In the late 1990s, the internet boom and a parallel telecom expansion unfolded at the same time. Carriers and new market entrants invested heavily in network infrastructure on the assumption that data demand would rise without limit. The Telecommunications Act of 1996, deregulation, aggressive valuations, and inflated traffic forecasts all contributed to one of the largest infrastructure overbuilds on record.

Capital arrived in sequence: venture funding, then capacity swaps used to inflate reported sales and earnings, then IPOs that financed still more construction. When demand failed to match those forecasts, market value collapsed. A widely circulated datapoint from an internal spreadsheet—that internet traffic would double every 100 days—helped justify the buildout. Actual growth was much lower.

The fiber overbuild

Hundreds of billions of dollars flowed into fiber-optic networks. New carriers such as Global Crossing, Level 3, Qwest, and WorldCom joined incumbents in laying tens of millions of miles of terrestrial and undersea cable. Dense wavelength division multiplexing (DWDM) made it possible to extract far more capacity from each strand, which further amplified supply.

By the early 2000s, the combination of the dot-com collapse, accounting scandals, and a severe telecom downturn left the sector with excess capacity and collapsing prices. Bandwidth rates fell by as much as 90 percent. Large portions of the newly installed infrastructure sat unlit. Estimates put utilized or lit capacity from the boom-era build in the low single digits—often cited in the 2.7 to 5 percent range.

What looked like stranded investment later became foundational infrastructure for the next computing cycle.

From unused fiber to cloud backbone

The fiber didn’t disappear. It remained a durable physical asset. As broadband adoption, video, mobile data, e-commerce, and cloud computing increased demand, the capacity was acquired at steep discounts.

The effect on cloud economics was material:

  • Low-cost, abundant bandwidth made it practical to place hyperscale data centers where power, climate, and geography were favorable, rather than next to every customer.
  • Centralized computing became commercially viable. Organizations could move from capital-intensive on-premises environments to scalable platforms such as AWS, Google Cloud, and Azure.
  • Improved global connectivity supported SaaS delivery, data-intensive workloads, and streaming at scale.
  • The cloud operating model—elastic capacity and usage-based pricing—depended on this pre-built transport layer.

The original investors absorbed the losses. The infrastructure itself proved durable. Assets that were underused in 2002 later supported a multi-trillion-dollar cloud market – and proved particularly reliable during the COVID cloud adoption era.

Ideas can be right before the market is ready

The same period produced another pattern that is easy to miss if the only lesson taken from 2000 is “avoid hype.” Some companies failed because the product was weak. Others failed because the product was early.

HomeGrocer.com, founded in 1997, is a clear example. The company built an online grocery service with warehouses, delivery fleets, and a reputation for fresh produce. The concept was coherent. The timing wasn’t. Broadband access was still limited. Online ordering was unfamiliar. Many households were unwilling to let a stranger select produce. The common objection was simple: customers didn’t want someone else picking their tomatoes.

HomeGrocer was acquired by Webvan in 2000. Webvan filed for bankruptcy in 2001. The category looked like a cautionary tale. Years later, the behavior became ordinary through Amazon Fresh, Instacart, and retailer-owned delivery programs. The idea wasn’t permanently wrong. Consumer habits, last-mile logistics, payment systems, and network access just needed to catch up.

That distinction might be significant in the current AI cycle. A capability can be commercially valid and still fail as a business if adoption, cost, workflow integration, or infrastructure lag behind the pitch. The reverse is also true: a failed first wave doesn’t prove the underlying use case was worthless.

Comparing the AI boom with the dot-com bubble

The current AI infrastructure cycle is frequently compared with the late-1990s internet boom. The comparison is useful only if the two episodes are separated clearly.

Where the cycles look similar

Both periods featured a widely accepted technology thesis, compressed timelines, and large capital commitments made before unit economics were fully proven. In the late 1990s, investors funded websites, software platforms, and the networks those services would require. Today, capital is concentrating in model development, GPUs, data centers, and power. In both cases, a small number of narratives—traffic growth then, model capability and token demand now—have been used to justify outsized spending.

Both cycles also produced circular commercial relationships. In the telecom era, capacity swaps and related-party transactions inflated reported activity. In the current cycle, chip suppliers, model labs, cloud providers, and data-center developers are increasingly interdependent through investment, capacity commitments, and purchase agreements. That doesn’t make today’s activity fictitious. It does mean reported demand may be less independent than it appears.

A third similarity is the gap between long-term usefulness and near-term return. The internet thesis was directionally correct. Many of the companies and capital structures built around it were not. AI could follow a similar pattern: the technology can be transformative while individual projects, vendors, and financing structures fall short.

Where the cycles differ

First, business models. Many dot-com companies had little revenue, weak unit economics, and no clear path to cash generation. The current AI build is being led in large part by profitable hyperscalers and established enterprises with existing customers, distribution, and free cash flow. That doesn’t eliminate overbuild risk. It does change who can absorb a downturn.

