The previous article laid out a method for studying technology revolutions. Unlike mature industries, technology revolutions are supply-constrained first. Leading-edge fabs take three to five years to build. Advanced packaging yields need several quarters to ramp. Critical equipment has delivery lead times of two to three years. No matter how strong the price signal, capacity cannot expand instantly.
That shifts the research focus from forecasting demand to analyzing supply.
The framework has three layers:
- Supply signal: is the bottleneck still there?
- Capital signal: are companies still investing?
- Earnings signal: is commercialization actually materializing?
This article skips the framework introduction and applies it directly to the AI supply chain. One question: where is the AI industry actually in its cycle right now?
I. Layer One: Supply — Is the Bottleneck Still There?
Why look at supply first, not demand?
In traditional industries, demand leads the cycle. Downstream orders rise, companies expand, the boom accelerates — a familiar chain.
In supply-rigid industries, that logic doesn't hold. What determines cycle duration is not how strong demand is, but how fast supply can release. When capacity can't expand with price signals, strong demand only translates into higher prices, not more output. So the first thing to watch is the point in the chain closest to the physical bottleneck.
Why memory chips as the indicator?
HBM and DRAM sit at the very top of the AI supply chain. Their price changes are the earliest reflection of the chain's supply-demand dynamics.
When demand runs ahead of supply release, upstream prices rise first. When supply constraints ease, prices typically turn first. So instead of tracking GPU shipments directly, watch memory prices. They function as the supply chain's thermometer.
The latest data
Memory prices are still rising. Q3 DRAM contract prices are up 13%–18% sequentially, and HBM4 is expected to double next year. The important point is not whether every individual price forecast proves correct. It is that the upward slope has not yet clearly broken — a direct signal that supply still can't keep up with demand.
GPU delivery timelines keep stretching. H100 and H200 spot supply is tight, with lead times pushed to 2027. If demand had clearly weakened, companies wouldn't accept wait times exceeding a year. The stretching lead time itself signals that supply still lags demand.
Compute rental prices hit new highs. The one-year lease price for an H100 rose from $1.70/GPU/hour in October 2025 to $3.09 by end-June 2026 — an increase of over 80%. Lead times tell you whether the constraint has lifted; rental prices tell you how severe the constraint is.
Layer One Conclusion
The supply signal clearly points to Tight. Memory prices keep rising, GPU lead times are stretching, compute rental prices are at all-time highs. The core bottleneck upstream has not meaningfully eased. The AI cycle still has a physical foundation underneath it.
II. Layer Two: Capital — Are Companies Still Betting?
Why is CAPEX the second key indicator?
Supply constraint only tells you things are tight right now. It doesn't tell you they'll stay tight. If companies stop investing, the current squeeze is a short-term phenomenon.
CAPEX answers a different question: are industry participants willing to keep taking risk and expanding capacity?
One easily overlooked point: capital investment growth does not equal immediate supply release.
In traditional manufacturing, higher capex usually means capacity arrives in a few quarters. AI infrastructure is different. New supply has to pass through equipment procurement, installation, process qualification, and yield ramp — a cycle measured in years.
So CAPEX growth actually means future supply will increase, but in the short term it reinforces the boom. Capital spending creates new demand for equipment, components, power infrastructure, and construction before the resulting capacity reaches the market.
The latest data: capital investment is accelerating
SEMI's latest forecast puts 2026 global wafer fab equipment spending at $138.1 billion, rising above $150 billion in 2027. From 2026 to 2028, cumulative 300mm fab equipment spending will reach $374 billion. Memory equipment investment over the next three years will exceed $175 billion.
The primary market is equally active. IPO and SEO proceeds for non-financial corporates sit at historical highs. Risk appetite is spilling from a handful of leaders into a broader range of AI assets.
What does this mean?
Companies aren't pulling back — they're accelerating. More importantly, these investments won't translate into supply immediately. Fabs take three to five years from groundbreaking to volume production. Equipment delivery runs two to three years. Today's CAPEX surge locks in future supply additions, not current capacity release.
That means the current supply tightness is not a short-term phenomenon. It's the structural result of sustained capital investment meeting capacity that can't release quickly.
Layer Two Conclusion
The capital signal clearly points to Expansion. Cloud CAPEX keeps getting revised up, fab equipment spending hits new highs, primary market financing is active. Companies are putting real money behind AI's future — and in the short term, that investment reinforces the supply constraint rather than relieving it.
III. Layer Three: Earnings — Is Commercial Value Materializing?
Why is ARR the core indicator for this layer?
Capital investment can continue for a long time. But if it never converts into commercial revenue, the chain is just a cash-burning game.
ARR (annual recurring revenue) reflects not one-off sales but sustainable, predictable revenue streams. It's the key metric for whether commercialization is genuinely landing.
But note: the earnings signal is still in a validation phase.
Compared to infrastructure investment, AI application commercialization remains early. The more important question right now is not "does AI have revenue" but whether revenue growth can eventually cover the enormous capital investment.
The latest data: commercialization is accelerating, but still early
Leading model providers show strong ARR growth. Anthropic reported ARR of $47 billion as of May 2026, signaling that enterprise AI demand is growing fast.
Major overseas cloud providers' cloud revenue is becoming a new growth curve. CSP capex is the most direct forward-looking indicator of chain momentum; cloud revenue is the leading validation indicator of whether that capex is effective.
AI model token usage is exploding. Weekly call volumes have hit all-time highs. AI coding and video generation are among the scenarios where commercialization is advancing fastest.
What does this mean?
