Fiber Lasts Decades, GPUs Only Three Years: The AI Boom Is Not the Dot-Com Bubble

A Misread Analogy

Whenever AI investment hits new records, warnings about a "dot-com bubble repeat" come roaring back. A discussion on Hacker News offers a counterintuitive judgment: the AI boom is not the dot-com bubble. The argument isn't about valuations or fundraising pace, but a difference at the physical level—fiber optic cable lasts for decades, while GPUs become obsolete in three years.

This comparison is powerful because it points to a fundamental difference in asset structure between the two waves. In the dot-com era, massive amounts of capital were sunk into long-lived infrastructure like fiber optics, data centers, and backbone networks. After the bubble burst, those assets remained in service, paving the way for the later cloud computing and streaming era. In this round of AI arms racing, the most expensive assets are GPU clusters, whose economic lifespan is tightly locked to the pace of hardware iteration.

The Divide Between Long-Lived Assets and Short-Cycle Hardware

Fiber's "Sunk Cost" Is Actually a Legacy

After the dot-com bubble burst, large amounts of already-laid fiber were resold at low prices. It was precisely this "excess capacity" that supported broadband adoption and the later hyperscale data centers. Infrastructure depreciation cycles are measured in decades; the cost of investment mistakes gets spread out over time, and can even transform into the public foundation for the next growth cycle.

GPU Depreciation Curves Are Much Steeper

By contrast, the economic lifespan of a GPU is generally believed to be only about three years. The reason isn't hard to understand: model sizes keep expanding, new-generation chips make generational leaps in compute power and energy efficiency, and old cards quickly lose their cost-effectiveness for training tasks. This means AI infrastructure investment must recoup its costs within a much shorter window, or it's a real impairment.

This difference changes how we characterize the "AI bubble." If capital is sunk into short-lived assets, the damage from a bursting bubble isn't a bunch of idle fiber optic cable—it's a bunch of rapidly depreciating electronic equipment.

The reason GPU economic lifespans are anchored at around three years is directly related to the pace of AI hardware iteration. Manufacturers represented by NVIDIA basically maintain an architecture update cadence of about two years per generation (Ampere→Hopper→Blackwell), with each generation making significant leaps in power per unit of compute, memory bandwidth, and interconnect capability. For training clusters, old cards aren't incapable of running—they just produce significantly lower effective compute under the same power and data center space, and opportunity costs rise quickly. Accounting depreciation periods (commonly 3-5 years) aren't entirely equivalent to true economic lifespan; the latter also factors in memory pressure from expanding model sizes and the depreciation rate in the secondary market. Understanding this is essential to judging whether a GPU purchase is buying "compute" or buying "compute that won't be obsolete for a period of time."

Why This Doesn't Necessarily Mean a Crash

The Demand Side Is Different in Its Authenticity

In the dot-com era, many companies had no verifiable business model—traffic and clicks were treated as leading indicators of revenue. Current AI spending, while aggressive, is backed by real productivity demand from cloud providers and large enterprises. Inference calls, model training, and agent workflows are all generating actual bills. According to the viewpoint in the Hacker News discussion, the authenticity of demand is one of the key differences between this round and the last.

A commonly used observational dimension for measuring the authenticity of AI demand is the ratio between revenue and capital expenditure. During the dot-com period, many carriers' backbone network utilization remained well below design capacity for extended periods—investment was severely disconnected from actual traffic. While the AI-related revenue growth of leading cloud providers today is hard-pressed to fully cover their capex growth, inference call volumes, API bills, and enterprise subscriptions do constitute auditable cash flow. It's important to note that "real demand" and "reasonable return on investment" are two separate propositions: the former shows the market exists, while the latter depends on whether the price per unit of compute can cover depreciation and power costs. Discussions often conflate the two, but what really determines the cycle's direction is the latter.

Competition Compresses Returns but Doesn't Zero Them Out

Interestingly, short-lived hardware also means the industry has no cushion for "coasting on a win." Players must keep investing to stay competitive, and any generational lag can quickly show up in the cost structure. This pressures profit margins, but it's not the same as demand disappearing—it's more like a capex treadmill than a castle in the air.

The economic meaning of the capex treadmill can be understood through the "prisoner's dilemma": if a single manufacturer slows procurement, short-term cash flow improves, but it may fall behind in the next-generation model training race, losing customers and pricing power. So all major players are forced to maintain high-intensity investment, industry-wide profit margins get consumed by capex, and no individual can unilaterally opt out. This resembles the competitive dynamics of capital-intensive industries (like aviation and telecom)—demand persists, but excess profits are hard to sustain. The difference is that AI hardware depreciates faster, leaving manufacturers a shorter window to correct course.

Practical Implications for Developers and Enterprises

This framework has direct value for technical decisions. When evaluating AI infrastructure investment, putting depreciation cycles into your model is more useful than watching fundraising news: projects under a three-year payback assumption and those under a ten-year payback assumption have completely different risk profiles.

For users, this also explains why renting GPUs on demand is more rational than building your own in most scenarios—transferring hardware obsolescence risk to cloud providers and bearing only usage costs. For suppliers, whoever can maintain the highest utilization rate within this short cycle holds the key to winning this round of competition.

Conclusion

The value of the phrase "fiber lasts for decades, GPUs only three years" isn't in predicting when the bubble will burst, but in reminding us: the balance sheet structures of the two technology waves are fundamentally different. Using the dot-com script to project the endpoint of the AI cycle may cause us to miss the truly important variable—the mismatch between asset lifespan and return windows.

(This article's viewpoints are synthesized from relevant Hacker News discussions.)