Second, the nature of the asset. Dot-com spending included a large share of marketing, software, and brand value that disappeared when funding stopped. Telecom fiber lasted for decades. AI infrastructure sits between those extremes. Data-center shells, power interconnects, and cooling plants are long-lived. GPUs, networking gear, and optimized server designs are not. Hardware refresh cycles in AI are measured in years, not decades. An overbuild in accelerators is more perishable than an overbuild in dark fiber.

Third, demand timing. In the late 1990s, much of the network was built ahead of proven mass-market demand. Today, AI workloads, cloud consumption, and enterprise pilots are already generating real usage. The open question is not whether demand exists, but whether it will grow fast enough, and at high enough margins, to justify the current pace of capital expenditure.

Fourth, financing. The telecom bubble was heavily debt-financed and vulnerable once capital markets closed. The AI cycle is mixed: large operators are funding a substantial share from operations, while specialists, neoclouds, and project-level vehicles are more dependent on credit and vendor arrangements. A correction would therefore be uneven. Balance-sheet strength will matter as much as technology quality.

Fifth, concentration. Dot-com losses were spread across thousands of issuers. AI exposure is more concentrated in a smaller set of model providers, chip vendors, and hyperscalers. That can cushion the broader economy if the largest firms remain solvent. It also means operational and pricing shocks can transmit quickly through the rest of the stack.

What “residue” actually means—and why GPUs are not fiber

The fiber era showed that two results can happen at once. Investors can lose large amounts of capital. The physical foundation they financed can still be useful later.

That is the residue: the part of the spend that remains after valuations collapse. In telecom, the residue was cable in the ground, empty conduits, and long-haul routes that could be lit when traffic finally arrived. Cloud providers did not need the original business plans to be right. They needed the glass, the rights of way, and the interconnection points. Those assets had long useful lives. Buying them after the crash was cheaper than building them from scratch.

The same logic does not apply evenly to today’s AI stack.

A data-center building, a substation, a power contract, a parcel with grid access, and the fiber that reaches the site can serve more than one generation of compute. If training demand slows, those assets can often be reused for inference, conventional cloud, storage, or other high-density workloads. Their value may fall. It does not reset to zero on a two-year chip cycle.

Accelerators are different. A GPU generation is a short-lived tool, not a 20-year backbone. If too many chips are purchased against demand that doesn’t arrive, the unused inventory ages quickly. Newer architectures, higher memory bandwidth, and better performance per watt reduce the resale and reuse value of last year’s equipment. An empty fiber route in 2004 could wait. A warehouse of last-generation GPUs can’t wait in the same way.

That’s why comparison between the 1990s overbuild and the current AI build isn’t apples to apples. Overinvestment in power, land, connectivity, and building structure can still leave behind reusable capacity. Overinvestment in a single generation of accelerators faces likelihood to become a write-off than a gift to the next wave.

What enterprise IT leaders should take from the outcome

Opus Interactive was founded in 1996 and has operated through the internet boom, the telecom collapse, the rise of cloud, and the current AI infrastructure cycle. The company even purchased some office furniture from lucy.com during the bust. The table in our conference room is a small reminder that the crash was local, material, and that survival often depends on changing the operating model rather than defending the original one.

The historical record is consistent: technology waves can create both durable platforms and expensive false starts.

The sharpest losses from the dot.com and telecom bubble fell on pure-play overbuilders that sold speculative capacity rather than contracted services:

  • WorldCom filed the largest U.S. bankruptcy at the time in 2002, later reorganized as MCI and was acquired by Verizon in 2006.
  • Nortel Networks, once a major share of the Toronto Stock Exchange, declined by roughly 99 percent and entered bankruptcy in 2009.
  • Lucent Technologies also lost nearly all of its peak equity value before a series of combinations that ultimately folded into Nokia.

The more useful lesson is not the list of failures. It’s the profile of the firms that endured. Survivors typically had:

  • Recurring revenue tied to actual customer workloads, not circular capacity sales
  • Manageable leverage, or the ability to reduce debt without disrupting operations
  • Liquidity sufficient to fund day-to-day operations through a downturn
  • Products or services that were difficult to replace
  • Operational discipline under compressed margins
  • Reliable partner ecosystems rather than one-way vendor dependence
  • The capacity to change architecture, pricing, and go-to-market plans quickly

Lucy.com is one version of that last point. HomeGrocer.com is the other side of the same decade: a service model that later became mainstream, but not on the timetable or capital structure of 1999. Capacity without demand is not an asset. An idea without adoption is not a market yet.

For enterprise IT, the practical standard remains the same across cycles: match infrastructure spend to contracted or clearly evidenced workloads, separate long-lived site decisions from short-cycle compute purchases, and avoid concentrating critical operations on a single vendor, financing model, or hardware generation.

Opus Interactive designs and operates purpose-built cloud, colocation, and AI-ready infrastructure with a relationship-driven delivery model. The company has more than 30 years of operating history and maintains PCI-DSS, HIPAA, SOC 2, and ISO 27001 compliance. To discuss hybrid architecture, capacity planning, or compliance-bound workloads, schedule a consultation with the Opus Interactive team.