Dark fibre and infinite interns

Technology is an industry powered by competing narratives, but the dominant one this season is structural panic. If you read the essays driving the loud, exhausting 'Let AI Burn' narrative sweeping the tech industry—and the inevitable Hacker News doomerism trailing in its wake—you will find a perfectly rational fear that hyperscaler GPU spending is an unrecoupable capital furnace. The underlying math is genuinely terrifying. Sequoia Capital points out a massive $500B+ gap between the capital going into data centers to buy Nvidia chips and the actual software revenue coming out the other side. People look at the concrete being poured in Iowa and the power contracts being signed with nuclear facilities, and they do the back-of-the-envelope calculations. The financial logic dictates that venture capitalists, sovereign wealth funds, and tech giants are currently funding a bonfire, and that they will inevitably lose their shirts. This is mostly irrelevant.

The last time the industry built a capital bonfire of this magnitude, telecom companies laid millions of miles of optical cable for an anticipated wave of early internet traffic that simply failed to arrive on schedule. The resulting 1999 telecom crash wiped out trillions of dollars in equity and bankrupted the immediate investors, but the physical cables remained under the ocean. They became unused 'dark fibre' that subsequently drove the marginal cost of global bandwidth down to practically zero. Global Crossing went bankrupt, but the cheap structural substrate they left behind is what made YouTube, Netflix, and the entire architecture of cloud computing economically viable. You can map the exact same dynamic back to the 1840s UK Railway Mania, a cycle that fits neatly into Carlota Perez’s framework of 'installation versus deployment' phases. The early speculators always lose everything in the installation crash. The tracks, however, stay in the ground. They permanently alter the physical constraints of the adjacent industries, creating national mail services, commuting suburbs, and catalogue retail. People lost a fortune laying the iron, but Sears Roebuck built an empire on top of it.

Focusing on the immediate financial bubble completely misses the structural consequence of its inevitable bursting. The physical byproduct of this current capital furnace will be the total collapse of the marginal cost of computing logic. If you were to map this dynamically onto a whiteboard, you would draw one line showing hyperscaler capex exploding upward to the right, crossing violently with a second line tracking the cost-per-million tokens for foundation model APIs plummeting toward the floor. The cost of deployment is already in free-fall. OpenAI recently dropped their API pricing by a full order of magnitude in mere months with GPT-4o mini (and the floor continues to fall away month by month, moving rapidly toward fractions of a cent). This isn't just a rounding error in an IT budget—it changes the fundamental nature of what software is allowed to do.

When inference drops to essentially zero, we do not suddenly get a grumpy omniscient machine running the economy—we simply get a vast utility grid of infinite interns. Software currently requires highly structured inputs to do anything useful—drop-down menus, precise database queries, rigidly formatted spreadsheets. But if you can rent the cold start for fractions of a penny, you can suddenly deploy a system to write a hundred mediocre first drafts, parse ten thousand messy PDFs, or classify a million angry customer service emails without ever worrying about the compute bill. It doesn't matter if the machine occasionally hallucinates, any more than it matters if an intern misfiles a document; you just apply the Newspaper Test and print a daily retraction next to the crossword. The physical limits of hyperscale architecture cease to be the interesting part of the equation. What matters is the grueling, unglamorous work of re-architecting legacy enterprise platforms to assume reasoning is free.

We know with absolute certainty that this infrastructure is being built, and we know that many of the people paying for it will lose a fortune. But does this zero-cost utility accrue to the platform layers in Silicon Valley, or does it become a retailing and distribution problem solved by legacy incumbents who already own the customer relationships? What exactly does the enterprise software market look like when you can query an infinite supply of free, slightly mediocre interns from any application on earth? Is this something that gets solved by three engineers in a garage, or is it a workflow problem that takes a decade of slow, boring integration by global systems integrators? We know what the utility looks like when it arrives, but mapping exactly how the rest of the economy reorganises around it is simply an exercise in guessing.