Thesis
AI Infrastructure, Semiconductor Cyclicality, and the 2026 Correction Window
This page is for informational and educational purposes only and is not financial advice. 🙂
On This Page
- I. The Core Argument
- II. Why AI Specifically
- III. The Constraint Layer
- IV. The Semiconductor Cascade — Why They Fall in Sequence
- V. Probability Assessment
- VI. The Case Against
- VII. Trigger Framework — Where We Are
- VIII. Expected Timeline
- IX. What This Thesis Is Not
- X. How Newsletters Connect to This Page
I. The Core Argument
This is an early deceleration cycle, not a collapse.
The AI boom that began in 2023 is real. The models work. The hardware is being built. The capital is being deployed. None of that is in dispute.
What is in dispute is the timeline between investment and payoff — and whether current valuations across AI hardware, semiconductors, and hyperscalers already price in a decade of execution that has not yet occurred.
SGGI's position: they do. The gap between what markets are pricing and what physical, financial, and organizational systems can actually deliver in the near term is the setup for a meaningful correction in the late 2026 into early 2027 compression window.
The mechanism is specific. Q3 2026 is the inflection point — not because the correction begins on a fixed date, but because Q3 is when hyperscalers issue forward-looking 2027 capex guidance. The thesis does not require 2027 capex to fall. It only requires 2027 capex growth to fail to scale astronomically versus 2026. That comparison — 2027 versus 2026 — is the first-derivative read that the market has not yet priced.
Growth expectations are normalizing faster than systems can adjust. Markets are repricing the rate of change, not absolute levels. This is not a bubble thesis in the dramatic sense. It is a slope thesis: the rate of capex growth decelerates in the 2027 guide, and that deceleration — already visible in capital behavior, organizational decisions, and constraint data — is not yet priced.
The thesis doesn't require AI to fail. It doesn't even require capex to fall. It only requires next year's growth rate to be lower than this year's — which it always is, on its own schedule.
II. Why AI Specifically
The Circular AI Capital Loop is the mechanism at the center of the thesis. A small group of firms — NVIDIA, Microsoft, Google, Amazon, Oracle, OpenAI — now design, buy, and finance each other's compute. They build the chips, rent the servers, and underwrite the capital cycles that sustain one another's valuations.
Each GPU order, cloud-service contract, and infrastructure loan compounds valuation across the same interconnected balance sheets. What began as a technology race has matured into a self-reinforcing financial ecosystem where one company's expense becomes another's inflated revenue.
This creates two compounding risks. First, the loop inflates valuations without generating proportional external revenue. Second, when any part of the loop slows — GPU demand, cloud contract growth, model spending — the compression is synchronized rather than isolated.
The loop resolves when forward growth stops accelerating. It does not require collapse. It requires the next-year capex guide to come in below the current-year growth rate — at which point the multiples that priced perpetual acceleration begin to compress. That moment is the 2027 guidance cycle, which is why Q3 2026 carries the structural weight in the thesis.
When the same capital is flowing in a circle and everyone calls it growth, the question isn't whether it stops. It's when next year's number stops being bigger than this year's by the same margin.
III. The Constraint Layer
Software scales in weeks. Physical infrastructure does not. That asymmetry is the operational core of the thesis.
The five binding constraints that the AI build-out is running into — not in theory, but in current data:
1. Power and Grid
U.S. data-center load is projected to grow from ~25 GW (2024) to over 80 GW by 2030. Grid interconnection queues are years long. Transformer lead times have extended significantly. Power availability — not chip availability — is now the primary scaling constraint for AI infrastructure. This is broadly acknowledged and not yet fully priced.
2. Capex Outpacing Monetization
Based on earnings reports and analyst projections as of early 2026, the top U.S. hyperscalers are on track to deploy $650–700B+ in 2026 — a near-doubling of 2025 levels. The driver is AI compute, data centers, and networking infrastructure. That is the baseline, not the forecast. Consumer and enterprise AI revenue remains a small fraction of that investment. The gap between capital deployed and revenue generated is the widest it has been at any point in the cycle.
