A filing-based comparison of how operating margin and direct CapEx intensity differ across AI infrastructure value-chain positions in the current eight-company comparison set.
Core analytical output for AI Infrastructure Economics, organized from thesis to evidence.
Conclusion
AI infrastructure demand does not create one uniform economic model.
In the current eight-company issuer-level comparison, AI demand appears as platform margin capture, capital-heavy bottleneck economics, memory-cycle exposure, equipment bottleneck demand, and downstream infrastructure transmission.
The central finding is that AI demand changes economic form as it moves through compute, foundry, memory, equipment, networking, and power / thermal infrastructure.
This is not a stock ranking, valuation model, or investment recommendation. It is a filing-based business-model comparison.
Thesis
AI exposure is not an economic category.
The current eight-company comparison shows that AI infrastructure demand does not distribute economics evenly. It appears as platform operating leverage in compute, bottleneck economics in foundry and lithography, cycle-exposed demand in memory, and downstream transmission in networking, power, and cooling.
The analytical mistake this page rejects is treating all AI-linked companies as if they share the same profit structure, capital burden, or disclosure clarity.
Role map
| Layer | Economic signature | Companies | Why it matters |
|---|---|---|---|
| Platform capture | High operating margin with low direct CapEx burden. | NVIDIA | Clearest issuer-level benchmark for AI platform economics in the current sample. |
| Bottleneck economics | Scarce manufacturing or equipment layer with margin power. | TSMC, ASML | Shows that bottleneck control can support margin even when capital intensity or equipment cycles matter. |
| Less clean AI economics | Real AI relevance, but issuer-level economics are diluted by business mix, cycle exposure, or still-developing leverage. | AMD, Broadcom, Micron | Prevents the analysis from equating AI exposure with clean AI economics. |
| Demand transmission | AI buildout appears through networking, power, cooling, and facility infrastructure demand. | Arista, Vertiv | Extends the map beyond semiconductors but requires different margin expectations. |
Classification
| Economic role | Company | Interpretation |
|---|---|---|
| Margin-capture profile | NVIDIA | Clearest issuer-level platform-style profit capture case. |
| AI compute challenger | AMD | AI compute exposure exists, but current full-company economics do not yet show NVIDIA-style platform leverage. |
| Capital-heavy bottleneck | TSMC | Strong margin can coexist with heavy reinvestment when a company owns manufacturing bottleneck value. |
| Cycle-risk supplier | Micron | AI-linked HBM demand improves relevance, but does not remove pricing cycles, inventory effects, or heavy reinvestment needs. |
| Equipment bottleneck | ASML | Critical upstream equipment supplier; AI relevance is indirect through advanced manufacturing demand. |
| Mixed consolidated profile | Broadcom | AI relevance is real, but custom silicon, networking, infrastructure software, VMware consolidation, and non-AI semiconductor exposure make the issuer-level read less clean. |
| Networking transmission profile | Arista | Capital-light networking infrastructure profile that transmits AI data-center demand. |
| Power / cooling transmission profile | Vertiv | Facility infrastructure profile exposed to data-center buildout, but with different margin structure from compute or equipment. |
Bridge
Issuer-level bridge; not segment-level attribution.
The segment bridge exists because issuer-level metrics are useful but easy to misread. They make cross-company comparison possible, but AI exposure is rarely isolated cleanly at company level. The bridge separates three things that are often confused:
Issuer-level read
High operating margin and low direct CapEx intensity.
AI-relevant segment / proxy
Data Center compute and accelerator platform.
Main limitation or mixed exposure
Company still includes Gaming, Professional Visualization, Automotive, and other businesses.
Business implication
Clearest issuer-level AI margin-capture proxy in the current sample.
Issuer-level read
AI compute challenger, but issuer-level economics do not yet show NVIDIA-style platform leverage.
AI-relevant segment / proxy
Data Center segment, including EPYC and Instinct.
Main limitation or mixed exposure
Client, Gaming, and Embedded businesses affect consolidated results.
Business implication
Should be read as an AI compute challenger, not as a direct NVIDIA-style margin-capture analogue.
Issuer-level read
Strong margin coexists with high CapEx intensity.
AI-relevant segment / proxy
HPC demand, advanced nodes, advanced packaging / CoWoS.
