Jordi Visser — AI Capex Cycle Investment Thesis

Source: Your Capex Is My Opportunity: The AI Capex Cycle Has Replaced the Old Economy Business Cycle, 22V Research / Jordi Visser, May 3, 2026.

The Framework: The Five-Layer Cake AI Business Cycle

Visser's organizing model: the old consumer/industrial business cycle is obsolete. A new AI capex cycle has replaced it, with distinct phases matching traditional early/mid/late cycle dynamics.

Cycle Phase	Layer	Named Components	Investment Implication
Early	Rack / Advanced Packaging	NVDA, TSMC, Intel, Samsung	Already running; hoard phase
Early	Optical Fiber	Corning (GLW), Coherent (COHR)	Lock-in contracts (Meta's $6B Corning deal)
Mid	Chemicals	Entegris (ENTG), 17-name basket	Lower volatility, still early innings
Mid/Late	Power + Data Centers	Caterpillar (CAT), Bloom Energy (BE), Exxon, Chevron	Tripling by 2030; White House invoking Defense Production Act
Late	Agents / Humanoids / Apps	Speculative	Future cash flows; not yet priced

"The prior 15 years was about 'your margin is my opportunity.' The next 15 years is about 'your capex is my opportunity.'"

Investment Thesis #1: Benchmark Arbitrage — The S&P Is Structurally Wrong for the AI World

"The benchmarks are completely wrong for the AI world. They are still built around a world of software and partly the industrial and the consumer-based economy which is gone right now because this is being driven by AI."

The argument: Traditional portfolios — S&P 500, RAIA models, mutual fund benchmarks — are still overweight the prior cycle's winners: software (Salesforce, Adobe, IGV), hyperscalers, and consumer-led growth names. These are the spenders in the AI economy. The actual winners of the AI capex cycle are the receivers: semiconductors, optical fiber, power equipment, chemicals, and infrastructure. The S&P's weighting doesn't reflect this shift. Semis are now the largest Level-2 GICS sector at $10 trillion, yet most benchmark-following portfolios are underweight.

The consensus reads the all-time-high S&P as proof the portfolio is working. Visser shows that ~50% of S&P stocks are down YTD, the median stock is 13% below its 52-week high, and software (IGV) is near YTD lows. The index is being carried by a narrow set of physical buildout names that most portfolios barely own.

Trigger: Any quarter where software earnings or gross margin assumptions begin to disappoint while semi/power/chemical earnings accelerate — this validates the cycle rotation and forces benchmark reweighting.

Names: Overweight NVDA, TSM, GLW, COHR, CAT, ENTG, BE, TXN (power semis). Underweight Salesforce (CRM), Adobe (ADBE), ServiceNow (NOW) — all still in legacy benchmarks at elevated weights.

Investment Thesis #2: Nvidia at a 10-Year Forward PE Low Is the Clearest Valuation Anomaly in the Market

"Caterpillar is trading at 36 times next year's earnings. Nvidia is at 10-year lows in their forward PE. I don't think this is going to happen this time. Multiples are going to stay at higher levels."

"Micron is a perfect example... something like four and a half times next year's earnings. The product is in shortage. Earnings are being revised sharply higher."

The argument: Caterpillar — widely understood as a cyclical industrial — trades at 36x forward earnings because the market is correctly pricing in AI-driven demand through 2030 (power generation equipment sales forecast to triple). Meanwhile, Nvidia — the enabler of everything Caterpillar is building for — trades at a decade-low forward PE. Micron, with HBM in structural shortage and estimates being revised sharply higher, trades at just 4.5x next year's earnings. The market is repricing industrials correctly and leaving semis behind.

Bear case on Nvidia centers on efficiency gains reducing compute demand. Visser dismisses this: all efficiency gains over the last three years have increased total demand, not replaced it. The hyperscaler RPO backlog ($1.3T across AWS/Google/Microsoft) is contracted revenue they cannot yet deliver because of compute shortages. There is no efficiency scenario that makes $1.3T of backlog disappear.

