Sector Research · AI Supply Chain

The AI Supply Chain: Following the $725B

Coverage May 2026 · JK
KAAMOS
Interactive · The Stack

The $725B build-out — from the foundry making the chips, into the data center that runs them, out to the cloud, models and agents on top. Drag to orbit, scroll to zoom, and click any structure to jump to its section.

The supply chain
Inside the $725B build-out
Click anything in the scene to inspect it.
drag orbit · scroll zoom · click explore
Scroll to read

1 · Executive Summary

This report maps the AI supply chain end to end — eleven layers, from the applications people pay for (L10) down to the electricity that powers them (L0). The aim is structural rather than directional: to show what each layer does, how money flows between them, and where the economics actually accrue. The analysis starts from the logic of hyperscaler capital expenditure — roughly $725B in 2026 — and traces how that single flow cascades down the stack, layer by layer, into chips, memory, networking, data centers and power. Throughout, the focus is on fundamentals and the business relationships between layers rather than on individual securities; the report makes no stock recommendations and takes no investment stance. Each layer is examined on the same template: what it does, who pays whom, its bottleneck, market size, a value-added (margin and pricing-power) analysis, and a live comparison table of the key financial metrics for every public name in it.

LayerFunctionNames
L10Applications & Agents2
L9Foundation Models (LLMs)7
L8Cloud (Hyperscaler + Neoclouds)13
L7Data-Center Real Estate & Builders5
L6Networking + Silicon Photonics + Fiber + AI-RAN13
L5Servers & Power Supplies4
L4AI Accelerators (Chips)7
L3Memory + Storage6
L2Foundry & Advanced Packaging6
L1EDA + Equipment + Materials16
L0Power & Cooling17

Valuation and margin data are pulled live from Yahoo Finance via yfinance and stored in a local SQLite database (snapshot date shown on each table). Segment splits, contract terms, backlog/RPO and AI-specific revenue are hand-curated from company earnings releases and filings. P/S is computed as market cap ÷ TTM revenue (both USD) to avoid the currency mismatch that breaks Yahoo's reported P/S for ADRs and foreign listings. International market caps are converted to USD at the refresh-date FX rate. EV/EBITDA, P/FCF and FCF margin are computed from USD-converted free cash flow, EBITDA and enterprise value; EV/EBITDA is suppressed (“—”) for the handful of ADRs where Yahoo’s enterprise-value currency is unreliable (it sits outside 0.3–3× of market cap). P/FCF shows “neg” where free cash flow is negative (e.g. neoclouds mid-buildout).

2 · Hyperscaler Capex — the $725B keystone flow

The analysis begins with hyperscaler capital expenditure, the single flow from which the rest of the chain derives. The five largest buyers — Amazon, Microsoft, Alphabet, Meta and Oracle — are guiding to roughly $725B of capex in 2026 (Amazon ~$200B, Microsoft ~$190B, Alphabet ~$185–190B, Meta ~$135–145B, Oracle ~$50B).[1] This is the upstream demand signal for every layer below it: the capital committed here sets the throughput available to accelerators, memory, servers, networking, data centers and power, and it is the most useful leading indicator for the chain as a whole. Table 2.1 sets out how that spending is estimated to allocate across categories.

2.1 Where the $725B goes

My estimate of how the 2026 capex breaks down by category. I build this bottom-up from hyperscaler guidance and the bill-of-materials of a modern AI rack, then cross-check it against the sell-side; treat the ranges, not the point figures, as the signal.

Category% of capex$ allocationPrimary beneficiaries
AI accelerators (GPUs + ASICs)~40–45%$290–325BNVDA, AVGO, MRVL, AMD
HBM + system memory~10–12%$70–85BSK Hynix, MU, Samsung
Data-center shell, land, fiber~10–12%$70–85BEQIX, DLR, GLW
Power & electrical equipment~8–10%$60–75BSU.PA, VRT, ETN, GEV, BE, BWXT
Servers, CPUs, NICs, PSUs~8–10%$60–70BDELL, SMCI, Foxconn, Delta
Networking~6–8%$45–60BANET, AVGO, NOK, COHR, LITE
Storage (SSDs + HDDs)~3–4%$22–30BSNDK, WDC, STX
Cooling~3–4%$22–30BVRT, SU.PA, TT, JCI, Daikin
Software, EDA, security~3–5%$25–35BSNPS, CDNS; in-house

Source: author’s estimates built from company capex guidance and earnings disclosures,[1] cross-checked against IDC server data,[2] TrendForce memory data[3] and McKinsey’s data-center capital model.[4] Allocation shifts each earnings cycle.

Accelerators are the largest single bucket and the tightest bottleneck, while power & cooling is the fastest-growing line as rack density climbs from ~130 kW (Blackwell) toward ~1 MW (Rubin).

Where the $725B flows — capex allocation by category & the layer it feeds hover a flow
AI accelerators (GPUs + ASICs): $290–325B (~40–45%) → L4HBM + system memory: $70–85B (~10–12%) → L3Data-center shell, land, fiber: $70–85B (~10–12%) → L7Power & electrical equipment: $60–75B (~8–10%) → L0Servers, CPUs, NICs, PSUs: $60–70B (~8–10%) → L5Networking: $45–60B (~6–8%) → L6Storage (SSDs + HDDs): $22–30B (~3–4%) → L3Cooling: $22–30B (~3–4%) → L0Software, EDA, security: $25–35B (~3–5%) → L12026E hyperscalercapital expenditure~$725BAI accelerators (GPU + ASIC) L4$290–325B · ~40–45% of capexHBM + system memory L3$70–85B · ~10–12% of capexData-center shell, land, fiber L7$70–85B · ~10–12% of capexPower & electrical equipment L0$60–75B · ~8–10% of capexServers, CPUs, NICs, PSUs L5$60–70B · ~8–10% of capexNetworking L6$45–60B · ~6–8% of capexStorage (SSDs + HDDs) L3$22–30B · ~3–4% of capexCooling L0$22–30B · ~3–4% of capexSoftware, EDA, security L1$25–35B · ~3–5% of capex
Big-5 hyperscaler capex — history & forecast ($B) [1]
$144B2023$240B2024$391B2025$768B2026E
Amazon$200BTotal capex; AWS is the fast-growing part.
Microsoft$190BFY basis; Azure + OpenAI infra.
Alphabet$188BGoogle Cloud + DeepMind + TPU.
Meta$140BCaptive AI clusters (no external cloud).
Oracle$50BOCI / Stargate buildout.
Amazon capex ($B) [1]
2023$48B2024$83B2025$118B2026E$200B
2023$48B
2024$83B+73%
2025$118B+42%
2026E$200B+69%
Microsoft capex ($B) [1]
2023$28B2024$55B2025$95B2026E$190B
2023$28B
2024$55B+96%
2025$95B+73%
2026E$190B+100%
Alphabet capex ($B) [1]
2023$32B2024$52B2025$85B2026E$188B
2023$32B
2024$52B+63%
2025$85B+63%
2026E$188B+121%
Meta capex ($B) [1]
2023$28B2024$37B2025$72B2026E$140B
2023$28B
2024$37B+32%
2025$72B+95%
2026E$140B+94%
Oracle capex ($B) [1]
2023$8B2024$13B2025$21B2026E$50B
2023$8B
2024$13B+63%
2025$21B+62%
2026E$50B+138%

Is the $725B the top-level flow, and does it equal the market size?

The $725B is best understood as the top of the infrastructure funnel rather than the size of the AI market. It measures annual capital spending by the five largest buyers, and it cascades downstream into accelerators (L4), memory (L3), servers (L5), networking (L6), data centers (L7) and power (L0) — which is what makes it the most useful single figure to track. It is not, however, the total addressable market.

It understates the full AI economy in three respects. First, it is capex, not revenue: the market also includes the operating revenue generated at the top of the stack (apps ~$30–40B, models ~$100B ARR),[5] which is the demand the capex is built to serve. Second, it captures only the big five: adding neoclouds, sovereign AI, Chinese hyperscalers and enterprise clusters takes Goldman’s estimate of cumulative spend to $1.15–1.4T across 2025–27.[6] Third, it is partly circular: some of it flows to NVIDIA, which reinvests in customers that in turn buy more compute, so the headline figure double-counts (see §6). The figure that most repays attention is the gap between the two — roughly $725B of spend against ~$170B of AI revenue today. Whether that gap closes is the central bull/bear question of the cycle, taken up in §5.