Main limitation or mixed exposure
Smartphone, automotive, IoT, and other foundry demand remain mixed into issuer-level results.
Business implication
Capital-heavy bottleneck where reinvestment burden can coexist with pricing power.
Issuer-level read
High CapEx intensity with memory-cycle pressure.
AI-relevant segment / proxy
HBM, DRAM, and data-center memory.
Main limitation or mixed exposure
Traditional DRAM / NAND pricing, inventory cycle, and commodity memory exposure.
Business implication
AI demand exposure remains filtered through memory-cycle economics.
Issuer-level read
Equipment bottleneck with indirect AI infrastructure relevance.
AI-relevant segment / proxy
EUV and High-NA lithography enabling advanced semiconductor manufacturing.
Main limitation or mixed exposure
AI is not a direct revenue segment; demand comes through logic, memory, and foundry capex cycles.
Business implication
Critical upstream manufacturing enabler, not a direct AI revenue proxy.
Issuer-level read
Mixed consolidated profile with custom silicon, networking, and software effects.
AI-relevant segment / proxy
Custom accelerators, networking silicon, hyperscaler ASIC exposure.
Main limitation or mixed exposure
VMware, infrastructure software, and non-AI semiconductor businesses affect consolidated margins.
Business implication
AI relevance is real, but issuer-level economics require a less clean issuer-level read.
Issuer-level read
Capital-light networking transmission profile with strong margin profile.
AI-relevant segment / proxy
Cloud networking and Ethernet switching for AI clusters.
Main limitation or mixed exposure
Cloud, enterprise, and non-AI networking demand remain mixed; customer concentration matters.
Business implication
AI demand transmission profile, not a compute-platform or manufacturing-capex profile.
Issuer-level read
Power and cooling demand-transmission profile with different margin structure.
AI-relevant segment / proxy
Data center power, thermal management, liquid cooling, and facility infrastructure.
Main limitation or mixed exposure
Non-AI data center, industrial, and traditional infrastructure demand remain mixed.
Business implication
Exposed to AI data-center buildout, but should not be read as a semiconductor-style margin proxy.
This bridge is interpretive and issuer-level; it is not segment-level financial attribution or investment advice.
Method
The chart below tests the thesis with two issuer-level ratios: operating margin as a proxy for margin capture, and direct CapEx intensity as a proxy for capital burden.
Proxy for margin capture: operating income as a share of revenue at issuer level.
Proxy for capital burden: direct capital spending (e.g. property and equipment purchases) as a share of revenue.
Each issuer uses the latest fiscal year available in the project dataset.
Revenue is displayed in each issuer's reporting currency; the comparison relies on ratio metrics, not absolute revenue ranking.
Main chart
The chart maps issuer-level operating margin against direct CapEx intensity. It is a reading map for business-model position, not a fitted classifier, stock ranking, or valuation model.
Issuer comparison
NVIDIA
Strongest issuer-level margin-capture signal in the current sample.
Micron
Heaviest capital-burden signal, consistent with memory manufacturing intensity.
Arista
Most capital-light issuer in the current map.
Broadcom
AI relevance is real, but VMware, infrastructure software, custom silicon, and broader semiconductor exposure make the consolidated read less clean.
Evidence table
The table shows the company-level filing row behind the chart. Revenue remains in original reporting currency; ratio fields drive the comparison.