Trigger: Nvidia forward PE expanding back toward historical norms as earnings path becomes undeniable; Micron earnings beats + ASP guidance for HBM; Caterpillar maintaining 36x PE while NVDA re-rates upward.

Names: NVDA (Tier 1 — cheapest it has been in a decade relative to earnings growth), Micron (MU) (Tier 1 — 4.5x forward earnings, HBM product in shortage), Caterpillar (CAT) (Tier 1 — $62B backlog, power equipment 3x by 2030).

Investment Thesis #3: Power Equipment Has a Structural 3x Ceiling — Caterpillar and Power Semis Are the Trade

"US spending on power plant equipment expected to triple through 2030s. White House is putting in Section 303 of the Defense Production Act to help with grid infrastructure, equipment, and supply chain capacity."

"Strong demand for large power generation equipment used in data centers supporting AI. We now expect sales of our power generation equipment to be roughly three times larger by 2030." — Caterpillar management

The argument: Caterpillar's $62B backlog gives "excellent visibility into demand over coming quarters." Power generation for data centers is explicitly growing 3x by 2030 per management guidance. The White House has invoked Section 303 of the Defense Production Act — a wartime-era provision — to address grid infrastructure shortfalls. Power semiconductor lead times have stretched to extremes; six power semi names tracked by Visser just broke out to 5-year highs. This is not cyclical demand; it is structural, government-backed buildout of the physical layer that all AI runs on.

Most investors look at CAT at 36x PE and assume it's overbought. Visser argues this multiple is correct and appropriate given the visibility. The real valuation anomaly is that the power equipment and power semiconductor layer hasn't re-rated to match the compute layer's urgency.

Trigger: Caterpillar quarterly backlog updates; US power plant construction permit data; power semiconductor lead time disclosures from distributors; any additional Defense Production Act invocations.

Names: Caterpillar (CAT) — $62B backlog, 3x power gen by 2030; Texas Instruments (TXN) and 5 additional power semiconductor names (subscriber-level); Bloom Energy (BE) — fuel cells in the power layer; GE Vernova (GEV) — implicitly within the same 3x power buildout thesis.

Investment Thesis #4: Chemicals Are the Hidden Mid-Cycle AI Layer — 17 Names With Little Analyst Coverage

"Chemicals are not thought of as AI names, but they are."

"Chemicals will go up with power semis. They fit into the exact same side."

The argument: Advanced semiconductor packaging (CoWoS, HBM stacking) and optical fiber deployment both require specialty chemicals and polymers at scale. Visser has a 17-name chemical basket scoring companies by exposure across advanced packaging, fiber tubing, and polymers. Chemicals are mid-cycle — they follow the early semiconductor buildout but precede the late-cycle application layer. Because they're not categorized as "AI names," they are absent from most AI investment frameworks and remain under-owned. Entegris (ENTG) — semiconductor-specific chemicals and materials for advanced nodes — is explicitly in Visser's thematic basket. Another name ("Camores," exact ticker unconfirmed in transcript) is up ~50% and expected to potentially double.

Zero AI analysts cover specialty chemicals. The Terafab chemical demand Elon Musk described — the raw materials needed for chip factories at that scale — "is not an oil trade." Chemicals embedded in advanced packaging processes and optical fiber deployment have a structural demand growth story that is independent of software revenue.

Trigger: Entegris and peers reporting earnings with advanced packaging volume data; chemical names showing chart correlation with power semis (as Visser demonstrated with Texas Instruments overlay); any Terafab-related procurement announcements.

Names: Entegris (ENTG) — semiconductor materials, chemicals for advanced packaging; "Camores" (exact ticker unclear from transcript; small/mid-cap chemical materials, up ~50%); European chemicals basket (chart overlay with TXN confirmed by Visser).