3 · How the Layers Connect & Who Captures Value

With the capex sized, the next step is to trace where it goes. A single hyperscaler capex dollar fans out across all 11 layers, and the chain is not strictly linear: hyperscalers (L8) own custom chips (L4: Trainium, TPU) and apps (L10: Copilot), and a bottleneck at any layer (HBM in 2024–25, CoWoS today, power in 2026–27) caps the throughput of everything above it. Two questions determine where the money stays: how the dollar physically moves (§3.1), and who keeps the most of it as profit (§3.2). The synthesis of both is in §3.3.

3.1 The money flow — a capex dollar’s journey down the stack

The capex Sankey in §2.1 shows the first hop of this journey — how the $725B splits across categories and the layer each one feeds. The table below traces the rest of the chain, step by step, from the end-customer down to power.

StepWho pays whomExample flow
1. End-customers → Apps (L10)Consumers + enterprises pay subscriptions / per-seat licenses$20/mo for ChatGPT Plus → OpenAI
2. Apps → Foundation Models (L9)Apps are the model, or pay per-token via APICursor pays Anthropic for Claude API
3. LLMs → Cloud (L8)Labs are the biggest customers of hyperscalers + neocloudsAnthropic $100B/10-yr AWS; OpenAI $300B/5-yr Oracle
4. Cloud → Data Centers (L7)Own DCs or lease from REITs; hire buildersMSFT leases Equinix space; pays Quanta to build
5. Cloud/DC → Networking (L6)Each rack needs switches, optics, fiberAWS buys Arista switches + Corning fiber + Lumentum 1.6T optics
6. Cloud → Servers (L5)ODMs assemble the physical racksMSFT pays Foxconn to assemble GB200 NVL72 racks
7. Servers → Accelerators (L4)GPUs/ASICs = 40–45% of server costFoxconn buys NVIDIA Blackwell GPUs
8. Chips → Memory (L3)Every GPU needs HBM + DRAM + storageNVIDIA buys HBM3E from SK Hynix
9. Chips → Foundry (L2)Fabless designers send designs to TSMCNVIDIA pays TSMC to fab + CoWoS-package Blackwell
10. Foundry → Equipment (L1)Foundries buy WFE (wafer-fab equipment) + materials + softwareTSMC buys ASML EUV + Hoya mask blanks
11. Everything → Power (L0)Every layer needs electricity; cooling removes heatAWS signs Talen 960 MW nuclear PPA
How a capex dollar cascades down the stack — and how the deep layers get paid hover a flow
Hyperscaler capex → L4 Accelerators: ~$307BHyperscaler capex → L3 Memory + storage: ~$104BHyperscaler capex → L0 Power & cooling: ~$95BHyperscaler capex → L7 Data centers: ~$78BHyperscaler capex → L5 Servers: ~$65BHyperscaler capex → L6 Networking: ~$52BL4 → L2: ~$40B — NVIDIA pays TSMC to fabricate + CoWoS-package the chipsL3 → L1: ~$22B — memory makers buy fab tools (Lam, ASML, bonders)L4 → L1: ~$5B — chip designers pay EDA (Synopsys, Cadence)L2 → L1: ~$15B — TSMC buys EUV / WFE from ASML, AMAT, LamHyperscalercapex 2026E~$725BAccelerators L4~$307BMemory + storage L3~$104BPower & cooling L0~$95BData centers L7~$78BServers L5~$65BNetworking L6~$52BFoundry L2~$40B from aboveEquipment & EDA L1~$42B from above

Direct-spend ribbons (left) are proportional $; the thin upstream ribbons (right) are the slices that cascade deeper — accelerators (L4) pay the foundry (L2), and foundry + memory pay the equipment makers (L1). Every layer also draws on power (L0). Widths are illustrative, built from the §2.1 allocation and bill-of-materials estimates.

3.2 Value capture — who keeps the most of each capex dollar

Value capture is where the structural analysis pays off. Two metrics together explain who keeps the most of each capex dollar: wallet share (what % of capex flows into the layer) and gross/operating margin (how much of that the layer keeps as profit). High wallet share × high margin is what defines the strongest businesses in this chain. The pattern is barbell-shaped: the largest profit pools sit at the chip (L4) and its irreplaceable inputs (L2 foundry, L1 EUV, L3 HBM), while the hardware-integration middle (L5 servers/ODMs) does real and fast-growing work at commodity margins.

LayerWallet shareRepresentative margin (live)Pricing powerWhat it tells us
L10 — Appsdemand-side*Microsoft 46% opMedium — model-gatedWhere cash enters, but the model below keeps most of it. Winners need distribution (MSFT) or proprietary data. *end-user revenue, not a capex slice.
L9 — Foundation modelsdemand-side*Alphabet 32% opHigh — real but transient~$100B ARR; top-3 take ~70%. Inference ~50–60% gross, but pure labs run GAAP-negative while training the next model. Margin shown = Alphabet, the owns-the-stack proxy. *largest L8 customer, not a capex slice.
L8 — Cloud (hyperscaler)captures L9→L8Microsoft 46% opHigh — three-player oligopolyCaptures the labs' compute spend at 35–50% margins; $1.8T+ combined backlog.
L7 — DC REITs~3–4%*Equinix 21% opHigh — interconnection moatsHigh recurring revenue, capital-intensive. *real-estate component.
L6 — Networking + Optics~6–8%Arista 43% opHigh — capacity-constrainedOptics (LITE/COHR) supply-bottlenecked through 2027; switching leaders 40%+ margins.
L5 — Servers / ODMs~8–10%Foxconn 6% gmVery low — near-commodityFoxconn ~6% gross margin on AI servers; volume game, margins compress with scale.
L4 — Accelerators~40–45%NVIDIA 60% opHighest — monopolyLargest single profit pool: ~$120B net income on $216B revenue (FY26).
L3 — HBM/Memory~10–12%SK 49% opVery high — sold-out oligopolySK Hynix at cycle peak; HBM TAM $35B→$100B (2028). Priced for mean-reversion (fwd P/E ~5×).
L2 — Foundry~5–7%*Taiwan 51% opVery high — near-monopolyHighest-quality monopoly; sole leading-edge + CoWoS. *chip-cost component.
L1 — EUV~3–4%*ASML 35% opHighest — pure monopolySole EUV supplier; High-NA $380M/unit, no substitute. *litho component.
L1 — EDA~1%Synopsys 13% opVery high — duopolySoftware economics: ~95% incremental gross margins on every chip designed.
L0 — Power & Equip~8–12%Vertiv 19% opMedium-high — supply-constrainedMultiple suppliers, but transformer/turbine lead times = pricing power.

3.3 What I take away from this

  • NVIDIA captures the largest single profit pool in the entire chain — roughly $120B of net income on $216B of revenue (FY26), equivalent to capturing on the order of a third of all AI capex as profit. No other layer comes close in absolute dollars.
  • The highest-quality monopolies cluster at the input end (L1, L2): ASML (EUV), TSMC (leading-edge + CoWoS), Hoya (EUV mask blanks), Hanmi (HBM bonders). Their margins are durable because the barrier to replicate them is measured in decades and tens of billions of dollars, not quarters.
  • The lowest-value layer is hardware integration (L5 branded OEMs and ODMs): Foxconn runs 5–7% gross margins on AI servers. The work is real and volumes are exploding, but pricing power is near zero — margins compress as scale rises. The value sits in the scarce inputs, not the assembly.
  • The binding constraint of the current cycle has migrated to the physical layer — power and the data-center shell. With grid-interconnect queues at 7–10 years and transformer and turbine lead times beyond 100 weeks, the scarce input is increasingly the ability to energize and cool a building rather than to buy the chips that fill it, so electrical equipment and on-site generation (L0) and powered-shell capacity (L7) now gate how fast the rest of the chain can deploy. After silicon, power is the next bottleneck of the buildout.

4 · Forward Growth vs Valuation Map

Trailing multiples systematically misprice this universe: a name on 45× forward earnings compounding 40% is cheaper than one on 15× going nowhere. This map plots every public name with a usable forward estimate on the two axes that matter — a two-year forward revenue CAGR (horizontal) against EV / forward sales (vertical, log scale). Bubble size is market cap; colour is supply-chain layer. The dashed line through each layer is its own peer regression — distance above the line flags a name priced richly for its growth, below flags one priced cheaply.