| Company | Role | Fiscal Year | Currency | Revenue | Gross Margin | Operating Margin | R&D Intensity | CapEx Intensity | Analytical Read |
|---|---|---|---|---|---|---|---|---|---|
| NVIDIA | Compute platform | 2026 | USD | $215,938.0M | 71.10% | 60.38% | — | — | Compute platform profile; strongest platform-margin capture signal in this comparison on this eight-company issuer-level view (issuer-level proxy, not pure AI revenue). |
| AMD | Compute challenger | 2024 | USD | $25,785.0M | 49.35% | 7.37% | 25.03% | 2.47% | Compute challenger profile; included to prevent the false shortcut that AI compute exposure automatically creates NVIDIA-style platform economics. |
| TSMC | Foundry manufacturing | 2024 | TWD | NT$2,894,307.7M | 56.12% | 45.68% | 7.05% | 33.03% | Foundry manufacturing profile; strong operating margin alongside heavy direct manufacturing CapEx burden. |
| Broadcom | Custom silicon / networking / infrastructure software | 2024 | USD | $51,574.0M | 63.03% | 26.10% | 18.05% | 1.06% | Mixed issuer-level read: custom silicon, networking, software, and VMware consolidation complicate issuer-level AI interpretation. |
| Micron | Memory / HBM / cycle risk | 2024 | USD | $25,111.0M | 22.35% | 5.19% | 13.66% | 33.40% | Memory-cycle exposure: AI/HBM demand can improve mix without removing memory-cycle economics at issuer level. |
| ASML | Lithography equipment / EUV bottleneck | 2024 | EUR | €28,262.9M | 51.28% | 31.92% | 15.23% | 7.31% | Equipment bottleneck profile; strong margin capture with lower direct CapEx intensity than foundry or memory manufacturing. |
| Arista | Networking fabric | 2024 | USD | $7,003.1M | 64.13% | 42.05% | 14.23% | 0.46% | Networking fabric profile; capital-light infrastructure profile with strong operating leverage in the latest collected row. |
| Vertiv | Power and thermal infrastructure | 2025 | USD | $10,229.9M | 36.32% | 17.89% | 4.32% | 2.15% | Power and thermal infrastructure profile; facility-level AI infrastructure exposure with lower margin profile than platform or equipment profiles. |
Chart interpretation
The chart separates the current sample into four economic zones.
This is the most capital-light margin zone in the current framework. It suggests operating leverage without large direct manufacturing burden.
NVIDIA is the benchmark case. ASML and Arista also sit in attractive margin / capital-light territory, but their sources of margin are structurally different: ASML reflects equipment bottleneck economics, while Arista reflects networking demand transmission.
This is bottleneck manufacturing economics.
TSMC shows that heavy reinvestment does not automatically destroy margin capture when the company owns a scarce manufacturing layer. CapEx burden and pricing power can coexist.
This is the most difficult economic zone.
Micron shows that AI/HBM demand can improve exposure without eliminating memory-cycle risk, pricing volatility, inventory effects, and heavy reinvestment needs.
This zone contains companies where AI exposure exists but does not yet translate into NVIDIA-style issuer-level platform leverage.
AMD and Vertiv sit in this zone for different reasons: AMD reflects still-developing issuer-level compute leverage, while Vertiv reflects lower-margin facility infrastructure economics.
Analytical claims
Claim 1
Claim 2
Claim 3
Implication
The same AI buildout can produce different financial signatures:
The analytical mistake this page rejects is treating all AI-linked suppliers as if they share the same profit quality, capital burden, or disclosure clarity.
Finding
The current issuer-level evidence suggests that AI demand is economically uneven: it concentrates visible margin capture in platform or bottleneck positions, shifts capital burden toward manufacturing-heavy companies, leaves memory exposed to pricing cycles, and transmits downstream into networking, power, and cooling without preserving compute-platform economics.
The central finding is that AI demand changes form as it moves through the value chain.
This finding does not estimate pure AI revenue, segment-level profitability, valuation, or forward returns.
Appendix
Scope limits, dataset construction, and planned validation.
Separate issuer-level metrics from reported segment disclosures.
Map how hyperscaler CapEx reaches compute, memory, foundry, networking, equipment, and power/cooling suppliers.
Test whether additional networking, cooling, power, or data-center infrastructure issuers change the comparison.
Compare AI-related claims with reported financial evidence.
Source trail
Each profile preserves filing notes, metric definitions, fiscal-year treatment, and issuer-level context for the plotted row.
| Company | Role | Latest fiscal year | |
|---|---|---|---|
| NVIDIA | Compute platform | 2026 | View profile |
| AMD | Compute challenger | 2024 | View profile |
| TSMC | Foundry manufacturing | 2024 | View profile |
| Broadcom | Custom silicon / networking / infrastructure software | 2024 | View profile |
| Micron | Memory / HBM / cycle risk | 2024 | View profile |
| ASML | Lithography equipment / EUV bottleneck | 2024 | View profile |
| Arista | Networking fabric | 2024 | View profile |
| Vertiv | Power and thermal infrastructure | 2025 | View profile |