A chart recently circulated by Christian Mueller-Glissmann, portfolio strategist at Goldman Sachs, delivered the kind of quiet shock that forces you to recalibrate your entire macro model.
It examined the distribution of equity sensitivities within the S&P 500. The finding: the proportion of S&P 500 constituents exhibiting a negative beta to the index has spiked toward 50%.
In plain English, half the large-cap equities in the United States now systematically zag when the benchmark zigs. They are no longer merely moving slower or faster than the market—they are actively moving in the opposite direction. As Mueller-Glissmann laconically noted, this phenomenon of widespread negative equity beta is “very similar to the Tech Bubble.”
As Toby Nangle highlighted over at FT Alphaville, the beta for the top-eight US mega-caps has drifted to 1.23, while the remaining 492 stocks in the S&P 500 carry an average beta of just 0.87. If you run an equity portfolio benchmarked against the S&P 500, you face a brutal dilemma: either you match the mega-cap tech cohort at its dizzying market-cap weight, or you find yourself running an accidental low-beta, defensive portfolio.
This is not normal market behavior. It is the signature of extreme structural distortion.
When you look back at the late 1990s through March 2000, you find the exact same mechanical footprints. But how deep does the analogy go? Is the 2024–2026 AI infrastructure boom a rerun of the 1999–2000 telecom and dot-com bubble, or do fundamental cash flows justify the divergence?
Let’s dismantle the mechanics across breadth, capital expenditure, vendor financing loops, valuation multiples, and monetary policy.
1. Market Breadth & Concentration: Beyond 1999
The dominant defense of the current bull market is liquidity and index performance. The headline S&P 500 and Nasdaq push forward, leading casual observers to assume broad economic health.
However, technicians and quantitative strategists remember the warning sign of 1999: the Advance-Decline (A/D) Line divergence.
Concentration Peak (Top 10 S&P 500 Weight)
-----------------------------------------------------------
1990 Historical Baseline: ~17%
Dot-Com Peak (March 2000): ~27% – 29%
Great Financial Crisis: ~20%
2024-2026 AI Expansion: ~38% – 41%
-----------------------------------------------------------In late 1998 and throughout 1999, the cumulative Advance-Decline line peaked and deteriorated sharply even as the cap-weighted S&P 500 and Nasdaq Composite sprinted higher. The rally narrowed down to a shrinking cabal of infrastructure and networking giants—Cisco, Lucent, Nortel, Microsoft, and Intel. The median stock was already entering a quiet bear market well before the index crested in March 2000.
In 2025 and 2026, this divergence has repeated in an even more concentrated format:
- Top-10 Weight: At the peak of the dot-com bubble, the top 10 stocks accounted for roughly 27% to 29% of the S&P 500. Today, the top 10 represent between 38% and 41% of the entire index—a concentration unseen since the early 1930s.
- Equal-Weight (RSP) vs. Cap-Weight (SPY): In 1999, the cap-weighted S&P 500 outstripped the equal-weighted index by double digits as capital concentrated into large tech. When the dot-com bubble burst, the dynamic inverted: during the “Lost Decade” (2000–2009), the cap-weighted S&P 500 posted a cumulative total return of approximately -9%, while the equal-weighted S&P 500 returned over +60%.
- The “Negative Beta” Symptom: Because capital flows are increasingly channeled through passive market-cap-weighted vehicles (ETFs, target-date funds), liquidity concentrates disproportionately in the top cohort. When macro shocks hit or tech rotates, non-tech constituents are treated not as fellow equities, but as funding sources or shock absorbers.
2. The Capex Treadmill: Telecom Fiber vs. GPU Clusters
Bubbles are rarely born from sheer fiction; they are almost always born from real, revolutionary technologies paired with an unsustainable capital cycle.
In 1998–2000, that technology was the commercial Internet and optical telecommunications. The narrative was unassailable: internet traffic was supposedly doubling every 100 days, and the world needed infinite fiber-optic bandwidth. Telecom operators—WorldCom, Global Crossing, Qwest, 360networks—embarked on an unprecedented multi-hundred-billion-dollar infrastructure binge.
The suppliers of that infrastructure—chiefly Cisco Systems, which became the world’s most valuable company in March 2000 at a market capitalization north of $550 billion (approx. 5.5% of US GDP)—could not manufacture routers and switches fast enough.