A deliberate feature of this universe: growth alone does not predict the multiple. The premium names are the wide-moat, high-margin franchises (networking, connectivity, analog) rather than the fastest growers — which is why the regression is fit per layer, not as a single line across the chain. Read each name against its own layer's peers, not the whole field.

Growth vs forward valuation — full chain[7][8]
0.5x0.3x1x1.5x2x3x4x5x6x8x10x12x15x20x25x30x40x60x0%20%40%60%80%Forward revenue growth — 2-year CAGR (FY0→FY+2 consensus)EV / forward sales (NTM, log scale)ALABCCJTSM6146AVGO6920LRCXANET688981STXLITEDLREQIXNVDAINTCGOOGLMSFTNEE2308QMETAFCEL4063005930NOKASXAMZNVSTPWRAMKREME99886367DELLSMCI2317NVDAAMDAVGOMRVLMPWRINTCMUSNDKWDCSTX000660005930MSFTAMZNGOOGLMETAORCL9988EQIXDLRPWRFIXANETCSCONOKLITECOHRAAOIGLWALABCRDODELLSMCI23172308TSM688981ASXAMKRSNPSCDNSARMASMLAMATLRCXKLAC8035685769206146042700BESI77414063QVRTSUETN6367JCITTGEVBEFCELBWXTCEGVSTTLNNEECCJ0700BIDUEMELayer (click to filter)L9 · Foundation modelsL8 · CloudL7 · DC real estateL6 · NetworkingL5 · ServersL4 · AcceleratorsL3 · MemoryL2 · FoundryL1 · EDA & equipmentL0 · Power & coolingBubble ∝ market capDashed = peer regressionAbove line = rich for growthBelow = cheap for growthClick a layer · hover a dot

EV, market cap and the FY+1 revenue estimate are from the project database (Yahoo Finance snapshot; curated overrides where consensus is unreliable).[7] The horizontal axis is the two-year forward revenue CAGR = (FY+2 ÷ FY0)½ − 1, a steadier signal than the base-effect-prone one-year figure. Second-forward-year (FY+2) revenue is hand-collected from analyst-consensus forecast tables;[8] most are read directly off the per-year table, the remainder derived from the published consensus 3-year revenue CAGR (see the dataset for per-name sourcing). The vertical axis uses near-term (FY+1) forward sales; a handful of stale +1y figures are repaired by interpolating between FY0 and FY+2. Six capacity-financed names (CIFR, CRWV, HUT, IREN, NBIS, WULF) are held out: their enterprise value is debt-heavy and their forward revenue is a contracted build-out ramp, so EV/forward-sales understates how richly they are priced — they belong on a separate EV/contracted-RPO lens (next iteration). Foreign and ADR names (TSMC, ASML, SK Hynix, Samsung, Tokyo Electron, Arm, Foxconn, Tencent and the rest) are included using their Simply Wall St consensus 3-year revenue CAGR — a currency-independent figure — anchored on their reliable USD trailing revenue; the three whose Yahoo enterprise value is currency-broken (TSMC, ASE, Baidu) use a clean EV from stockanalysis.com. Only a handful with genuinely missing data remain off.

Companion lenses live inside each layer below (a layer-only growth-vs-valuation chart plus a valuation read). The held-out neoclouds — capacity-financed names whose EV/forward-sales understates them — are covered with an EV ÷ annualized-RPO lens inside Layer 8, and a forward P/AFFO view sits in Layer 7.

5 · The Supply Chain — Layer by Layer

Each table is horizontally scrollable. Margins use the latest fiscal year (TTM where FY is stale). P/S = market cap ÷ TTM revenue. Market cap converted to USD. “—” = not available / not meaningful.

Layer 10 — Applications & Agents

Layer ProfileL10 · APPS & AGENTS

What this layer does

Where end-users actually pay for AI. Apps either bundle a foundation model behind a UI (ChatGPT, Claude.ai) or wrap one via API to solve a specific job (Cursor for coding, Harvey for legal).

Who pays whom

Consumers pay $20–200/month subscriptions; enterprises pay per-seat licenses (Microsoft 365 Copilot $30/user/mo; GitHub Copilot Business $19). Apps in turn pay LLM providers (L9) per token, or — if they own the model — pay the cloud (L8) directly.

Key Metrics to Track

Track ARR growth + net revenue retention for the private names (via funding rounds), and seat counts + attach rates for the public proxies. The key question for any app: how much of its gross margin survives the next round of model price cuts and competition?

Margin that mattersGross margin is what matters here — an app keeps only the spread above what it pays the model provider, so gross margin is the business model.
Valuation lens that mattersAnchor on EV/Sales + the Rule of 40 (growth % + FCF margin), not P/E — most apps are private and reinvesting, so they are priced on growth and net revenue retention, not earnings. The trap: an app's revenue multiple is only justified if its gross margin survives the next model price cut.

Analyst’s Take

The application layer is where cash physically enters the AI economy, yet it keeps the least of it: the model underneath captures most of the economics, so an app’s margin is only the spread it can hold above its model bill. The defining characteristic of the layer is therefore moat durability rather than growth. Distribution (Microsoft embedding Copilot across 400M+ M365 seats) and proprietary data or workflow (Harvey in legal, Glean in enterprise search) compound, while undifferentiated ‘GPT wrappers’ are compressed every time model prices fall. The second structural feature is that the layer is overwhelmingly private — almost every pure-play is venture-funded — so listed exposure is concentrated in the platform owners rather than the fast-growing native apps.

ReadProprietary-workflow and distribution-advantaged apps look like the durable winners; thin wrappers with no data moat are structurally squeezed as model prices fall. Largely a private-market layer today, with listed exposure limited to the platform owners.

Key risk

App-margin compression. An app keeps only the spread above what it pays the model, and switching costs are low (swap the model underneath). If model prices don't fall as fast as competition pushes app prices down, the middle gets squeezed.

Bottleneck LOW

  • Not a capacity bottleneck — apps depend on L9 model availability and L8 compute, both of which the $725B capex is solving.
  • The real constraint is customer acquisition and retention economics: Cursor, Cognition, Windsurf and others compete fiercely, and switching costs are low because they can swap the underlying model.
  • Margin compression risk: an app's gross margin is whatever is left after paying the model provider; if model prices don't fall as fast as competition pushes app prices down, the middle gets squeezed.

Market Size

Share of the $725B capexNot a slice of the $725B. L10 is the demand the capex is built to serve — the ~$30–40B app market is end-user revenue, not capital spending.
  • ~$30–40B in 2026E[5] — the fastest-growing software category in history.
  • Cursor reached $2B ARR within ~24 months; Claude Code hit $1B+ ARR within 6 months of launch; GitHub Copilot has 4.7M paid subscribers.[1]
  • Still tiny relative to the $100B+ model layer below it — most of the value today is captured by the model, not the app wrapper.

Value Added & Margins

  • VERY HIGH for winners, LOW for losers. Winning apps earn 70–90% SaaS-like gross margins — but only on the spread above model spend.
  • The durable moats are distribution (Microsoft bundles Copilot into 400M+ M365 seats), proprietary workflow/data (Glean, Harvey), and habit.
  • Most pure-plays remain private (Anysphere/Cursor, Glean, Perplexity, Cognition); public-market access is essentially limited to Microsoft and Alphabet.
Selected AI-app annualized run-rates [1]
M365 Copilot (run-rate)$5BCursor (Anysphere)$2BPerplexity$1BClaude Code$1B
M365 Copilot (run-rate)$5B~$5B+
Cursor (Anysphere)$2B$2B in ~24 mo
Perplexity$1B~$1B
Claude Code$1B$1B+ in 6 mo

Valuation readApps & agents

Most of the application layer is private (OpenAI's and Anthropic's apps, the agent start-ups), so the only public plots are the platform owners (Microsoft, Alphabet) already covered in L8. There is no meaningful public app-layer valuation scatter yet; the layer's economics surface in the LLM-revenue ARR figures above rather than in listed equities.