The forward question — and the one that drives the inflection — is what 2027 capex guidance looks like when it is issued in Q3 2026. If hyperscalers guide to continued double-digit growth on top of the 2026 base, the loop holds. If 2027 guidance steps down to single-digit growth or flat — even while remaining historically large — the slope thesis activates. CFO-level ROI scrutiny is rising. Procurement cycles are lengthening. The leading indicators of a guidance step-down are visible. The market has not yet priced the comparison.
3. Semiconductor Cyclicality
No evidence has emerged that AI demand has eliminated semiconductor equipment cyclicality. AMAT, LRCX, and ASML have each experienced 30–60% drawdowns during prior downcycles. Front-loaded accelerator demand, HBM bottlenecks crowding out DRAM, and China's reliance on extended DUV tooling all mirror prior overbuild dynamics. The cycle does not skip this phase.
4. Organizational Absorption
AI capability is advancing faster than enterprises can absorb it. The demo-to-production gap remains wide. Integration costs are high. Internal ROI hurdles have risen. Hiring has filtered rather than collapsed — but enterprise behavior in the current cycle is measurably more cautious than 2024. Adjustment is occurring through delay, phased rollouts, and longer decision cycles.
5. Consumer and Credit Stress
Household debt and delinquency metrics have continued rising. The demand signal that would justify accelerated AI monetization timelines — broad consumer spending growth — is not present. The economic base is narrower than headline figures suggest.
These aren't tail risks. They're current frictions. The distinction matters.
IV. The Semiconductor Cascade — Why They Fall in Sequence
Semiconductor stocks do not fall all at once. They fall in a specific order — the same order, every cycle — because the economic chain that connects them has a fixed direction. Equipment makers feel the slowdown first. Chip designers feel it second. Megacap tech feels it last. Understanding that sequence is critical to understanding both the timing of the thesis and where the buy opportunities appear.
This sequence has repeated in 2001, 2008, 2012, and 2018. It is not a prediction. It is a pattern with a mechanism behind it.
Why Equipment Falls First — AMAT, LRCX, ASML
Semiconductor equipment companies sit at the beginning of the supply chain. When a chip company — NVIDIA, AMD, Broadcom — decides to slow or cancel a fab expansion, the first call they make is to the equipment supplier. Orders are canceled or deferred before a single wafer is affected. Equipment backlog shrinks before revenue does. Guidance gets cut before earnings do.
This is why AMAT, LRCX, and ASML are the leading indicators of the cycle, not the lagging ones. Their order books are the earliest readable signal of what chip companies actually believe about future demand — not what they say on earnings calls.
AMAT currently trades at a forward P/E meaningfully above its five-year average. That multiple prices in continued semicap expansion through 2027 and beyond. SGGI's AMAT put position — November 2026 expiry — is a direct expression of the view that this multiple compresses before the earnings justify it.
The timing maps onto a specific AMAT-related catalyst. Hyperscalers issue forward-looking 2027 capex commentary on their Q2 2026 earnings calls in late July and August — that is the Q3 2026 inflection signal. AMAT's own forward 2027 WFE projection lands with the November 2026 annual filing, when management is required to articulate the year-ahead outlook in formal disclosures. That window is when the 2027-versus-2026 comparison becomes legible at the company level — and the multiple either holds or compresses. SGGI's AMAT put expiry is positioned for that window precisely.
Historically, AMAT, LRCX, and ASML have each experienced 30–60% drawdowns during equipment downcycles. The against-case argues they are more diversified now — service revenue, multi-year backlogs, government reshoring contracts. That is true. It may limit the depth of the drawdown. It does not eliminate the cycle.