Investment Thesis #5: Software's Three-Step Death March — The Benchmark Underweights the Short

"First the market compresses the multiple. Then the fundamentals start disappointing. What usually happens next is an earnings miss... analysts still have models that assume things like roughly 80% gross margins in '27, '28, '29 — very little structural change in profitability. Those are too static for an AI-disrupted world." — Adam Parker (cited by Visser)

"Every single tech person that said this is a mispricing, this is panic — well, now you've had a fall, you've had a rally. Adam Parker's right in my opinion."

The argument: Software companies (Salesforce, Adobe, ServiceNow, IGV constituents) are priced at 80%+ gross margin permanence through 2029. Parker's sequence: (1) multiples compress first as AI disruption becomes visible, (2) earnings miss follows as companies spend more to remain competitive, (3) sales miss follows as AI-native alternatives displace legacy products. The sequence has already started — IGV is near YTD lows while the S&P makes all-time highs. Naval Ravikant cited on Apple's terminal value in a similar framework: the certainty of Apple's cash flow model is broken by an AI hardware transition.

Most "buy the dip in software" calls assume this is a temporary rotation. Visser argues this is a structural re-rating — not a cyclical one. The benchmarks actively mask how bad the damage is because the software underperformance is averaged into index returns. Every RAIA holding the S&P has embedded software risk that is not being flagged.

Trigger: First major software company (Salesforce, Workday, ServiceNow) reporting gross margin below 75% or guiding to accelerated AI spend; any Salesforce or Adobe announcement of revenue displacement from AI-native alternatives.

Names: Short/avoid Salesforce (CRM), Adobe (ADBE), ServiceNow (NOW) — benchmark overweights; avoid Apple (AAPL) — terminal value risk flagged by Naval Ravikant; IGV (iShares Expanded Tech-Software ETF) — the collective expression of the overweighted legacy software benchmark.

Ecosystem Map

Primary long positions (Visser's thematic basket, named directly):

NVDA — cheapest on forward PE; core to entire five-layer cake
Caterpillar (CAT) — $62B backlog; power gen 3x by 2030; AI-driven industrial
Micron (MU) — 4.5x forward; HBM in shortage; estimates revised sharply higher
TSMC (TSM) — one of three fab end results for global AI compute
Corning (GLW) — optical fiber; Meta's $6B committed purchase
Coherent (COHR) — optical fiber and photonics infrastructure
Entegris (ENTG) — semiconductor chemicals; advanced packaging layer
Bloom Energy (BE) — fuel cell power for data centers; within power layer
Texas Instruments (TXN) — power semis; confirmed 5-year breakout
Entegris (ENTG) — bought by Visser after earnings drawdown
Bitcoin (BTC) via MicroStrategy (MSTR) — inflation hedge; calls purchased
Ethereum (ETH) — Clarity Act catalyst; 200-week MA $2,456 is the trigger level

Structural short / underweight:

Software (CRM, ADBE, NOW, IGV) — three-step deterioration in progress
Apple (AAPL) — terminal value risk from AI hardware disruption
Hyperscalers (AMZN, GOOG, MSFT) — the "spenders"; underperforming since 2025

Macro watches (neither long nor short, but catalysts):

Chevron (CVX) — unchanged despite oil $66 → $90; lagging but notable for energy rotation
Airlines — first leveraged victim if inflation/fuel prices spike further

Key Risks

Inflation spike breaks the cycle: Oil nearing $90, gas futures at $87 for September, crop prices breaking out, 30Y UK gilts + 10Y JGB at new highs — "I think this is the 1970s." CPI above 4-5% forces rate hike risk and damages P/E multiples across the board.
Market breadth collapse: Median S&P stock 13% below 52-week high; 50% of S&P stocks down YTD; financials still below 200DMA. Market is being held up by very few names — vulnerable to de-risking.
Overbuild narrative gains traction: "The number one reason people say this is an overbuild is because magically we're going to come up with some efficiency gain." Visser dismisses this but acknowledges the bear case drives positioning.
Turbulence model cross-asset signal: Visser's proprietary turbulence indicator issued its first warning; a trigger would precede a deleveraging event similar to early February.
Compute scarcity prevents revenue realization: Anthropic forecasting $50-100B by year-end but White House officials worried it can't access enough compute to serve the demand. This is a risk to the entire AI revenue acceleration thesis.
China compute independence: Elon Musk warned the US could soon produce more chips than it can turn on, while China does not have the same constraint. Export control policy is binary.
Software benchmark drag: If legacy software holdings (Salesforce, Adobe) in index-tracking portfolios undergo Adam Parker's three-step deterioration (multiples → earnings → sales misses), broad index exposure amplifies the downside. Visser's thesis works best when overweighting the physical buildout names relative to the S&P benchmark weight.