Sub-segments

General-purpose chatbots
ChatGPT, Claude.ai, Gemini App (~$20B).
Coding assistants/agents
Cursor ($2B ARR), GitHub Copilot (4.7M subs), Claude Code ($1B+); fastest-growing.
Enterprise productivity
M365 Copilot (~$5B+ run-rate), Google Workspace AI.
Vertical agents
Harvey (legal), Glean (search), Sierra (CX), Cognition/Devin (coding); mostly private.
AI search
Perplexity (~$1B ARR); creative — Midjourney, Runway, ElevenLabs.
L10 · AppsWatch: gross margin (the spread above model cost). Value on EV/sales; most names are private.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
MSFTMicrosoftCopilot revenue is bundled inside MSFT's Productivity + Cloud segments (not broken out). M365 Copilot ~$5B+ run-rate ($30/seat); GitHub Copilot ~$450–850M ARR (4.7M paid subs, 90% of F100).~2% (est.)$281.7BFYE Jun ’25$384Bconsensus+36%69%46%$3,067.2B24.6×21.4×1.323%9.6×7.4×17.0×82.9×
GOOGLAlphabetGemini App + Workspace AI bundled within Google Services / Cloud; not separately disclosed. Embedded across Search, Workspace, Android.~3% (est.)$402.8BFYE Dec ’25$575Bconsensus+43%60%32%$4,766.6B30.0×27.2×1.582%11.3×9.9×29.0×170.7×
Glossary — 2 terms
ARR (Annualized Run-Rate) — Current monthly revenue × 12 — how private AI companies report scale. Forward-looking, not GAAP.
Net revenue retention — Revenue this year from last year's cohort ÷ what they paid last year. >100% means existing customers expand.

Layer 9 — Foundation Models (LLMs)

Layer ProfileL9 · FOUNDATION MODELS

What this layer does

Foundation-model labs train the large neural networks (GPT-5, Claude Opus 4.5, Gemini 3, Llama 4, Grok 4) that every app in Layer 10 builds on. The capital-intensive heart of the AI stack.

Who pays whom

End-users pay labs directly (subscriptions, seats) or indirectly (apps pay per token via API). Labs are in turn the largest customers of the cloud (L8) — Anthropic ~$100B/10-yr AWS, OpenAI $250B Microsoft + ~$300B Oracle + $38B AWS — and that compute spend is funded substantially by hyperscaler equity.

Key Metrics to Track

No pure-play public exposure. Four proxies: MSFT (27% of OpenAI), AMZN (largest Anthropic holder), GOOGL (Gemini, fully owned stack), META (Llama, monetized via ads). Track each lab's ARR, enterprise API share (Menlo/Ramp), gross margin trajectory, and — critically — who is funding its compute.

Margin that mattersGross margin (inference unit economics, ~50–60%) is the read to watch. GAAP operating margin is deeply negative mid-training and misleads.
Valuation lens that mattersThere is no public multiple — labs are priced in private rounds on EV / forward-ARR (OpenAI ~$852B on ~$25B ARR ≈ 34×; Anthropic ~$800B–1T on ~$30B+ ARR ≈ 27–33×, after re-rating from ~$350B in early ’26). The special characteristic: GAAP earnings are deeply negative by design while training the next model, so any P/E is meaningless — you are underwriting revenue growth and capability, financed by a hyperscaler's balance sheet.

Analyst’s Take

Foundation models are the most contested layer in the chain, and the fundamentals are deliberately unsentimental: the labs combine extraordinary revenue growth with deeply negative GAAP economics, because each model costs billions to train while the prior generation depreciates toward commodity pricing. Two structural features matter more than the benchmark leaderboard. First, enterprise API share — where Anthropic has quietly moved ahead — is a cleaner read on durable demand than consumer mindshare. Second, and more important, every lab has sold a stake in itself to one or two hyperscalers in exchange for compute, so the layer cannot be analysed in isolation from the L8 balance sheets that fund it. With no clean public pure-play, the economics surface through the backers.

ReadOwning the full stack — model plus cloud plus distribution — is the structurally advantaged position, since standalone labs are gated by whoever funds their compute. There is no public pure-play, so the layer’s economics are read through the hyperscaler backers; the variables that matter are enterprise API share and compute funding, not benchmark wins.

Key risk

Circular financing. NVDA→OpenAI→NVDA, AMZN→Anthropic→AWS and similar loops exceed $150B in announced flows — vendors funding the customers who then buy their product. It's legal and disclosed, but it inflates apparent demand and concentrates counterparty risk on a handful of labs.

Bottleneck MEDIUM-HIGH

  • Frontier training runs need tens of thousands of the latest GPUs in a single coherent cluster — gated by L4 chip supply and L0 power, not by talent or data alone.
  • Capital is the second constraint: a frontier run + serving infrastructure costs billions, which is why labs trade equity for compute (OpenAI–NVIDIA, Anthropic–Amazon).
  • Resolution path is the hyperscaler equity-for-compute deals — but that ties each lab's fate to one or two backers and creates the circular-financing risk.

Market Size

Share of the $725B capexNot a slice of the $725B. L9 is the largest customer of the capex (labs are the biggest buyers of L8 compute); the ~$100B of LLM ARR[5] is the revenue the buildout is chasing.
  • ~$100B in LLM ARR (Q1’26)[5] — from under $1B in Jan 2023.
  • Forecast $400B+ by 2030 if scaling continues to translate into capability and demand.[5]
  • Concentration is extreme: the top three (Anthropic, OpenAI, Google) take ~70%+ of revenue (see charts).

Value Added & Margins

  • HIGH for the top three, brutal for everyone else. Anthropic + OpenAI together capture ~60% of LLM revenue.
  • Gross margins on inference are ~50–60% once a model is trained, but GAAP margins are deeply negative while training the next model and amortizing compute — the labs are burning billions.
  • The model itself is depreciating inventory: each generation makes the last one nearly worthless, so labs must keep spending just to stay frontier. Pricing power is real but transient.
Global LLM revenue share — Q1 2026 [5]
Anthropic31.4%$30B+ ARR; ~$800B–1T val (2026 round) — overtook OpenAI Apr’26
OpenAI29%$25B ARR; $852B PBC val
Google (Gemini)12.1%Embedded across Workspace + Cloud
Microsoft (Phi)7.2%Separate from its OpenAI stake
Tencent (Hunyuan)4.8%#1 Chinese by revenue
Baidu (Ernie)3.6%Enterprise + search
Alibaba (Qwen)2.9%Strong open-weight; cloud bundle
Meta (Llama)1.4%Open-source; indirect/ad revenue
xAI (Grok)1.4%X-tied; ~$3B ARR
Others6%Perplexity + long tail
Enterprise API spend share — vs end-2023 [9]
Anthropic40%OpenAI25%Google20%Meta (Llama)9%DeepSeek + others14%
Anthropic40%rising from 12% → +28 pts; now the enterprise leader
OpenAI25%down from 50% → −25 pts
Google20%+15 pts
Meta (Llama)9%open-source, ~stable
DeepSeek + others14%fragmented long tail
Global LLM industry revenue — annual trend & forecast [5]
2023$3B2024$12B2025$45B2026E$100B2028E$230B2030E$400B
2023$3Bfrom <$1B in Jan'23
2024$12Bfirst scaled year
2025$45Binflection
2026E$100B~$100B ARR by Q1'26
2028E$230Bif scaling holds
2030E$400B$400B+ base case
Growth vs forward valuation — Foundation models (7 names)[7][8]
1.5x2x3x4x5x6x8x10x0%20%2-yr forward revenue CAGREV / fwd sales (NTM, log)GOOGLMSFTMETA0700AMZNBIDU9988

Valuation readFoundation models

Public exposure to the model layer is almost entirely indirect — the leading pure-play labs (OpenAI, Anthropic, xAI) are private, so the names here are the hyperscaler parents that own or rent to them (Microsoft, Alphabet, Amazon, Meta) plus the Chinese platforms. On that basis Meta (~5× forward sales / ~18%) and Alphabet (~8×/18%) look reasonable for their growth, while the Chinese platforms screen cheap on a growth-adjusted basis — Tencent (~4×/9%) and Baidu (~2×/5%) carry the China/regulatory discount. For the US names this read is really an L8 read in disguise; the genuine model-layer valuation debate lives in private markets and the circular-financing web (§6).