Why GPU and Compute Fall Second — NVDA, AMD, AVGO
GPU and compute names feel the slowdown after equipment because their revenue depends on hyperscaler purchasing decisions, not fab expansion timelines. As long as Microsoft, Google, Amazon, and Meta keep ordering GPUs, NVIDIA's revenue holds. The question is what happens when those orders slow.
Hyperscaler capex has nearly doubled from 2025 to 2026 and is on track to exceed $650–700B in the current year. That spend is being justified by AI revenue projections that have not yet materialized at scale. When CFOs start asking whether the ROI justifies continued acceleration into the 2027 plan — and they are already starting to ask — GPU order growth slows before GPU revenue does. The multiple compresses before the earnings do. The 2027 guidance cycle in Q3 2026 is when that question gets answered publicly.
NVIDIA is the load-bearing pillar of the entire AI capital loop. Its market cap has surpassed the combined value of the five largest industrials without producing comparable free cash flow. That gap is belief, not earnings. When the belief adjusts, the adjustment is not gradual.
Why Megacap Tech Falls Last — AAPL, MSFT, GOOG
Megacap tech names fall last because they have the strongest buffers: massive cash reserves, diversified revenue, aggressive buyback programs, and the ability to keep spending even when the market is skeptical. They also have the most narrative support — AI is still seen as a tailwind for Microsoft Azure, Google Cloud, and Amazon AWS regardless of what happens to equipment suppliers.
But they do fall. When AI investment stops looking like a growth engine and starts looking like a cost center, the multiple on cloud and platform revenue compresses. That compression is slower and shallower than equipment or GPU names — but it is real, and it represents the final phase of the cascade.
Megacap names also recover fastest, for the same reasons they fall last: cash, buybacks, and the ability to acquire competitors and consolidate talent during the downturn. This is why GOOG, MSFT, and AMZN appear on the Wave 1 crash buy list rather than the Wave 2 list — the entry window is shorter and the recovery is faster.
Why Energy and Grid Are Not Part of This Cascade
Energy generation, grid infrastructure, and nuclear names — NEE, CCJ, LEU, PWR, ETN — do not belong in a semiconductor correction sequence. Including them would imply the thesis is bearish on them. It is not. It is structurally bullish.
The constraint layer of the thesis argues explicitly that power availability is the binding limit on AI expansion — not chips, not software, not capital. If that constraint is real, then the companies building and supplying that power are not correction victims. They are the correction's beneficiaries. The AI build-out slowing does not reduce electricity demand. The grid still needs to be upgraded. The contracts are already signed. The backlog does not disappear because NVIDIA's multiple compressed.
These names appear on the Wave 1 crash buy list precisely because a broad market correction may create a temporary entry point — broad selling affects everything indiscriminately in the early phase. But that is a tactical opportunity created by macro noise, not a structural decline caused by the thesis. The reason to buy them during the correction is that the market will have sold them for the wrong reason.
If the power grid is the constraint, the companies building it don't fall because AI slows. They become more necessary. The correction is an entry point, not a thesis-driven decline.
| Group | Falls | Bottoms | Recovers | Why |
|---|---|---|---|---|
| Semicap AMAT, LRCX, ASML |
Q4 2026 | Q1–Q2 2027 | Mid 2027–2028 | Order books lead the cycle. Equipment is first canceled, last reinstated. |
| GPU / Compute NVDA, AMD, AVGO |
Q4 2026 – Q1 2027 | Q2 2027 | Late 2027 | Hyperscaler purchasing decisions slow before revenue does. Multiple compresses first. |
| Megacap Tech AAPL, MSFT, GOOG |
Q1–Q2 2027 | Q2–Q3 2027 | Late 2027–2028 | Strongest buffers — cash, buybacks, diversification. Fall last, recover fastest. |
The cascade is not a theory. It is the same sequence that played out in 2001, 2008, 2012, and 2018. The names change. The order does not. Energy and grid are excluded because the thesis is bullish on them — not because they're immune to market noise, but because their structural demand case is independent of what happens to AI hardware multiples.