Fast forward to 2024–2026:
| Indicator | Dot-Com Era (1999–2000) | AI Infrastructure Era (2024–2026) |
|---|---|---|
| Infrastructure Anchor | Cisco Systems (Routers/Switches) | Nvidia (Accelerated Compute/GPUs) |
| Peak Valuation Scale | Cisco: ~$550B–$569B (~5.5% US GDP) | Nvidia: ~$5.4T–$5.5T (~11.7% US GDP) |
| Big Hyperscaler/Telco Capex | Telecom Capex peaked at ~$100B–$120B/yr | Hyperscaler Capex: ~$250B (2024) $\rightarrow$ ~$410B (2025) $\rightarrow$ $700B–$800B+ (2026) |
| Capex vs. Free Cash Flow | Telecoms took on massive junk debt | 2026 Capex now eclipses combined operating cash flows for leading hyperscalers |
| Physical Bottleneck | Laying trench/rights-of-way fiber | Power & Megawatts (Grid interconnection, nuclear PPAs) |
The numbers in 2026 dwarf the late 1990s. The combined capital expenditures of the primary hyperscalers—Amazon, Alphabet, Microsoft, and Meta—have ballooned from ~$250 billion in 2024 to more than $700 billion in 2026.
Crucially, projected 2026 CapEx is now beginning to outstrip operating cash flow for several hyperscalers. To sustain the buildout, tech titans that previously sat on pristine net-cash fortresses are issuing long-term corporate debt, executing leaseback structures, and underwriting private credit facilities.
3. The Return on Investment Problem: The $600 Billion Question
The critical question for any capital expenditure boom is simple: Where is the end-user cash flow to pay for it?
In June 2024, David Cahn, partner at Sequoia Capital, published his seminal analysis, AI’s $600B Question. Cahn pointed out the arithmetic dilemma:
- Take the annualized revenue run-rate of chip hardware providers.
- Multiply by 2x to account for total data center cost of ownership (cooling, energy, land, backup power, optic transceivers).
- Multiply by another 2x to grant the buying software layer a viable 50% gross margin.
By mid-2024, Cahn estimated that the software ecosystem needed $600 billion in incremental annual revenue just to justify the hardware CapEx run-rate. By 2026, with hyperscaler CapEx exceeding $700B annually, that required payback run-rate has climbed toward the trillion-dollar mark.
This tension was articulated within Wall Street itself in the landmark Goldman Sachs report, Gen AI: Too Much Spend, Too Little Benefit?.
In the report, Jim Covello, Goldman Sachs’ Head of Global Equity Research, laid out the fundamental flaw in the comparison:
Unlike historical transitions where expensive human workflows were replaced by low-cost automated solutions (such as PCs replacing typist pools or spreadsheets replacing accounting books), generative AI attempts to replace relatively low-wage human labor with ultra-expensive, power-hungry compute.
Covello compared the economics to “building a nuclear reactor to power a toaster.” If a technology is astronomically expensive to run, it cannot survive on low-value summarization, code autocompletion, or chat widgets. It requires high-margin, mission-critical autonomous problem-solving—a capability still hindered by reasoning horizons and hallucinations.
In 2000, Cisco’s revenue did not decline because businesses stopped using the internet; it cratered because their telecom customers ran out of money and realized they had overbuilt capacity by five to seven years. Between 2001 and 2003, an estimated 85% to 95% of the optical fiber laid in the US remained “dark” (unlit).
If enterprise AI adoption scales linearly while hardware infrastructure scales exponentially, the hyperscaler depreciation wall looms as an inevitable margin compression event.
4. Vendor Financing and the Circular Revenue Engine
One of the least understood facets of the dot-com collapse was vendor financing.
In 1999, equipment manufacturers like Lucent Technologies, Nortel, and Cisco loaned billions of dollars directly to startup competitive local exchange carriers (CLECs). The carrier used the loan to purchase routers from the vendor; the vendor booked immediate revenue and profit, inflating their share price, which allowed them to issue more stock-backed credit. By late 2000, vendor financing exposure reached roughly $25 billion. When the CLECs defaulted, Lucent and Nortel were decimated, with Nortel ultimately collapsing into bankruptcy.
In the 2024–2026 AI expansion, this mechanic has reemerged in a more sophisticated, equity-and-cloud format:
- The GPU Neocloud Flywheel: Hardware vendors invest hundreds of millions into specialized cloud infrastructure operators and AI model developers. Those companies immediately turn around and use those balance sheet commitments to purchase the vendor’s latest GPU architecture.
- Cloud Credit Round-Tripping: Hyperscalers invest billions into foundation model labs (e.g., billions invested by Big Tech into OpenAI, Anthropic, Mistral), which is largely returned directly to the investor in the form of dedicated compute hours on their proprietary cloud platforms (Azure, GCP, AWS).
- Asset-Backed Debt Facilities: Compute clusters are now routinely collateralized. Companies borrow against their hardware assets to buy more hardware, creating a leverage layer that assumes hardware values do not suffer sudden technological obsolescence when new architectures debut.
While the balance sheets of today’s mega-caps are vastly superior to Lucent or WorldCom, the circularity of revenue recognition creates an artificial demand signal that obscures organic, end-market adoption.
5. The Physical Constraint: Speed to Power
In 1999, the ultimate bottleneck was physical right-of-way: digging ditches alongside highways and railroad tracks to lay conduit.