L9 · LLMWatch: gross margin (inference unit economics). Private labs trade on EV/forward-ARR; P/E is meaningless mid-training.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
GOOGLAlphabetOwns Gemini outright (100%) and the full stack — model + TPU + Google Cloud. The only frontier lab that owns its own accelerator and cloud, which is why Cloud margins inflected.~15% (est.)$402.8BFYE Dec ’25$575Bconsensus+43%60%32%$4,766.6B30.0×27.2×1.582%11.3×9.9×29.0×170.7×
MSFTMicrosoftHolds a ~27% economic stake in OpenAI (~$135B, post-PBC restructure Oct'25) + IP rights to 2032; also builds in-house Phi / MAI models. Cleanest public OpenAI proxy.~15% (est.)$281.7BFYE Jun ’25$384Bconsensus+36%69%46%$3,067.2B24.6×21.4×1.323%9.6×7.4×17.0×82.9×
AMZNAmazonLargest outside shareholder of Anthropic — ~$46B+ invested ($8B+$13B+$25B); convertible/non-voting, est. high-teens % economic, non-controlling. Also owns Nova models. Cleanest Anthropic proxy.~5% (est.)$716.9BFYE Dec ’25$930Bconsensus+30%50%11%$2,924.1B32.4×27.6×1.875%3.9×6.6×18.9×298.2×
METAMeta PlatformsOwns Llama outright (open-weight). Monetized indirectly via ad-targeting uplift, not API fees — so there is no clean 'LLM revenue' line.indirect (ad uplift)$201.0BFYE Dec ’25$302Bconsensus+50%82%41%$1,557.9B22.3×17.0×0.962%7.2×6.4×14.3×61.0×
0700.HKTencent HoldingsOwns Hunyuan (100%); #1 Chinese lab by revenue. Embedded across WeChat / Tencent Cloud.~3% (est.)$110.9BFYE Dec ’2556%33%$500.8B15.6×11.4×1.323%4.4×3.0×12.1×26.1×
BIDUBaiduOwns Ernie (100%); monetized via search + Baidu AI Cloud.~15% (est.)$19.0BFYE Dec ’2544%8%$44.5B14.3×0.7-59%2.3×1.1×37.9×
9988.HKAlibaba GroupOwns Qwen (100%) — the leading open-weight family — bundled with Alibaba Cloud. China's strongest open-model + cloud combination.~10% (est.)$151.0BFYE Mar ’26$165Bguide/curated+9%40%6%$304.5B19.6×12.8×0.4104%2.0×1.9×18.5×neg

Notes, deals & disclosures

The labs are mostly private, so the table shows public proxies. Microsoft is the cleanest OpenAI proxy (27% economic stake worth ~$135B); Amazon is the cleanest Anthropic proxy (~$46B+ invested, largest holder); Alphabet is the only frontier lab that owns its entire stack — model (Gemini), accelerator (TPU) and cloud — which is why its cloud margins inflected so sharply. Meta is a different animal: it open-sources Llama and monetizes indirectly through ad-targeting uplift rather than API fees.

Glossary — 3 terms
Foundation model — A large neural network trained on internet-scale data that many downstream apps build on top of.
Tokens — The unit of text models read/write (~0.75 words each). APIs charge per million in/out tokens (Claude Opus ~$15/M in, ~$75/M out).
PBC — Public-Benefit Corporation — OpenAI's restructured for-profit entity (Oct'25) that crystallized Microsoft's ~27% stake.

Layer 8 — Cloud (Hyperscaler + Neoclouds)

Layer ProfileL8 · CLOUD

What this layer does

Cloud providers own data centers and chips and rent them by the hour or via long-term contract. Two sub-types: hyperscalers (general-purpose, full-stack, own custom silicon) and neoclouds (GPU-as-a-service pure-plays, faster to spin up GPU capacity).

Who pays whom

Model labs (L9) are the largest customers; enterprises buy AI services on top. Clouds in turn pay L7 (real estate), L6 (networking), L5 (servers), L4 (chips) and L0 (power). Neoclouds raise debt against their hyperscaler/lab contracts to buy GPUs from NVIDIA.

Key Metrics to Track

Hyperscalers: cloud revenue growth YoY, RPO/backlog (the leading indicator), AI revenue run-rate, and — the stress test that defines this cycle — the free-cash-flow squeeze. ‘Looking at FCF’ means watching how much cash is left after the capex, because capex this large can turn a cash machine into a cash sink. Three ratios make it concrete: (1) capex-to-OCF (capex ÷ operating cash flow) — how much of the cash the business generates is being plowed straight back into buildout; BofA estimates 2026 reaches ~94% across the big four, versus a ~40% ten-year norm, so almost nothing is left over. (2) FCF/capex (free cash flow ÷ capex) — below ~1.0× the company is spending more than its free cash flow and funding the gap with debt or customer prepayments. (3) FCF margin (FCF ÷ revenue) and FCF conversion (FCF ÷ net income) — both compress hard when capex outruns earnings. The signal is the trajectory: rising capex-to-OCF with falling FCF margin is the clearest tell that the buildout is outrunning cash generation (the revenue-to-capex charts above visualize the same tension; the debate is taken up in §5). Neoclouds: contracted RPO vs market cap, customer concentration, contract duration, cost of debt, and the gap between gross margin (good) and operating margin (often negative).

Margin that mattersOperating margin is the franchise read (AWS ~35%, Azure ~50% on AI). For neoclouds, watch the gross-to-operating gap — ~70% gross collapses to negative operating on GPU depreciation.
Valuation lens that mattersHyperscalers: watch EV/EBITDA and — above all — free cash flow. This cycle's defining feature is that the capex surge is compressing FCF hard: BofA estimates 2026 capex eats ~94% of operating cash flow after dividends, vs a ~40% ten-year norm.[10] So track FCF margin, FCF conversion (FCF ÷ net income) and capex-to-OCF — a rising capex/OCF with falling FCF is the single most important risk signal here. Neoclouds are GAAP- and FCF-negative, so value them on EV / contracted RPO instead.

Analyst’s Take

Cloud is the toll booth on the whole chain: the hyperscalers convert the labs’ compute spend into 35–50% operating margins, and the most reliable fundamental indicator here is RPO/backlog — over $1.8T contracted across the big four[1] signals that the demand is largely signed rather than hoped for. The neoclouds are a structurally different business: economically they are securitized compute contracts in equity form. Their ~70% gross margins look healthy, but depreciation and interest drive operating margins negative, so their viability rests on a handful of mega-contracts and the debt raised against the leases behind them — a high-beta, high-binary-outcome profile rather than a quality-compounding one.

ReadOn fundamentals the hyperscalers are the higher-quality, backlog-backed expression of the buildout; the neoclouds and ex-miners are the high-beta exposure, with economics that turn on a few mega-contracts and the debt financing them. Track RPO/backlog and capex-to-OCF, not headline revenue.

Key risk

The revenue-to-capex gap, and what it does to free cash flow. ~$725B of 2026 capex sits against only ~$170B of AI revenue (~4–5× gap), and BofA estimates capex now eats ~94% of operating cash flow after dividends vs a ~40% ten-year norm[10] — increasingly debt- and prepayment-financed. The bull counter is $1.8T+ of contracted backlog, but it's concentrated in a few labs funded partly by their own suppliers.

Bottleneck MEDIUM

  • Capacity is sold out across AWS/Azure/GCP for 2026 — the constraint is upstream (L4 GPUs, L7 buildings, L0 power), not demand.
  • The $725B capex IS the resolution; neoclouds and ex-miners add capacity faster than hyperscalers can build, which is why they exist.
  • For neoclouds specifically, the binding constraint is capital and customer concentration: they live or die on a handful of mega-contracts and the debt raised against them.