V. Probability Assessment
SGGI assigns probability bands to outcomes, not certainties to predictions. The thesis is held with conditional conviction — sized for the probability, not the conclusion.
70–75% is not a guarantee. It is a bet worth sizing — and sizing carefully.
VI. The Case Against
SGGI runs the against-case explicitly. These are the scenarios that would invalidate or materially weaken the thesis:
Government backstop / industrial policy
The U.S. will not let AI leadership slip to China. CHIPS Act funding, accelerated depreciation, defense procurement, and potential direct subsidy could sustain AI hardware spending beyond what market economics alone would support. A government-supported plateau is not a correction.
Inference wave as the next demand cycle
Training is expensive and concentrated. Inference is cheap and broad. If inference demand scales faster than expected — through enterprise deployment, edge automation, or sovereign AI buildouts — it could absorb excess capacity and generate the revenue that training economics cannot.
Semicap diversification buffers the cycle
LRCX, AMAT, and ASML are more diversified than in prior cycles. Service revenue, multi-year backlogs, and government-backed reshoring projects could limit downside to a sideways period rather than a correction. Sideways for 18–24 months is not a crash.
Monetization converts faster than modeled
If AI investment converts cleanly into earnings — through enterprise productivity gains, new revenue lines, or margin expansion — the denominator catches up to the multiple. The thesis depends on delayed payoff. Early payoff invalidates it.
A thesis without an against-case is a prediction dressed as analysis. These scenarios are real. The 25–30% probability reflects them.
VII. Trigger Framework — Where We Are
SGGI tracks 28 predefined triggers across three phases. Triggers fire based on evidence, not opinion. The framework is reviewed monthly.
| Phase | What It Measures | Status | Interpretation |
|---|---|---|---|
| Phase 1 Early Cracks (12 triggers) |
Internal structural weakening — power constraints, capex with low ROI, cloud revenue flattening, grid and cooling shortages | 9 fired 3 partial | Foundation weakening on schedule. First-stage cracks visible and consistent with late-cycle overheating. |
| Phase 2 Pre-Crash Accelerators (8 triggers) |
Semiconductor order collapse, megacap earnings cuts, hyperscaler layoffs, GPU procurement freezes, cloud contract cancellations | 0 fired | Cold. The pre-crash window has not opened. The Q3 2026 hyperscaler earnings calls (late Jul–Aug) carry directional 2027 commentary; the AMAT November 2026 annual filing — with its forward FY2027 WFE projection — is the company-level activation gate. |
| Phase 3 Collapse Triggers (8 triggers) |
Fab cancellations, GPU inventory glut, cloud recession, AI infrastructure writedowns, forced hyperscaler capex cuts | 0 fired | Entirely dormant. The crash has not begun. System is still in late-cycle expansion masking structural weakening. |
Phase 1 firing is not the crash. It is the precondition. Phase 2 cold means we are not yet in the break. That distinction is everything for positioning.
VIII. Expected Timeline
Q3 2026 was SGGI's original target for the sentiment and capex inflection — established as part of the initial thesis framework, before FY25 data confirmed it. The crash buy list was determined in December 2025, also before the correction window, so that deployment decisions would be made in calm conditions rather than reactive ones. Neither was retrofitted from events that came later.
To be direct about the sequencing: Q3 2026 was SGGI's original target window, established as part of the core thesis before the evidence base accumulated through FY25. It was not reverse-engineered from data that came later. The crash buy list — Wave 1 and Wave 2 deployment targets — was defined in December 2025, also ahead of the correction window, specifically so that buy decisions would be made in calm conditions rather than under pressure during a drawdown.
That sequencing matters. A thesis that predates its own confirmation is more useful than one assembled after the fact. The monthly newsletters document whether the original framework is holding — not whether a new one needs to be built.