Commercialization is moving from expectation to realization, but hasn't fully matured. ARR growth, cloud revenue expansion, and token usage acceleration are improving simultaneously — the reliability of earnings realization is rising.
But there's still a gap before self-sustaining cash flow that covers investment. Current earnings growth is driven more by capital-fueled demand expansion than by the maturation of intrinsic business models. The harder test comes later: can the economic value generated by AI eventually cover the cost of building the system?
Layer Three Conclusion
The earnings signal clearly points to Improving, but still in validation. Commercialization is accelerating but not yet mature. The industry has moved from a pure investment phase into a phase where investment and realization run in parallel. Whether earnings can durably cover capital investment remains an open question.
IV. Cross-Validation: What Do the Three Signals Combined Tell Us?
Each signal alone can mislead. Only combined do they locate the industry's true stage.
| Stage | Supply | Capital | Earnings | Characteristics |
|---|---|---|---|---|
| Early investment phase | Tight | Rising | Unvalidated | Capital flooding in, commercialization not yet started |
| Current AI | Tight | Rising | Improving | Supply constraint persists, capital expanding, commercialization accelerating but immature |
| Mature phase | Easing | Slowing | Stable | Supply ample, growth decelerating, earnings stable |
All three point positive — but this is not the "golden phase." Because earnings haven't fully matured and commercialization is still accelerating. The weak link is still the third signal.
More precisely, this is the transition from investment phase to realization phase: supply constraint provides the time window, capital investment fills the capacity gap, and earnings realization is validating the business model. That makes the current phase interesting — but also vulnerable to disappointment if earnings fail to accelerate further.
Stage Assessment
Supply constraint persisting + capital investment expanding + commercialization gradually materializing.
The AI industry is currently in the transition from investment phase to realization phase, not the mature phase. This is not the golden phase. It's a window where physical support, capital momentum, and early commercial validation coexist.
V. Bubble Assessment: Industry Cycle and Investment Return Are Not the Same Question
Against the backdrop of a booming chain, AI-related assets are seeing extraordinary trading activity in capital markets. The three-layer framework can locate the industry's stage, but it can't directly answer whether valuations are reasonable.
That requires a supplementary dimension: bubble assessment.
Industry investment expansion: warming, but not yet a bubble
IT equipment investment as a share of GDP sits at the 80th–100th percentile, but overall equipment investment as a share of GDP is only at the 20th–40th percentile. Corporate debt growth is at the 0th–20th percentile. The current boom is not yet primarily sustained by broad corporate leverage. This is a clear contrast with the dot-com era, when telecom carriers borrowed heavily to lay fiber.
Capital and sentiment: active but not out of control
Primary market financing is at historical highs, but volatility indicators remain in neutral territory. The market has not yet entered a state of full euphoria, retail-ization, or lottery-ticket trading.
Valuation: clearly elevated
U.S. equities — especially the tech giants — have already priced in years of future growth. The margin of safety on returns has narrowed significantly. As expectations move further into the future, the margin for error becomes smaller.
Bubble Assessment Summary
Overall assessment: industry investment expansion is in a warming phase (not yet a bubble), market sentiment is active but not out of control (not yet dangerous), but valuations are already elevated (return expectations need to come down).
An important distinction
Industry cycle and investment return are not the same question.
An industry moving in the right direction doesn't mean every entry point produces excess returns. The industrial trend determines direction. Valuation determines returns. An industry can continue to improve while future investment returns deteriorate. The technology can work, demand can rise, profits can increase — and the stock can still underperform because the valuation started too high.
The current state of the AI supply chain: industry fundamentals remain strong, but capital markets have already priced in that strength. That means even if the industry keeps improving, investment returns may be compressed by excessive valuations.
These are two questions. They need to be considered separately.
VI. What the Case Study Tells Us
Combining the three signals, the AI supply chain currently shows:
Core Conclusion
The industry is strong. The valuation is expensive.
Two questions. Separate them.
The first tells us what the industry is doing. The second tells us what investors may earn from owning it. Industry fundamentals determine direction. Valuation determines returns.
Without the three-layer framework, you might fall into one of two extremes: seeing memory prices spike and calling "AI overheated," or seeing capital pour in and declaring "the trend is eternal." The framework's value is that it forces you to watch all three dimensions simultaneously, avoiding the misjudgments that come from any single signal.
- Supply constraint tells you: the physical support for the boom is still there.
- Capital signal tells you: the trend's momentum is still there, and won't ease in the short term.
- Earnings signal tells you: value realization is happening, but hasn't completed.
Stacking the three produces the "transition phase" judgment. It's not a simple answer to "does AI have a bubble" — it's a separation of two different questions: how are the industry fundamentals, and is the valuation reasonable?
Find the bottlenecks. Watch the capital. Wait for the earnings.
Closing
AI is not the research objective. It's the test case for the framework.
The question that matters is not "how much longer can AI rise." It's whether, when the next technology revolution arrives — whether new energy, biotech, or quantum computing — we have a method for understanding it.
Through this three-layer signal framework, what we see in the AI industry is not a simple "bubble" or "golden track." It's a complex system in the middle of supply constraint, capital expansion, and commercial realization.
The goal is not to predict markets, but to understand them.
Studying an industry is not about chasing the hottest theme. It's about understanding how value moves through the supply chain.
*Magic Econ · Common Sense Jack*
This article is based on public market data and publicly available research materials (sources include TrendForce, SEMI, Omdia, SIA, WSTS, and company filings). It is intended to validate the effectiveness of the industry research framework through case analysis and is for educational exchange only. It does not constitute investment advice.
If this case study sharpened your research framework, follow Magic Econ
