Dark fiber and burning GPUs
Equating the AI bubble to the dot-com crash ignores that, unlike passive dark fiber, power-hungry, rapidly decaying GPUs simply get turned off.
By Hugh Askell
Sparked by Treasury Has an Internal Report Warning About the Dangers of an AI Bubble · discussion
A leaked US Treasury report warning of a $750 billion AI bubble recently prompted the usual, exhausted hand-wringing on Hacker News about the 1999 dot-com crash. On the pure financial math of the capital destruction, the regulators are largely correct, yet they are committing a classic widget fallacy by treating all infrastructure crashes as structurally identical. We have a tendency in tech to assume that because the financial graphs look the same, the underlying physical assets must behave the same way. This is mostly nonsense.
To understand the error, we have to look at the exact historical precedent driving the current panic. From 1996 to 2001, the telecommunications industry spent hundreds of billions of dollars laying fiber-optic cables across the ocean floor and under city streets, assuming infinite, exponential demand for bandwidth. (At the time, the consensus analyst forecasts were simply charts that went up and to the right forever). When the financial reality caught up with the physical build-out, titans like Global Crossing went bankrupt, and the market evaporated.
But that massive capital destruction triggered a profound second-order consequence for the next wave of computing: it left behind millions of miles of dark fiber. Because the cables were already physically in the ground, they became a nearly free, zero-margin utility. Bankrupt institutional investors inadvertently subsidized the cheap, ubiquitous bandwidth that made Web 2.0, cloud infrastructure, and high-definition video streaming economically viable over the next two decades. You could build YouTube, or AWS, because someone else had already paid for the plumbing.
Today, we are looking at an identical financial trajectory. Analysts at Sequoia and Goldman Sachs point out that the technology industry is aggressively building out a $600 billion to $1 trillion AI infrastructure footprint, while actual ecosystem revenues lag drastically behind. The comforting assumption across the valley is that, just like 1999, even if the current hyperscalers and foundation model companies collapse, they will leave behind a permanent, cheap compute utility for the next generation of software startups to build upon. Rent the cold start, ride the bankruptcy subsidy, and build your SaaS.
This fundamentally misunderstands the physical mechanics of a modern data center. 'Compute' and 'fiber' are not interchangeable industrial materials. If you plot the two eras on a basic chart, you find a stark two-axis contrast: the 1999 telecommunications boom laid down passive assets with a thirty-year lifespan and negligible operating expenses, whereas the 2024 AI build-out relies on highly active hardware that decays in under five years and carries an operating overhead that rapidly eclipses the initial capital investment.
Fiber optic cable is entirely passive. It is simply a glass tube sitting in the dirt that costs almost nothing to maintain once the trench is dug. A massive cluster of H100 GPUs, by contrast, is an active industrial furnace. The ongoing power and cooling costs required to run these server farms often outstrip the purchase price of the IT equipment itself. You cannot run an abandoned GPU cluster at zero margin, because someone still has to pay the local power grid to keep the fans spinning and the coolant flowing.
Furthermore, the underlying silicon math inevitably moves on. An H100 chip will be hopelessly obsolete long before the decade ends, replaced by architecture that processes language models at a fraction of the energy cost. In 1999, the infrastructure was largely permanent; in 2024, the infrastructure is a rapidly depreciating space heater. You are not putting down a permanent road network — you are buying a fleet of trucks that will rust into uselessness in 48 months. And this ripples outward into real estate and energy markets: the bottleneck isn't just buying the chips, but finding a municipal utility willing to guarantee 500 megawatts of continuous power for a facility whose tenant might not exist in three years.
And so this comes back to the second-order ripples crashing into the software ecosystem. If you are building a product on top of large language models today, your business model implicitly assumes that the cost of your infinite automated interns will trend toward zero, exactly as broadband did. You expect to ride the subsidized wave, outsourcing the massive capital expenditure of intelligence to Microsoft or Google, and reaping the software margins.
But what happens if the physical operating floor of a data center is simply too high to subsidize indefinitely? If the venture capital dries up and the utility bills come due, that compute does not gradually get cheaper in bankruptcy. It gets turned off. We have spent the last two years assuming that a financial bubble will automatically yield a permanent, cheap utility, mostly because that is exactly what happened the last time we had a bubble. So, what happens to the software ecosystem when the subsidized compute gets unplugged and the interns go home? What is the baseline cost of intelligence when no one is willing to pay the cooling bill? I don't think we know.