In 2026, the constraint has shifted to electrons and baseload power.
Data centers cannot run on intermittent renewables alone; AI training and inference clusters require 24/7/365 baseload electricity. With transmission grid interconnections facing multi-year queues across PJM, ERCOT, and CAISO, tech hyperscalers have been forced into unprecedented capital commitments:
- Constellation Energy entered into a landmark 20-year power purchase agreement with Microsoft to resurrect the shuttered 835-megawatt Unit 1 reactor at Three Mile Island, rebranded as the Crane Clean Energy Center.
- Amazon expanded its nuclear energy partnerships, including massive agreements to source gigawatts of direct power from commercial nuclear stations like Talen Energy’s Susquehanna facility.
When hyperscalers are compelled to underwrite 20-year nuclear restarts and small modular reactor (SMR) development just to feed compute clusters, capital intensity ceases to be an asset-light software game. It transforms tech into a heavy capital-goods utility.
6. Valuation Multiples: Where Do We Stand?
Are valuations as crazy as 1999? The answer is nuanced:
Valuation Metric Comparison
-----------------------------------------------------------------
Metric Dot-Com Peak (1999-2000) Current (2026)
-----------------------------------------------------------------
Shiller CAPE (S&P 500) 44.2 (Dec 1999) ~40.6
Top-10 Median Forward P/E ~30x – 35x ~32x – 38x
Headline Leader P/E Cisco >150x Nvidia ~30x – 35x
Tech Sector Unprofitable % >35% of listed tech <15% of top tiers
------------------------------------------------------------------ Shiller CAPE Ratio: The cyclically adjusted P/E ratio for the S&P 500 sits at ~40.6, higher than the 1929 peak (~32.6) and second only to the all-time record of 44.2 reached in December 1999.
- Torsten Slok’s Median Metric: As Apollo Global Management Chief Economist Torsten Slok demonstrated, if you look at the median forward P/E of the top 10 companies, today’s market is actually more expensive than the median of the top 10 in 1999. In 2000, the average was skewed by insane outliers like Cisco (150x+) and Yahoo, but several large-cap leaders (like Exxon, GE, Wal-Mart) traded at lower multiples. Today, premium multiples are ubiquitous across the entire leadership cohort.
- The Multiple Trap: Bulls point out that Nvidia trades at ~30–35x forward earnings, nowhere near Cisco’s 150x multiple. But this ignores the denominator: Cisco’s ‘E’ (earnings) looked reasonable right up until quarterly capex cuts vaporized its order book overnight. A multiple of 30x on earnings that are peak-cycle and debt-financed is not cheap.
7. The Macro Setting: The Fed Easing Echo
The final, eerie parallel lies in the Federal Reserve’s posture.
In late 1998, following the Russian debt default and the collapse of Long-Term Capital Management (LTCM), Alan Greenspan’s Federal Reserve cut interest rates by 75 basis points to stabilize liquidity. That liquidity injection acted as high-octane fuel poured directly onto an already simmering tech rally, propelling the Nasdaq on its historic vertical climb through late 1999 and early 2000. It was only when the Fed reversed course and hiked rates six times (reaching 6.5%) that the liquidity sponge was wrung out, puncturing the bubble.
In 2024–2026, after raising rates to tame post-pandemic inflation, the Federal Reserve embarked on an easing cycle into an economy with record asset prices, tight credit spreads, and massive fiscal deficits. By lowering the cost of capital while market concentration is at an all-time high, monetary policy has once again provided an exogenous tailwind to the speculative frontier.
Conclusion: How the Playbook Unfolds
The lesson of March 2000 is not that the Internet was a fad. The Internet transformed global commerce, created trillion-dollar ecosystems, and lived up to every techno-optimist vision of the 1990s.
The lesson is that the timing of the economic payoff was separated from the capital investment by almost a decade. The companies that built the physical pipes in 1999 were wiped out, while the capital was reallocated at fire-sale prices to the next generation of builders (Google, Amazon, Netflix) years later.
When nearly 50% of the S&P 500 has a negative beta to the index, the market is broadcasting that it is no longer an economic index—it is an AI capital-cycle derivative.
For investors, the playbook of 2000–2002 offers a clear roadmap:
- The “Zaggers” are your defense: When mega-cap momentum broke in 2000, value, dividend-growth, industrial, and equal-weight strategies crushed the benchmark for years.
- Watch the Capex inflection: The turn will not come from Fed commentary; it will come when the first major hyperscaler cuts infrastructure guidance due to free-cash-flow exhaustion.
- Depreciation is real: The trillions spent on H100s, Blackwells, and data centers must eventually be amortized. When those depreciation charges begin hitting quarterly income statements without offsetting software revenue, multiple compression follows.
The market has given us its warning in the beta distribution. The only question is whether investors are willing to heed it before the zigs become zags.