Market Size

Share of the $725B capexL8 is the $725B — this layer does the spending, it does not receive a slice. The capex is L8 converting its balance sheet into the L0–L7 stack below.
  • Total cloud infrastructure ~$520B in 2026E.[11] AWS FY25 $128.7B (+19%); Azure +40%; Google Cloud ~$59B FY25 (+~50%); Oracle cloud ~$35B run-rate.[1]
  • Backlog is the leading indicator and it is enormous: Microsoft RPO $627B, Oracle $553B, Google $462B, Amazon ~$200B — over $1.8T of contracted cloud combined.[1]
  • Neocloud contracted HPC (high-performance computing) revenue exceeds $100B+ (CoreWeave $66.8B RPO alone).[1]

Value Added & Margins

  • HIGH for hyperscalers, THIN and volatile for neoclouds. AWS runs ~35% operating margin; Azure ~50% on the AI run-rate; Google Cloud inflected to ~33%.
  • Neoclouds run 5–15% operating margins at best — many are GAAP-unprofitable today (CoreWeave op margin ~-7%) because GPU depreciation and interest swamp early revenue, even at ~70% gross margins.
  • The ex-miners (IREN, TeraWulf, Hut 8, Cipher) are the highest-beta exposure: they own power and shells, sign 10–15-yr hyperscaler leases, and finance the GPUs with debt — huge upside if contracts perform, severe risk if a counterparty wobbles.
Cloud infrastructure services — revenue share & $ , 2026E (~$520B market) [11]
AWS (Amazon)30%≈$156B 2026E · $128.7B FY25, +19%
Microsoft Azure23%≈$120B 2026E · +40% YoY
Google Cloud13.5%≈$70B 2026E · ~$59B FY25
Alibaba Cloud4%≈$21B 2026E · largest non-Western
Oracle (OCI)3%≈$16B 2026E · $553B RPO
Others + neoclouds26.5%≈$138B 2026E · CoreWeave, Nebius, IBM, Tencent
The FCF squeeze — 2026E capex as % of operating cash flow [10]
Oracle140%Meta108%Microsoft95%Amazon88%Alphabet85%Big-4 average94%
Oracle140%spends well above OCF — debt- & prepayment-funded
Meta108%captive clusters, no external-cloud revenue offset
Microsoft95%Azure + OpenAI infrastructure
Amazon88%AWS cash generation cushions it
Alphabet85%strongest FCF coverage of the five
Big-4 average94%vs a ~40% ten-year norm — almost no FCF left over
Growth vs forward valuation — Cloud (7 names)[7][8]
1.5x2x3x4x5x6x8x10x0%20%40%2-yr forward revenue CAGREV / fwd sales (NTM, log)GOOGLMSFTORCLMETA0700AMZN9988

Valuation readCloud

Among the hyperscalers Oracle screens cheapest-for-growth — ~7× forward sales on a ~39% 2-yr CAGR, the steepest in the group, driven by the Stargate/OCI backlog — while Microsoft and Alphabet sit ~8× on mid-teens growth and Amazon carries the lowest multiple (~3×) because cloud is a minority of a retail-heavy top line. Alibaba (~1.8×) wears the China discount. The neoclouds that also belong to this layer are on the separate EV/contracted-RPO lens below — EV/forward-sales flatters them. Net: a toll-booth oligopoly priced on backlog quality, with Oracle the growth-cheap outlier.

EV ÷ annualized contracted revenue (RPO ÷ term)[1]
CRWV7.5×HUT11.8×WULF11.8×IREN11.9×NBIS14.4×CIFR36.7×
CRWV7.5×$66.8B total RPO; $22.4B OpenAI through May 2031 (3 tranches)
HUT11.8×$7.0B + $9.8B Fluidstack/Google-backed, 15-yr, 3% escalators
WULF11.8×~$12.8B Fluidstack, 10-yr, Google-backstopped
IREN11.9×$9.7B Microsoft, 5-yr, 20% prepaid
NBIS14.4×~$17.4B Microsoft (→$19.4B) + ~$3B Meta, through 2031
CIFR36.7×$5.5B AWS, 15-yr

Valuation readNeocloud — EV vs contracted backlog

The six capacity-financed names sit off the master map because EV/forward-sales understates them: their enterprise value is debt-heavy and their revenue is a contracted build-out ramp. The right lens is EV against contracted backlog (RPO). But total EV/RPO flatters the ex-miners' long-dated leases — Cipher's $5.5B runs 15 years, CoreWeave's $66.8B about five. Dividing RPO by its contract term gives an annualized contracted revenue, and EV ÷ that (the chart above) is the comparable multiple.

NameEV $BContracted RPO $BTerm yrsEV / RPOAnnualized RPO $BEV / ann. RPO
CRWV90.766.85.51.36×12.17.5×
HUT13.216.8150.79×1.111.8×
WULF15.212.8101.18×1.311.8×
IREN23.19.752.38×1.911.9×
NBIS53.520.45.52.62×3.714.4×
CIFR13.55.5152.45×0.436.7×

Contracted RPO and contract terms are hand-curated from company filings and earnings releases;[1] terms are approximate weighted averages where a name carries several deals. EV is from the database snapshot.[7] On a total-RPO basis Hut 8 and CoreWeave look cheapest, but once annualized the long 10–15-year ex-miner contracts re-rate sharply — CoreWeave is the cheapest on signed annual economics, the 15-year deals the richest. Strip circular financing before underwriting any of these: some contracted demand is the same dollars cycling between a few balance sheets.

L8 · US hyperscalerWatch: operating margin + free cash flow. Value on EV/EBITDA and FCF, not P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
MSFTMicrosoftMicrosoft Cloud $54.5B in FQ3'26 (+29% YoY); Intelligent Cloud segment $34.7B (+30%). Azure $ not separately disclosed; Azure growth +40% (+39% cc). AI run-rate >$37B (+123%). Microsoft Cloud ≈ ~half of MSFT's $281.7B FY25 (Jun) revenue.~50% (Microsoft Cloud)$281.7BFYE Jun ’25$384Bconsensus+36%46%$3,067.2B24.6×21.4×1.323%9.6×7.4×17.0×82.9×Commercial RPO $627B (+99% YoY; ~+26% ex-OpenAI)
AMZNAmazonAWS FY25 $128.7B (+19%), op income $45.6B (35.4% margin). Q4'25 $35.6B → Q1'26 $37.6B (+5.6% QoQ, +28% YoY); $150B annualized run-rate. AWS ≈ 18% of Amazon's $716.9B FY25 revenue but ~57% of consolidated operating income.~18% (AWS)$716.9BFYE Dec ’25$930Bconsensus+30%11%$2,924.1B32.4×27.6×1.875%3.9×6.6×18.9×298.2×Amazon RPO ~$200B (multi-year commitments, mostly AWS)
GOOGLAlphabetGoogle Cloud FY25 ~$59B (Q4'25 $17.7B, +48%); Q1'26 $20.0B (+63% YoY, +13% QoQ); Cloud op income $6.6B (33% margin) vs $2.2B a year ago. Cloud ≈ 15% of Alphabet's $402.8B FY25 revenue.~15% (Cloud)$402.8BFYE Dec ’25$575Bconsensus+43%32%$4,766.6B30.0×27.2×1.582%11.3×9.9×29.0×170.7×Google Cloud backlog $462B (≈2x YoY)
ORCLOracleTotal cloud $8.9B in Q3'FY26 (+44%): OCI infrastructure $4.9B (+84% YoY), cloud apps $4.0B (+13%). Pure cloud (OCI+SaaS) is ~$24-25B of Oracle's $57.4B FY25 (May) revenue — ~45% and the fastest-growing part, on track to become the majority.~45% (cloud, rising)$57.4BFYE May ’25$67Bguide/curated+17%31%$544.6B33.9×23.6×1.224%8.5×16.2×24.9×negRPO $553B (+325% YoY; +$29B QoQ) — mostly large-scale AI contracts
METAMeta PlatformsMeta sells no external cloud — its enormous capex builds captive infrastructure for Llama + ad/recommendation AI. Shown here as a top-5 capex spender (~$135–145B 2026E), not a cloud vendor. Revenue is ~98% advertising.~0% (captive; ad-funded)$201.0BFYE Dec ’25$302Bconsensus+50%41%$1,557.9B22.3×17.0×0.962%7.2×6.4×14.3×61.0×n/a (no external cloud RPO)
L8 · Chinese cloudWatch: operating margin. EV/EBITDA, with a China policy discount.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
9988.HKAlibaba GroupAlibaba Cloud ~$5B/quarter, ~10% of Alibaba's $150.9B FY (Mar'26) revenue; ~4% global cloud share (largest non-Western). AI-related product revenue +triple-digit YoY for 11 straight quarters.~10% (Alibaba Cloud)$151.0BFYE Mar ’26$165Bguide/curated+9%6%$304.5B19.6×12.8×0.4104%2.0×1.9×18.5×neg
0700.HKTencent HoldingsTencent Cloud is ~12% of Tencent's revenue (est.) and the #2 Chinese cloud behind Alibaba; bundled with the Hunyuan model and WeChat ecosystem. Group revenue is dominated by games + fintech + ads.~12% (Tencent Cloud, est.)$110.9BFYE Dec ’2533%$500.8B15.6×11.4×1.323%4.4×3.0×12.1×26.1×
L8 · NeocloudWatch: the gross-to-operating GAP (~70% gross → negative operating on GPU depreciation). Value on EV / contracted RPO.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
CRWVCoreWeave pure-playPure-play GPUaaS. FY25 revenue $5.13B (fastest cloud platform ever to $5B). TTM GM ~69%, op margin ~ -7% (heavy D&A/SBC), EBITDA margin ~+49%.100% (pure-play)$5.1BFYE Dec ’25$25Bconsensus+385%72%-1%$55.9B-163.5×9.0×11.5×30.0×negRPO $66.8B (Dec'25, >4x YoY)
NBISNebius Group pure-playEuropean challenger. FY25 revenue ~$0.5B but ARR exited 2025 at $900M–$1.1B (Q2 rev +625% YoY). TTM GM ~72%.100% (pure-play)$530MFYE Dec ’2569%-115%$52.7B80.5×574.9×0.660.1×7.4×negMulti-year contracted backlog ramps through 2026
IRENIREN LimitedEx-bitcoin miner pivoting to AI cloud; FY ends Jun. Owns power + builds liquid-cooled DC at Childress TX (750MW campus).~60%→100% (pivoting)$501MFYE Jun ’2568%4%$23.2B84.3×-48.3×3.130.6×8.8×157.1×neg$9.7B contracted (Microsoft)
WULFTeraWulfHPC revenue ($21M in Q1'26) overtook bitcoin ($13M) for the first time — the pivot inflection point.~60% (HPC>BTC)$168MFYE Dec ’2551%-101%$12.6B84.1×75.2×76.9×neg~$12.8B contracted HPC
HUTHut 8Bitcoin miner pivoting to AI DC leasing.ramping (lease)$235MFYE Dec ’2554%-41%$13.2B-63.2×46.5×9.1×neg~$16.8B base-term contract value (two 15-yr leases)
CIFRCipher MiningBitcoin miner pivoting to AI DC leasing.ramping (lease)$224MFYE Dec ’2528%-151%$9.6B35.1×45.9×13.7×81.3×neg$5.5B contracted (AWS, 15-yr)

