2026 Analytical Notes:
Tracking AI Infrastructure Trends.
A personal tracking framework for how AI demand moves through physical and financial constraints.
The AI cycle is moving from exposure to conversion. In 2026, the harder question is whether AI demand can become deliverable capacity, recognized revenue, margin evidence, and cash-flow quality.
Core Thesis: From Exposure to Conversion
Exposure answers a simple question: whether an issuer is close enough to the AI theme to be included in demand narratives, customer planning, or capital spending cycles. Conversion asks a harder question: whether that exposure becomes deliverable compute, energized facilities, recognized revenue, margin evidence, and cash-flow quality that can survive an investment-heavy period.
AI-related demand is no longer visible only in narrative; it increasingly appears in corporate capital budgets, supplier commentary, data-center planning, power procurement, and segment-level disclosure. The harder work is tracing whether that demand actually moves through the physical and financial system.
That is the purpose of this analytical note. It is not ranking securities, assigning price targets, or forecasting market returns. It defines the structural variables that matter when AI demand meets real-world constraints: semiconductor capacity, memory bandwidth, advanced packaging, power availability, data-center delivery, software monetization, and financial evidence visible in public filings.
An Analytical Theme: Structural Convergence
Structural convergence means that technical progress, physical deployment, and financial evidence begin to collapse into the same analytical problem.
In the earlier AI cycle, these layers could be discussed separately. Model capability was one topic. GPU demand was another. Data-center power was an infrastructure footnote. Free cash flow and depreciation appeared later in financial statements. That separation is becoming less useful.
In 2026, the same thesis has to travel through the full chain: model progress → compute demand → semiconductor supply → HBM and packaging → networking → power → cooling → facility delivery → software usage → revenue quality → margin and cash-flow evidence.
If one layer fails, the failure does not stay local. Packaging latency can delay cluster activation. Power constraints can delay data-center revenue. Heavy CapEx can compress free cash flow before utilization is visible. AI software usage can grow without producing durable revenue if pricing, retention, or gross margin do not follow.
AI exposure is no longer a sufficient analytical category. The more useful question is whether exposure converts through the system without breaking the economics somewhere else.
The Four Conversion Tests
These tests are not a scoring model. They are a discipline for reading public evidence. An issuer can pass one test and fail another; the research task is to identify which link is converting, which link is delayed, and which link is only narrative.
| Conversion test | Core question | Evidence to watch |
|---|---|---|
| Compute Conversion | Can model and workload demand convert into usable compute capacity? | Accelerator shipments, HBM availability, advanced packaging commentary, networking supply, backlog language, cluster deployment timing |
| Power Conversion | Can data-center plans convert into energized, cooled, and reliably powered facilities? | PPA disclosures, interconnection updates, behind-the-meter or co-located power arrangements, facility energization timing, regional power constraints |
| Capital Conversion | Can CapEx convert into durable revenue, margin evidence, and cash-flow quality? | CapEx / revenue, property and equipment additions, depreciation growth, cloud or infrastructure margins, free cash flow versus stated investment cycle |
| Software Conversion | Can AI usage convert into recurring, high-quality revenue? | Commercial revenue growth, usage-based revenue quality, renewal and retention commentary, gross margin impact, implementation cost, RPO or backlog quality where disclosed |
Stress Scenarios
These are schematic stress scenarios, not forecasts and not outputs of a hidden model. A stress scenario matters only if it can be tested. Each row should eventually resolve through filings, segment commentary, infrastructure disclosures, or the absence of evidence where evidence should have appeared.
| Scenario | What breaks | Evidence to watch |
|---|---|---|
| Supply latency persists | Demand is real, but HBM, packaging, networking, or installation delays slow usable capacity and revenue recognition. | Lead-time commentary, inventory behavior, backlog quality, shipment timing, packaging capacity, margin follow-through |
| Power access tightens | Data-center plans outrun grid access, interconnection timing, or contracted power availability. | PPA and power procurement disclosures, project slips, regional grid commentary, energization delays, power-cost sensitivity |
| CapEx digestion fails | Heavy infrastructure investment does not translate into revenue durability, margin improvement, or free-cash-flow normalization within the horizon management describes. | CapEx / revenue staying elevated, depreciation growing faster than operating income, cloud margins stalling, weak unit-cost evidence |
| Software monetization lags | AI usage, pilots, or customer interest do not convert into recurring, high-quality commercial revenue. | Weak AI revenue disclosure, margin pressure from inference cost, long implementation cycles, low renewal visibility, narrative metrics without financial conversion |
Evidence I Track
On CM Terminal, this evidence is pursued through AI Infrastructure Economics, issuer-level Analytics profiles, methodology notes, and the Archive of structural research. The emphasis is whether the same story survives across operating data, physical constraints, and financial statements—not one headline metric.
| Evidence category | What it indicates |
|---|---|
| CapEx / revenue | Whether capital intensity is rising faster than revenue conversion, or beginning to normalize after an investment wave |
| Property, equipment, and depreciation | Whether infrastructure spending is becoming a durable cost burden or a platform for future operating leverage |
| Cloud and infrastructure margins | Whether scale economics are visible after power, depreciation, facility, and mix effects |
| HBM and advanced packaging commentary | Whether the bottleneck is moving upstream from demand into memory bandwidth and packaging capacity |
| Power, PPA, and interconnection disclosures | Whether data-center plans have a credible path to energization and reliable operation |
| Data-center deployment timing | Whether physical capacity arrives on the timeline implied by demand and CapEx language |
| Enterprise AI revenue quality | Whether AI usage becomes recurring revenue with margin support, rather than pilot activity or narrative adoption |
Methodology & Boundaries
This outlook is based on public issuer filings, earnings commentary, infrastructure disclosures, energy-system research, semiconductor supply-chain commentary, and CM Terminal Analytics outputs.
- Issuer-level by default. Metrics are whole-company unless a page explicitly states otherwise.
- No pure-AI revenue unless disclosed. AI exposure is not treated as the same thing as AI revenue.
- No price targets or return forecasts. This page does not assign valuation targets.
- No buy, sell, or hold recommendations. It is not a trading note.
- Not a portfolio allocation framework. It does not rank securities for investment action.
- Not an audit. Figures and classifications remain subject to source limitations, reporting differences, and later correction.
- Ratios over false precision. Cross-issuer comparison emphasizes ratios, disclosed reporting currencies, and explicit methodology rather than artificial consolidation.
The purpose is narrower and more useful: to define the variables that can prove, weaken, or falsify AI infrastructure narratives as 2026 evidence arrives.
Personal analytical notes for 2026. Not investment advice.