Fiber in 1999, GPUs in 2026

People keep telling me GPU servers are the new fiber. I think that is half right. The wrong half decides whether a lender gets repaid.

What 1999 actually looked like

More than a dozen companies built expensive networks on a belief that traffic would keep growing. They paid for it with debt and equity, ahead of revenue. The demand came, but not on schedule.

By mid-2002, 97 percent of existing lines went unused at any single moment. TeleGeography found that only 3.9% of long-haul fiber through Chicago was lit.

The bills came due. Global Crossing filed for bankruptcy on January 28, 2002. It was then the fourth-biggest bankruptcy in US history. 360networks also filed, and Level 3 warned on covenants.

The glut was technical as much as financial. Wavelength-division multiplexing multiplied the capacity of each strand. The same glass suddenly carried far more traffic. Supply jumped without anyone digging a new trench.

Where the analogy holds

The shape of the build is the same. Capex goes in ahead of revenue. Debt and equity pay for it. Many players overbuild. The customers are themselves funded by capital markets. Then consolidation arrives.

Gartner expects the same ending. Its AI infrastructure report forecasts that over 80% of neocloud providers cease independent operations by 2030. I covered the write-down side of that forecast in The free-for-all is over.

Operators fail for different reasons, and the reason matters to a lender. I will come back to that.

So far the analogy is fair. Now the break.

Where it breaks

Asset life. Fiber plant lasts 25 to 40 years. The glut cleared because demand grew into cheap capacity that still worked. Time was on the side of whoever held the fiber by then, even where the company that laid it had gone bankrupt.

A GPU server does not work that way. Gartner assumes semiconductors depreciate over 5 years and other infrastructure over 20. That is the same source I just used for the consolidation forecast. Large cloud providers depreciate their servers over about five to six years: Microsoft and Alphabet over six, Meta over 5.5, and Amazon over five for part of its fleet. We underwrite a six-year useful life, with lower rental rates assumed at re-lease. Actual life may be shorter.

A new chip generation arrives, and the old one loses economic value. You cannot grow into a GPU glut. The hardware gets replaced first. There is no dark fiber option on a B200.

GPU server: accounting lives at large cloud providers; Gartner assumes five years for semiconductors.

Source: GPU Capital; Gartner, September 14, 2026; company filings as summarized by Global Data Center Hub, September 11, 2026.

Show data

Fiber gluts clear. GPU gluts expire.

Asset	Low	High
Long-haul fiber plant	25 years	40 years
GPU server	5 years	6 years
Typical chip loan tenor	3 years	4 years

Supply mechanism. WDM made fiber supply cheaper and bigger overnight. GPU supply is held back by chips, power and memory. Those inputs are getting more expensive, not cheaper. The oversupply risk is real, but the mechanism is different. It comes from the next chip generation, not from a free multiplier on the old one.

Buyers. Fiber-era CLECs had little contracted, creditworthy demand. Many sold on a hope of traffic. In the deals I see, the better operators sign 3 to 5 year contracts. Large neoclouds such as Nebius collect prepayments that cover 50 to 60 percent of capex on many deals. That is a real improvement.

Visibility. Fiber in the ground is passive and hard to price. GPUs have a live secondary market with measurable prices. A lender can see collateral value instead of guessing it.

Exhibit 2. Fiber in 1999 and GPU servers in 2026

Feature	Fiber, 1999	GPU servers, 2026
Economic life	25 to 40 years	About 5 to 6 years
What caused oversupply	WDM multiplied each strand	Next chip generation replaces the old one
Supply limits	Cheap to light more capacity	Chips, power, memory, all costly
Contracted demand	Thin	3 to 5 year contracts at better operators
Collateral pricing	Hard to observe	Live secondary market

Source: GPU Capital.

Where the analogy still bites

The 1999 CLEC is alive today. It is the neocloud on a short contract with a VC-funded customer, financed against an asset that loses value every year. The lender in that structure is holding the same risk as the 1999 lender. The difference is that the asset will not wait.

Prepayments do not change this. A 50 to 60 percent prepayment buys repayment speed. It also moves counterparty risk onto the prepaying customer. That is useful, but the risk is moved, not removed. The tail is still the lender’s: re-contracting and residual value after year 3.

What it means for lenders

I use four rules.

1. Tenor inside the asset’s economic life. If the loan runs past the point where the hardware is worth the debt, the collateral cannot cover it.
2. Customer credit ahead of the operator story. A named, creditworthy customer on a signed contract is worth more than a good pitch deck.
3. Prepayments as repayment speed, not validation. Faster paydown is good. It does not prove the tail is safe.
4. Never lend on the next generation’s price curve. Value the whole node in a working site with a named customer. Do not underwrite a recovery that needs a future chip price to hold.

My view

The fiber lesson is not “do not lend.” Plenty of fiber lenders were repaid, because they lent against assets that lasted. The lesson is to match tenor to asset life, put customer credit first, and refuse to lend on a price curve that has not happened yet.

Tomorrow’s winners will not repay a GPU loan. It is repaid only if today’s customers keep paying inside the hardware’s life, and collateral value and tenor decide how much is lost if they stop.

The glut will clear someday, as it did for fiber. The question is whether the loan is already repaid when it does.