Investment Opportunities at a Glance

Tier	Name / Category	Core Thesis	Conviction Signal
1	NVIDIA (NVDA)	10-year low forward PE; compute shortage confirmed; backlog $1.3T across hyperscalers	"Nvidia has gotten cheaper and cheaper" despite 3× demand growth
1	Caterpillar (CAT)	$62B backlog; power gen equipment 3× by 2030; AI capex mid-cycle beneficiary	Management guide: "three times larger by 2030" — explicit and dated
1	Micron (MU)	4.5× forward earnings; HBM in shortage; estimates revised sharply higher	Adam Parker: "Micron is a perfect example… something like 4.5× next year"
2	Corning (GLW)	Optical fiber early-cycle lock-in; Meta $6B committed; Jensen supply chain investment	Long-term take-or-pay contracts validate demand through 2030
2	TSMC (TSM)	One of three global fab end-results; severe shortage of advanced processes	"Severe shortages of Taiwan semis — pushing an all-out effort to build"
2	Texas Instruments (TXN)	Power semis at 5-year breakout; AI servers driving extreme lead-time stretch	Chart: broke 5-year resistance; lead times "extremely stretched" per report
2	Entegris (ENTG)	Semiconductor chemicals for advanced packaging; AI chemical mid-cycle layer	Visser bought after earnings drawdown; 17-name basket confirmed
2	Bloom Energy (BE)	Fuel cells in power layer; named alongside Exxon/Chevron	Explicitly placed in Visser's five-layer cake power tier
3	Coherent (COHR)	Optical fiber photonics; Jensen direct supply chain investment	Named alongside Corning as optical early-cycle lock-in
3	Chemical basket (17 names)	Advanced packaging + fiber tubing + polymers; not in AI benchmarks	"Chemicals are not thought of as AI names, but they are"
3	MicroStrategy (MSTR)	Bitcoin proxy; inflation hedge; calls bought for year-end	Trigger: CPI above 3-month yields
4	Ethereum (ETH)	Clarity Act + programmable money + stablecoin (DoorDash model)	Two closes above 200-week MA ($2,456) is the specific signal

Monitoring Checklist

Bottom Line

Nvidia at a 10-year forward PE low while Caterpillar trades at 36× is the single clearest valuation anomaly in the market. CAT correctly prices in AI-driven demand through 2030; NVDA has not re-rated to match. The market is doing the opposite of what logic requires. Own NVDA; the Caterpillar PE is the floor, not the ceiling.
Chemicals are not in a single AI investor's model — and they should be. Advanced packaging (CoWoS, HBM stacking) and optical fiber both require specialty chemicals at scale. Entegris (ENTG) is the most directly named play. The entire 17-name basket is mid-cycle, lower volatility, and effectively invisible to anyone running an AI portfolio. Buy the dip on earnings.
The benchmark is structurally short the AI world. Every RAIA or mutual fund holding the S&P 500 at market weight owns too much Salesforce, Adobe, and Apple — and too little semiconductors, optical fiber, power equipment, and chemicals. "Benchmark arbitrage" is not a trade; it's a structural reallocation. The window to close this gap is before the software earnings misses begin.
Micron at 4.5× forward earnings with HBM in structural shortage is the cheapest AI infrastructure stock in the market. This is not a meme trade — it is the market incorrectly pricing a commodity cyclical when the product is, structurally and physically, in shortage until 2027–2028.