Has Compute Become Cheaper Than Municipal Tap Water?

It is a well-established law of modern commerce that whenever an industry spends several hundred billion dollars constructing vast, cathedral-like facilities full of humming silicon and specialized liquid cooling infrastructure, the immediate market value of the resulting product ought to be at least moderately impressive. If one builds a deep-sea oil rig or a high-speed railway, one expects to charge something slightly more ambitious than three pence for a round trip.

Yet in the curiously inverted economy of artificial intelligence, we find ourselves contemplating a landscape where the cost of generating automated speech, complex reasoning, and digital conversation has quietly collapsed from thirty cents per transaction to less than a single fraction of a penny. Software developers, having applied a series of clever algorithmic compressions and specialized hardware tweaks, now routinely boast that an artificial intelligence model can conduct a full-throated, interactive conversation for roughly the price of half a grain of coarse sea salt.

This hyper-deflationary plunge in unit economics presents corporate boardrooms with a rather exquisite conundrum. On the one hand, lowering the operational cost of software by ninety-seven percent in a single afternoon is precisely the sort of achievement that earns a senior engineer a modest glass of warm sparkling wine at the annual general meeting. On the other hand, it raises uncomfortable questions for the executive who only last Tuesday persuaded institutional investors to underwrite a forty-billion-dollar bond issue for new power stations, on the strict understanding that synthetic thought was an exceedingly rare and costly luxury commodity.

(I was reminded briefly of a regional water board official who once attempted to invoice a suburban garden pond for unmetered atmospheric evaporation, only to discover that the rain had rendered his ledger entirely superfluous.)

Consider the sheer physical disproportion of the arrangement. In rural Virginia or northern Sweden, acres of reinforced concrete house tens of thousands of graphics processors, each drawing enough electricity to illuminate a mid-sized cathedral, all cooled by closed-loop chillers operating at peak industrial pressures. At the other end of this immense thermodynamic chain sits a mobile phone application, politely asking whether a language learner would like to order a croissant in conversational French, and consuming six-thousandths of a cent in the process.

The absurdity lies not in the efficiency itself—which is admittedly a triumph of mathematics—but in the economic fiction required to keep the entire apparatus upright. For several years, the prevailing consensus in Silicon Valley was that digital intelligence would behave rather like high-grade gold or aged mahogany: expensive to produce, carefully rationed, and billed to enterprise clients on generous monthly retainer agreements. Instead, it has taken on the economic characteristics of municipal tap water. It is endlessly available, piped quietly into every conceivable software interface, and so absurdly cheap per litre that nobody bothers to check the meter until the basement is under six inches of water.

This creates a peculiar dilemma for financial analysts attempting to calculate the return on investment for the ongoing infrastructure boom. When the capital expenditure is hyper-inflationary—with individual technology conglomerates committing fifty billion dollars a quarter to hardware purchases—and the unit price of the finished token is hyper-deflationary, the accounting department must engage in a high degree of creative narrative construction. One is forced to argue that while a single unit of artificial thought is essentially worthless, the world will eventually require so many quadrillions of them that the volume will neatly resolve the discrepancy.

(A retired bank auditor of my acquaintance once remarked that any business model relying entirely on the word "quadrillion" usually ends with someone handing over a ring of brass keys to a receiver.)

Moreover, the collapse in inference costs has begun to disrupt the carefully arranged hierarchy of enterprise software. When bespoke conversational models cost thirty cents a call, only major financial institutions and high-margin customer service desks could afford to deploy them. At a fraction of a cent, however, the technology instantly becomes cheap enough to embed inside digital alarm clocks, smart kettles, and automated parking meters. We are rapidly approaching the moment when a five-pound electric toaster contains enough spare computational capacity to offer unprompted advice on portfolio diversification while browning a slice of sourdough.

Whether this economic model can sustain the dizzying valuations currently assigned to data centre operators remains an open question. For the moment, the industry appears content to double down on scale, operating on the optimistic assumption that if one builds a sufficiently large reservoir, the sheer abundance of water will somehow alter the laws of arithmetic. In the meantime, consumers may enjoy the rare privilege of watching the world's most expensive machinery work tirelessly to deliver thoughts that cost less than the air used to whisper them.