The Silicon Boomerang

There is a lovely, almost pastoral simplicity to the traditional transaction. One party possesses a physical object—say, a particularly sturdy turnip—and another party possesses a coin. They meet in a draughty marketplace, exchange the turnip for the coin, and go their separate ways, each reasonably confident that they understand what has transpired. The turnip will be boiled; the coin will be spent on cider. The universe remains in a state of comprehensible equilibrium.

In the upper echelons of the artificial intelligence sector, however, the turnip has become highly abstract, and the coin has developed a habit of flying through the air in a perfect, self-contained loop. This has been neatly demonstrated by the recent revelation that Nvidia, a company currently valued at a sum that would make the Roman Empire look like a modest corner shop, has agreed to pay six billion dollars to license software from a startup called Poolside. For context, Poolside is a company whose primary product is an artificial intelligence designed to write computer code, and which was valued in its entirety at a mere three billion dollars not so very long ago.

To the untrained eye, this might look like a standard commercial arrangement. A large technology firm requires software; a small technology firm has written some; money is exchanged. But to view it in this light is to miss the exquisite, circular beauty of the modern silicon economy. It is a system that operates less like a marketplace and more like a highly sophisticated, multi-billion-dollar game of pass-the-parcel where the parcel is filled with cash and everyone is using the same wrapping paper.

Consider, if you will, the life cycle of the capital involved. A venture capital firm, having run out of sensible things to do with its investors' money, decides to back a promising AI startup. The startup, eager to prove its worth, immediately hands the vast majority of this capital to Nvidia in exchange for several truckloads of highly advanced graphics processing units. These chips are then installed in a data center that consumes enough electricity to power a medium-sized European principality, where they spend several months humming warmly while digesting the entirety of the public internet.

Eventually, the chips produce a model. In Poolside's case, this model is exceptionally good at writing Python scripts that do not immediately crash. Nvidia, observing this success, looks at its own mountain of cash—which is growing at a rate that is starting to cause structural concerns for the treasury department's floorboards—and decides that the best possible use of this money is to pay the startup six billion dollars to license the very model that was trained on Nvidia's own chips. It is the corporate equivalent of selling someone a spade, watching them dig up a chest of gold, and then paying them double the value of the gold just to borrow the spade back for the weekend.

One cannot help but admire the sheer, unbothered elegance of it all. It is a closed-loop ecosystem of pure, unadulterated liquidity. The money never actually has to leave the valley; it simply changes state, much like water evaporating from a puddle only to fall as rain on the exact same puddle. If one were to trace the journey of a single dollar bill through this system, one would likely find it has developed a severe case of motion sickness.

There is, of course, a perfectly serious institutional explanation for this. Nvidia is desperately keen to avoid the fate of previous hardware giants who found themselves relegated to the status of digital plumbers. No one wants to be the company that merely provides the pipes when they could be the company that owns the water. By licensing Poolside's code-generation models, Nvidia is attempting to build a comprehensive software ecosystem, ensuring that developers do not merely use Nvidia chips, but live entirely within an Nvidia-branded reality. It is a noble ambition, though one wonders if six billion dollars is a slightly steep price to pay to ensure that software engineers can generate mediocre boilerplate code three seconds faster than they could yesterday.

I recently spent an afternoon observing a small, automated vacuum cleaner attempt to navigate its way out from behind a Victorian sideboard. It was a masterclass in persistence. It would bump into the mahogany leg, pause for a thoughtful second, spin forty-five degrees, and then promptly bump into the skirting board. It did this for forty-seven minutes before its battery expired, at which point it let out a tiny, melancholic beep and fell asleep. I felt a deep, spiritual connection to that little machine. It was trying so hard to map a world that was simply too complicated for its sensors, using a series of pre-programmed collisions to find a freedom that did not exist. One suspects the boardrooms of Santa Clara are currently filled with a very similar kind of energetic bumping.

The sheer scale of the figures involved in these transactions has a curious numbing effect on the human brain. Six billion dollars is not a number that can be easily visualized. It is not a pile of banknotes; it is an abstract concept, a collective agreement to pretend that a specific sequence of ones and zeros in a banking database is worth more than the entire annual economic output of several small island nations. When a company can spend that sum on a licensing agreement without its shareholders so much as blinking, we have entered a realm of financial surrealism that would have made the Dadaists look like sober-minded accountants.

Perhaps the most delightful aspect of this arrangement is the concept of the 'licensing agreement' itself. Nvidia is not buying Poolside. It is merely renting the right to use its intellectual property. This means that, in a few years' time, when the current models have been rendered entirely obsolete by some new breakthrough involving quantum-entangled turnips, Nvidia will have to pay another six billion dollars to license the next version. It is the ultimate subscription model, a perpetual-motion machine of recurring revenue that ensures no one ever actually owns anything, but everyone is constantly paying for the privilege of holding it.

In the meantime, the hum of the data centers continues unabated. The chips grow hotter, the models grow larger, and the financial loops grow ever more intricate. We may not yet have achieved artificial general intelligence, but we have certainly succeeded in creating something far more impressive: an economy that can run entirely on its own steam, completely undisturbed by the messy, inconvenient realities of the physical world. One can only hope that when the algorithms finally take over, they remember to keep paying their licensing fees.