A technology revolution can be real while capital returns are overstated. The internet transformed the economy, yet markets still financed and valued a decade of genuine demand too early.
Core AI demand retains hard evidence: TSMC's Q2 2026 revenue and margins reached the top of guidance, Q3 guidance remained strong, and Alphabet and Microsoft continue to expand AI infrastructure investment. At the same time, SOXX and high-beta semis have undergone a meaningful valuation washout.
This report aligns the cycles through a dual timeline, peak-to-trough drawdowns, and a first-principles capital-cycle map, then reduces the 6–12 month decision to six indicators: unit purchases, AI revenue versus depreciation, utilization, order quality, post-earnings price response, and market breadth.
Dot-com vs. AI capital cycles: is this reset 1998 or 2000?
A technology revolution can be real while capital returns are overstated. This report aligns the 1995–2002 internet buildout with the 2022–2028E AI buildout to separate demand, valuation compression, and earnings risk.
The question is not whether AI is real. It is who converts CapEx into cash flow.
The dot-com lesson is not that the internet failed. The industry built, financed, and valued a decade of genuine demand too early. AI now follows a comparable sequence: usage growth, scarcity, a CapEx arms race, supplier margin expansion, synchronized capacity additions, and finally an ROI test.
The cycles are not identical. Today's hyperscalers have strong cash flows, AI revenue is visible, and leading-edge supply is concentrated. TSMC reported $40.2bn Q2 revenue and 67.7% gross margin, with $44.6–45.8bn Q3 guidance. Alphabet guided to $175–185bn of 2026 CapEx while Cloud grew 48%. CapEx also includes buildings, networks, replacement assets, and non-AI workloads, so it cannot all be translated into new AI silicon. Broad cancellations are still not visible.
Align the cycles by capital stage, not calendar year
Netscape's IPO turns the web into a commercial entry point.
ChatGPT makes frontier model capability directly usable.
Usage metrics run ahead of mature economics.
Enterprise trials and copilots lift compute demand.
Networking and optics become the first picks-and-shovels winners.
Compute, foundry, memory, and packaging monetize first.
Debt-funded fiber and data-center duplication.
GPU, ASIC, networking, power, and campuses scale together.
Concept websites, optical small caps, and IPOs rerate.
Neoclouds and second-tier suppliers gain narrative premium.
Liquidity shock; earnings and infrastructure spending keep rising.
Semis and fringe themes delever while the index stays milder.
Liquidity, CapEx, and estimates power a final melt-up.
The fork between a 1998 reset and 2000 distribution.
Cisco took a $2.77bn inventory and purchase-commitment charge in FY2001; Nasdaq ultimately fell 77.9%.
Utilization sustains the cycle; idle capacity triggers estimate cuts.
The index and the theme are not experiencing the same correction
Internet cycle: Nasdaq waves
Current cycle: max drawdown since May 2026
This is not a 2000-style Nasdaq crash, but high-beta semis have already endured a first-crash-sized valuation washout.
How a real technology becomes a capital bubble
1998 continuation
- Unit purchases rise with CapEx
- AI revenue outruns depreciation
- Orders convert into reported revenue
- SOXX regains relative strength
2000 transition
- CapEx rises mostly because prices rise
- Efficiency outruns compute demand
- Deferrals and duplicate orders emerge
- Good earnings produce lower highs
Strong similarities, materially different balance sheets
| Dimension | Dot-com cycle | AI cycle | Implication |
|---|---|---|---|
| Technology | The internet ultimately transformed the economy | AI is entering cloud, software, and R&D workflows | Long-term truth does not validate every price |
| Infrastructure | Fiber, routers, and IDC preceded app profits | GPU, HBM, network, and power precede mature AI profits | Suppliers monetize first and carry CapEx risk |
| Financing | Telecom and startups relied heavily on debt/equity | Hyperscalers self-fund; neoclouds rely more on financing | Lower core systemic risk, fragile fringe |
| Unit economics | Traffic rose while bandwidth prices collapsed | Tokens rise while inference cost falls rapidly | Elastic demand and price deflation coexist |
| Supply | Many suppliers and heavy duplication | Leading GPU, foundry, and HBM are concentrated | Scarcity can last longer, then unwind sharply |
| Evidence now | Orders and IT budgets weakened after the peak | TSMC and cloud results still show growth and constraints | Valuation reset, not confirmed earnings recession |
Six indicators decide the next 6–12 months
Is CapEx buying more servers, or merely absorbing component inflation?
Does monetization outrun depreciation, power, and operations?
A shift from supply shortage to idle compute changes the cycle.
Watch deferrals, cancellations, lead-time compression, and duplication.
Sustained gains support 1998; good news and lower highs support 2000.
SOXX vs. QQQ and leaders vs. fringe determine rebound quality.
Evidence trail
.jpg){kind=link}
Historical returns are close-to-close peak/trough calculations excluding dividends. Current drawdowns use Nasdaq closes from May 1 through July 16, 2026. All 2026E–2028E timeline entries are scenarios, not forecasts. Static research snapshot; not investment advice.