchain, probably), and promises would be kept.

Would that be enough?

Getting a new currency adopted is no easy task (just ask bitcoin). But compute currency would have several advantages in overcoming the cold-start problem it faces.

• Its backing seems likely to become one of the world's largest markets: Annual spending on compute may soon reach into the trillions of dollars.

• Hyperscalers could promote the use of their currency by offering discounts when it's used to pay for all that compute.

• Compute-backed currency would be ideal for AI agents, offering them a claim on the resource they need to complete their tasks.

• Issuing currency would be so valuable to hyperscalers, they could happily subsidize a payments system with no fees, undercutting the credit card system. 

But the biggest advantage of compute-backed currency might be what it could mean for the financial system itself. 

If Hyman Minsky is correct that an economy can be described by a set of interrelated balance sheets and income statements, compute-backed currency offers a way to make the economy more resilient.

Financial crises and crashes are often caused less by bad assets than by liabilities that can't survive a temporary decline in asset values.

To the extent the financial system adopted currency backed by compute, assets and liabilities would be more closely matched, and the system would be less likely to crash.

Because here’s something that’s never changed: Every economy is built on promises, and promises are safest when they're redeemable in the thing the borrower produces.

We've known that since at least the Mesopotamians.

Now may be the time to bring it back.

(Without the clay tablets.)

— Byron Gilliam