Notes, deals & disclosures

Hyperscalers — how the AI/cloud revenue compares to the whole company. The point of showing both segment and total revenue is to see how much of each franchise is actually cloud: AWS is only ~18% of Amazon's $716.9B revenue but ~57% of its operating income; Google Cloud is ~15% of Alphabet; Microsoft Cloud is roughly half of Microsoft. Note that Microsoft does not disclose Azure revenue in dollars — only its growth rate (+40%) — so Azure dollar figures industry-wide are estimates (~$30B+/quarter). Oracle is the outlier: cloud is now the majority of the company, and its $553B RPO (up 325% YoY) is ~8x current revenue, almost entirely large-scale AI contracts (mainly OpenAI Stargate).

Neoclouds & ex-miners — the contracts, with durations. These businesses are essentially securitized compute contracts, so the term length is the whole story. CoreWeave's $22.4B OpenAI commitment runs through May 2031 (built up in three tranches: $11.9B in Mar'25 on an initial 5-year deal, +$4B in May'25, +$6.5B in Sep'25), and its total RPO is $66.8B. Nebius's Microsoft deal is ~$17.4B through 2031 (expandable to ~$19.4B) plus a ~$3B / 5-year Meta deal. Among the ex-miners: IREN–Microsoft is $9.7B over 5 years (20% prepaid); Cipher–AWS is $5.5B over 15 years; Hut 8 has two 15-year Fluidstack/Google-backed leases ($7.0B + $9.8B, with 3% annual escalators); TeraWulf's Fluidstack deals run 10 years (plus extensions, ~$12.8B total, Google-backstopped) and its HPC revenue overtook bitcoin for the first time in Q1'26. The common thread: 10–15-year terms backstopped by a hyperscaler, against which the miner raises debt to buy GPUs.

Glossary — 4 terms
Hyperscaler — Global-scale cloud (AWS, Azure, Google Cloud, Oracle, Alibaba) that owns its DCs, designs custom chips, and sells a broad stack.
Neocloud (GPUaaS) — GPU-as-a-Service pure-play renting AI compute by the hour. Lower margin, less stack, faster to deploy GPUs.
RPO / Backlog — Remaining Performance Obligation — total contracted future revenue not yet recognized. The cleanest leading indicator for cloud demand.
GW / MW — 1 GW = 1,000 MW. A typical AI DC draws 100–500 MW; 1 GW of AI DC costs ~$50–60B to build (~$35B of it chips).

Layer 7 — Data-Center Real Estate & Builders

Layer ProfileL7 · DC REAL ESTATE

What this layer does

The physical buildings and the contractors who build them. Hyperscalers either own DCs (Microsoft, Google, Meta build a lot) or lease from REITs; specialty contractors handle the power/cooling/electrical build-out.

Who pays whom

Hyperscalers + enterprises pay DC REITs monthly rent for space + power. Hyperscalers + REITs pay construction/electrical contractors (Quanta, Comfort Systems, EMCOR) to build. REITs pay L0 for power and L6 for fiber.

Key Metrics to Track

REITs: leased backlog, MW (megawatts — the power a site can deliver, the real unit of data-center capacity) under construction, vacancy (NoVa, i.e. Northern Virginia, the world's largest data-center market, runs <5% = pricing power), and AFFO growth. AFFO = adjusted funds from operations: a REIT's operating cash flow after the maintenance capex needed to keep buildings rentable — it strips out the heavy depreciation that makes REIT net income (and P/E) misleading, which is why REITs are valued on price/AFFO and AFFO growth rather than earnings. Contractors: backlog, book-to-bill, and the data-center/advanced-tech mix within that backlog (rising fast for both PWR and FIX).

Margin that mattersEBITDA margin is the right lens — REITs carry heavy depreciation, so EBITDA / AFFO, not net or operating margin, reflects the economics.
Valuation lens that mattersREITs are valued on P/AFFO and EV/EBITDA, never P/E — heavy depreciation makes net income misleading, and AFFO is the real cash-flow proxy. Also watch dividend coverage and leased-backlog. Contractors (PWR, FIX, EME) trade on P/E + EV/EBITDA against backlog and book-to-bill.

Analyst’s Take

Data-center real estate has quietly become a binding constraint: GPUs cannot be deployed without a powered, cooled shell to house them, and supply is tight — Northern Virginia vacancy sits under 5% and interconnect queues run 7–10 years. The layer splits into two fundamentally different business models. The REITs carry genuine interconnection moats and highly recurring revenue, but they are capital-intensive and trade on real-estate metrics, so a headline P/E overstates how expensive they are — AFFO is the right lens. The construction and electrical contractors are the less obvious exposure: they trade on ordinary earnings multiples and are not yet priced as ‘AI’ names, even as their backlogs become increasingly data-center-driven at high incremental margins.

ReadOn fundamentals the contractors are the more attractive sub-segment — ordinary earnings multiples, not yet re-rated as AI names, with data-center-driven backlogs at high incremental margins — while the REITs offer the durable interconnection moat but are better judged on AFFO than on P/E.

Key risk

Buildout & power access. You cannot energize a shell where there's no power — Northern Virginia vacancy is <5% and grid-interconnect queues run 7–10 years, while electrician and long-lead-equipment (transformer, switchgear) shortages cap how fast contractors deliver even on secured sites.

Bottleneck HIGH

  • Northern Virginia vacancy is <5% and grid-interconnect queues run 7–10 years — you cannot build a DC where there is no power, regardless of capital.
  • Resolution is geographic: build in Texas, Wisconsin, Ohio, New Mexico where power is available; Stargate is explicitly a land-and-power solution. Permitting reform is helping at the margin.
  • Skilled-labor and long-lead-equipment shortages (electricians, transformers, switchgear) cap how fast contractors can deliver even when sites are secured.

