Perhaps a better analogy is the Y2K/Dot Com bubble bust.
Money flows to “opportunities” which don’t pan out. Companies go belly-up. Some people lose a pile of cash. But the technology emerges stronger and more solid and transforms the entire economy over the following decades.
"1. Equity Losses vs. Debt-Leveraged Backlogs
The 2000 Dot-Com crash was primarily an equity event. Venture capitalists and retail investors bought overvalued stocks. When Pets.com or Webvan went under, equity evaporated, investors took the loss, and the damage was largely contained to balance sheets. Today's AI boom is built on a complex credit chain: Big Tech hyperscalers are issuing real debt and committing $700B+ in annual capex against $2.1 trillion in "take-or-pay" contracts owed by startups that lose tens of billions a year. When equity dies, investors lose money; when debt-funded supply chains break, credit markets freeze.
"2. Telecom Dark Fiber vs. Public Utility Rate-Basing
In 2000, telecom companies laid thousands of miles of dark fiber. When those telecoms went bankrupt, private bondholders got wiped out, but the fiber stayed in the ground and regular citizens weren't forced to pay for it on their monthly bills. Today's data center buildout requires massive physical power generation, high-voltage transmission lines, and substations built by regulated public utilities. Under utility rate-basing rules, those multi-billion-dollar infrastructure costs are embedded into the regional rate base—meaning captive residential consumers and retirees pay for those assets through higher monthly electric bills regardless of whether the tech companies stay solvent.
"3. Market Concentration & Retirement Exposure
In 2000, the S&P 500 was far more diversified. Today, roughly 40% of the entire S&P 500 index sits in just 10 mega-cap technology companies, all heavily leveraged into the exact same AI capex trade. A broad repricing doesn't just hit speculative tech portfolios; it directly impacts standard, passive index retirement funds that everyday workers rely on.
"4. Sovereign Fiscal Backdrop (2000 vs. Today)
When the Dot-Com bubble burst in 2000, the US Federal Government was running an annual budget surplus and national debt was under 60% of GDP. The Fed had immense flexibility to slash rates without threatening sovereign debt stability. Today, the US operates under a $1.6 trillion annual deficit, $39.8 trillion in national debt, and over $12 trillion in Treasuries maturing within 18 months. An economic shock originating in tech hits a sovereign balance sheet that has zero fiscal margin for error.
"The Dot-Com crash was an equity reset in a financially healthy macro environment. The current scenario is a debt-stacked, utility-backed infrastructure expansion happening at the peak of sovereign fiscal strain. Calling it "just another Dot-Com bubble" ignores who holds the debt and who gets stuck paying for the physical grid."
Total telecom investment was approximately $1.3 trillion nominal, equivalent to roughly $2.5 to $2.6 trillion in 2026 dollars over about six years (approximately $430 billion per year average, adjusted). Roughly $320 billion was borrowed 1999-2001 alone, plus $500 billion in bonds issued 1996-2001. The buildout was not just optical cable physical mileage -- it was a multiplicative stack:
2024 spend was approximately $226 billion. 2025 was approximately $415 billion. 2026 is approximately $725 billion. Goldman Sachs projects approximately $1.15 trillion cumulative for 2025-2027, and approximately $5.3 trillion cumulative for 2025-2030 across the four largest hyperscalers (2026 dollars). The debt-funded share of capex has risen from 9% to 32% of spend in about 18 months, and a web of circular vendor-financing arrangements between chipmakers, cloud providers, and AI labs is now estimated north of $800 billion.