Q4 2025 – Q1 2026
Capital selectivity phase. Discipline visible in organizational decision-making. Phase 1 triggers accumulating. Observed on schedule.
Q2 – Q3 2026 — Signal Accumulation
Hyperscalers report Q1 and Q2 2026 earnings during this window, with management commentary increasingly focused on the 2027 capex trajectory. The Q3 2026 inflection signal prints when forward 2027 commentary on Q2 earnings calls (late July–August) reveals whether the 2027-versus-2026 slope is holding or breaking. Book-to-bill below 0.9 in any major equipment name or any hyperscaler guidance cut at this stage would fire Phase 2 early. Not yet reached.
Q4 2026 — AMAT Annual Filing / Price Action Begins
This is the primary activation gate. AMAT's November 2026 annual filing is when management is required to articulate the forward FY2027 WFE projection in formal disclosures — making the 2027-versus-2026 comparison legible at the company level. Equipment names lead the cascade lower in this window. SGGI's AMAT put expiry (November 2026) is positioned for this catalyst.
Late 2026 – Early 2027
Expected market compression window. Multiple deferred constraints become harder to defer simultaneously. This is the original correction window, unchanged from the December 2025 thesis review.
Crash buy deployment
Buy targets defined in December 2025 — before the correction window — to avoid making buy decisions under pressure. Wave 1 (AMZN, MSFT, GOOG, NEE, CCJ) deploys at 10–15% correction. Wave 2 (PWR, ETN, CEG, LEU, FLNC, GEV, CAT) deploys at 30–50% correction.
The buy list was built in December. The correction hasn't arrived. That's the point — decisions made before pressure are better than decisions made inside it.
This is not a fixed-date prediction. Q3 2026 is not a deadline. It is the point where multiple constraints that can be deferred individually become harder to defer simultaneously. The thesis survives a timing shift. It does not survive a structural invalidation — which is why the against-case is tracked alongside the for-case every month.
Structural theses don't care about your timeline. They care about the evidence. Watch the triggers, not the calendar.
IX. What This Thesis Is Not
This is stated explicitly because the thesis is consistently misread in one direction or the other.
This is not a prediction that AI fails. Model capability continues to improve. The technology is real. The use cases are expanding. None of that is disputed.
This is not a recession forecast. The economy can slow unevenly — as it did throughout FY25 — without entering a formal recession. The thesis is compatible with continued GDP growth alongside valuation compression in specific sectors.
This is not a short-AI thesis. SGGI holds long positions in AI-adjacent names and intends to increase exposure aggressively after the correction window — that is the entire purpose of the December 2025 crash buy list. The thesis is bearish on the multiple, not on the technology.
This is not a guaranteed event. 70–75% probability means a 25–30% chance of being wrong. Position sizing reflects that.
This is not a timing trade. The window is Q3 2026 inflection, late 2026 into early 2027 compression. Anyone treating it as a single dated event is misreading the framework.
X. How Newsletters Connect to This Page
Every monthly newsletter is a progress report against this page — not a new thesis, not a fresh take, not a reaction to the week's headlines.
Each issue advances the same framework at a different layer of visibility:
October established the setup. November showed capability does not equal replacement. January showed capital does not equal absorption. Each newsletter is an evidence layer, written before the first genuine Phase 2 test prints.
When the newsletters reference "the thesis," this is the document they mean. When they note that no probability adjustment has been made, they are being compared against the numbers on this page. When they describe Phase 2 as cold, they are reading from the trigger table above.
The thesis page is the spine. The newsletters are the monthly vertebrae.
If a newsletter ever contradicts this page without explanation, that's the signal something changed. So far, nothing has.
Editor’s note: Last update 25 April 2026 to correct timeline of inflection (Q3 2026) and correction (Q4 2026 to Q1 2027). The inflection point represents when SGGI believes 2027 forecasts for growth may stall or begin to come into question.