| Telecom (adjusted to 2026 dollars) | AI hyperscalers | |
| Span | Approximately 1996-2001 (6 years) | 2024-2030 (7 years, partly projected) |
| Total spend | Approximately $2.5 to $2.6 trillion | Approximately $5.3 trillion (projected) |
| Funding trend | Increasingly debt-funded, vendor financing | Increasingly debt-funded, circular vendor financing |
| Physical capacity multiplier | Roughly 100,000x via conduit, cable, fiber count, and DWDM stacking | No hardware equivalent -- Moore's Law has slowed to roughly a 3-year doubling cadence, a single decelerating curve, not a stack of independent multipliers |
| Effective capacity multiplier | Physical/hardware layer only | Software/algorithmic layer -- pre-training efficiency roughly 3x per year, inference cost for fixed capability down roughly 280x in two years |
| Demand vs. supply growth | Supply grew far faster than demand -- 2.7% fiber utilization at peak | Demand currently growing faster than supply -- training compute demand roughly 5x per year vs. roughly 3x per year efficiency gains; hyperscalers report capacity constraints, not idle capacity |
The layered, compounded overbuild I saw firsthand at Nortel has no clean hardware analog in AI. Chip and memory scaling is a single slowing curve, not a stack of independently multiplying physical layers. The closer analog to DWDM is the software/algorithmic efficiency stack -- architecture improvements, quantization, sparsity (mixture-of-experts), distillation, and inference-serving software -- which has compounded at roughly 3x per year for over a decade by rotating through techniques as each one saturates. Some individual levers are already hitting hard limits: KV-cache quantization is reportedly near its information-theoretic (Shannon) bound similar in kind to how DWDM itself hit the nonlinear Shannon limit around 2010 after about 15 years of compounding gains. Whether new AI efficiency levers keep appearing past 2030 the way new DWDM techniques did through the 2000s is genuinely disputed among researchers.
This is the single largest difference from telecom. The 2001-2002 crash happened because supply exploded far ahead of realized demand while heavily leveraged. Today, training compute demand is growing faster (roughly 5x per year) than algorithmic efficiency is reducing the compute needed for a given capability (roughly 3x per year), and hyperscalers report capacity constraints rather than idle infrastructure. Counterintuitively, if AI software efficiency gains slow down, that argues against an overbuild -- it would mean more physical compute is needed per unit of capability, tightening rather than loosening the supply-demand balance. The actual overbuild risk looks more like the opposite: a sudden efficiency breakthrough (a DWDM-style shock) that makes already-built infrastructure far more capable than currently needed faster than demand can absorb it. DeepSeek's 2025 cost-efficiency shock is a small-scale preview of that shape of event, though it did not trigger a broader capacity glut.
SpaceX has filed with the FCC for a solar-powered orbital data center constellation, and Starcloud already flew an Nvidia H100 into orbit in November 2025, running production workloads by February 2026. Musk claims space-based compute could be cost-competitive with terrestrial data centers within two to three years, citing near-free solar power and no land or water constraints. That is disputed: Starcloud projects roughly 20x cost savings over 10 years, while skeptics such as Varda Space Industries put orbital compute at roughly 3x more expensive per watt once realistic launch costs and unsolved cooling and synchronization challenges are included. The whole case depends on Starship launch costs falling sharply, which has not yet happened. If it works, it would be the first genuine hardware-side capacity multiplier in this buildout, easing the demand-supply tightness discussed above, or, if it arrives suddenly at scale, creating the kind of abrupt capacity unlock that produced the telecom mismatch. It belongs in this analysis as a tail risk and opportunity, not a base case.
The mechanics that actually broke telecom -- debt-funded overbuild, vendor financing, and supply running far ahead of demand -- are present in the AI buildout in the first two respects but not, so far, in the third. The physical fiber overbuild had no AI hardware equivalent; the closer analog is software efficiency gains, and those are currently being outpaced by demand rather than outpacing it, which is the reverse of the 2000-2001 setup. The clearest early-warning indicators to watch are not spending totals but a flip in that relationship -- efficiency gains suddenly outrunning demand growth -- combined with the debt and circular-financing exposure built up during this cycle. This is analysis of publicly available data, not a forecast or investment recommendation.
I’m sort or an AI optimist, back in dotcom boom/bust you had dozens of companies that had nothing more than a website engaged in E-Commerce, basically selling things via the internet, which was exploding in use primarily because high speed internet was starting to make inroads and that allowed more and more people to engage in E-Commerce.
I personally remember buying a few books from Amazon back when that was all they did.
Contrast to today, how many major AI companies are there, probably less than a dozen, not that many compared to the dotcom boom/bust.
The other major contrast is today most of the major AI companies are part of major corporations that are hugely profitable outside of their AI expenditures.
Google is hugely profitable, along with Microsoft, Amazon, Meta, Elon Musk’s various companies, according to reports Anthorpic is cash flow positive.
I do think the pace of AI expansion will slow down and perhaps a few of the major AI companies will fail.
I also think the growth of AI will be Agentic or the development of specific AI Agents or applications built on top of the various AI Models.
I think entire new industries will spring up using AI agents as the driver.