Market Size

Share of the $725B capex~$70–85B of the $725B (10–12%, the ‘shell, land & fiber’ bucket), against a researched global DC-infrastructure TAM of ~$200B+[2] — the gap is the non-hyperscaler and maintenance spend.
  • Global DC infrastructure ~$200B+ in 2026E; US colocation market ~$72B.[2]
  • McKinsey models a $5.2T global DC build by 2030, with ~$1.3T (25%) flowing to ‘energizers’ (utility + electrical).[4]
  • REIT capacity is pre-leased years out — Digital Realty had 769 MW under construction at end-2025.[1]

Value Added & Margins

  • MEDIUM-HIGH for REITs, LOWER but steady for builders. Equinix runs ~49% gross / ~51% adjusted-EBITDA margins; Digital Realty ~48% gross.
  • REIT economics are high-recurring but capital-intensive; the durable moat for Equinix is interconnection density (network effects), not just real estate.
  • Contractors (Quanta ~16% gross / 8% operating; Comfort Systems ~22% gross) earn modest margins but have record multi-year backlogs and very high incremental returns on capital.
Growth vs forward valuation — DC real estate (5 names)[7][8]
1.5x2x3x4x5x6x8x10x12x0%20%2-yr forward revenue CAGREV / fwd sales (NTM, log)DLREQIXFIXPWREME

Valuation readDC real estate

Two business models that demand different lenses. The contractors (Quanta ~3×, Comfort Systems ~5×, EMCOR ~1.9× forward sales) look genuinely cheap-for-growth and are not yet priced as AI names, even as their backlogs turn data-center-heavy. The REITs (Equinix, Digital Realty) screen ‘expensive’ at ~11–12× sales, but that is the wrong denominator for a capital-intensive REIT — the right lens is P/AFFO, shown in the companion below, on which both look reasonable for high-single-digit contracted growth. Read the contractors on EV/sales, the REITs on AFFO.

Companion lensREIT forward P / AFFO

On headline P/E the data-center REITs look extreme (~70–100×) because REIT net income is buried under depreciation. AFFO strips that out, and on forward P/AFFO they sit in the mid-20s× for high-single-digit, contracted, recurring growth — reasonable for the interconnection moat, and the reason AFFO (not earnings) is the right lens here. AFFO/share is hand-curated from each REIT's earnings supplement; treat as indicative.

Forward P / AFFO — data-center REITs[1]
DLR26.4×EQIX26.8×
DLR26.4×FY26E AFFO/share ~$7.0–7.5; ~6–8%/yr growth as AI leasing lifts
EQIX26.8×FY26E AFFO/share from guidance (~$37–42 range); ~8–9%/yr growth
REITPriceFY+1E AFFO/shFwd P/AFFOAFFO growth
DLR$192$7.3026.4×~7%
EQIX$1,071$40.0026.8×~9%

Sub-segments

Colocation REITs (EQIX retail, DLR wholesale)
own and lease buildings.
Wholesale developers (QTS/Blackstone, Vantage/DigitalBridge
private) — single-tenant powered shells.
Electrical & mechanical contractors (PWR, FIX, EME)
the physical build-out.
L7 · DC REITWatch: EBITDA margin. Value on price/AFFO and AFFO growth — never P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYEBITDA MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
EQIXEquinix~Pure DC REIT. FY25 revenue $9.2B; record interconnection. AFFO growth 9–13%.~100% (DC)$9.2BFYE Dec ’25$11Bconsensus+22%44%$105.6B74.1×55.7×3.520%11.1×7.4×30.0×38.7×Recurring lease revenue; >90% recurring
DLRDigital Realty Trust~Pure DC REIT. FY25 revenue $6.1B; 769MW under construction (Q4'25).~100% (DC)$6.1BFYE Dec ’25$7Bconsensus+23%60%$68.8B50.9×67.2×13.168%10.9×3.0×30.4×21.1×Multi-year leased backlog; >$1B annual bookings run-rate
L7 · BuilderWatch: operating margin + backlog/book-to-bill. Ordinary P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
PWRQuanta ServicesDC + grid electrical work is a fast-growing slice of Quanta's ~$25B revenue (Electric Power segment); largest US DC electrical contractor.~15% (est.)$28.5BFYE Dec ’25$40Bconsensus+39%6%15%$110.6B101.4×44.8×2.051%3.7×12.3×44.2×106.6×Record ~$35B+ total backlog
FIXComfort Systems USAData centers / tech are now a large and rising share of Comfort Systems' backlog (~$8B revenue); modular + DC mechanical.~35% (est.)$9.1BFYE Dec ’25$14Bconsensus+53%14%24%$66.8B54.8×36.3×1.439%6.6×103.2×39.6×60.8×Record backlog; DC + advanced-tech mix rising
EMEEMCOR GroupEMCOR is a ~$16B-revenue electrical/mechanical contractor; its fast-growing Network & Communications + mechanical-services work is increasingly data-center-driven (a slice of a diversified book).~20% (est.)$17.0BFYE Dec ’25$20Bconsensus+19%9%19%$38.8B29.2×26.9×1.130%2.2×10.0×20.5×43.5×Record remaining performance obligations (RPO), DC-weighted

Notes, deals & disclosures

Why the contractors look 'cheap' next to the REITs. Quanta and Comfort Systems trade on normal-looking earnings multiples and are not pure AI plays — data centers are a (fast-growing) slice of diversified electrical/mechanical contracting books. But their record backlogs are increasingly DC- and fab-driven, and incremental margins are high, so they offer a lower-multiple, lower-glamour way to own the build-out. The REITs (Equinix, Digital Realty) are the purer DC exposure and carry REIT-style multiples (P/AFFO, not P/E), which is why their headline P/E looks elevated.

The AFFO numbers (hand-curated — yfinance does not carry them). AFFO is a non-GAAP REIT metric, so it is not in the live valuation feed; you pull it from each REIT's quarterly earnings supplement, where management reconciles net income → FFO → AFFO. On that basis: Equinix guides to roughly $37–39 of AFFO/share for 2026 and trades around ~22–24× forward AFFO, with AFFO/share growing ~8–9% a year; Digital Realty sits near ~$7.0–7.5 AFFO/share at ~20–23× forward AFFO, with growth re-accelerating toward ~6–8% as AI leasing lifts. The read-through: both look expensive on P/E (heavy depreciation) but reasonable on P/AFFO for high-single-digit, contracted, recurring growth — which is exactly why AFFO, not earnings, is the right lens. (Figures are management-guidance / consensus estimates, refreshed manually; treat them as indicative, not a live snapshot.)

Glossary — 4 terms
Retail colo — Many customers lease individual cabinets/cages; Equinix is #1. Value is interconnection — connecting to many clouds in one place.
Wholesale colo — One large customer leases an entire building / powered shell; Digital Realty is #1.
DC REIT — Real-Estate Investment Trust that owns DC buildings, leases them, and distributes income as dividends.
Powered shell — A building delivered with power and cooling infrastructure but fitted out by the tenant.

Layer 6 — Networking + Silicon Photonics + Fiber + AI-RAN

Layer ProfileL6 · NETWORKING

What this layer does

Connects thousands of GPUs so they behave as one machine. AI training is bandwidth-bound — networking is ~10–15% of cluster cost but mission-critical. Four sub-segments: Ethernet switching, optical transceivers, optical fiber, and AI-RAN (AI compute inside telecom networks).

Who pays whom

Hyperscalers and neoclouds (L8) are the primary buyers — every new GPU cluster needs switches, transceivers and fiber. Switch and optics vendors buy ASICs from Broadcom/Marvell (L4) and lasers/InP from the optics makers.

Key Metrics to Track

A different metric matters in each sub-segment. Switching (ANET, CSCO): track design wins — being spec'd in as the chosen switch vendor for a hyperscaler's next data-center build, each one a contracted future revenue stream — plus overall data-center switch market share. Optics (LITE, COHR, AAOI): capacity utilization, LTA coverage (the share of a vendor's output pre-sold under multi-year long-term agreements, which is what gives revenue visibility), and InP supply — indium phosphide, the substrate the high-speed EML lasers are built on, whose limited fab capacity is the binding bottleneck. Fiber (GLW): optical-segment revenue growth + hyperscaler fiber supply deals. AI-RAN (NOK): quarterly Cloud & AI revenue + the count of carrier pilots converting to deployments. For all four: watch the next GPU-generation interconnect spec — it sets the size of the next demand step-up.

Margin that mattersOperating margin separates the winners — Arista ~43%, Broadcom networking ~50% — from the commodity box-shifters.
Valuation lens that mattersSwitching leaders (Arista) on forward P/E + PEG; optics names (Lumentum, Coherent, AAOI) on EV/EBITDA but with eyes open — they carry the chain's highest cycle risk, trading near ~50× forward earnings on the assumption the LTA-backed boom runs uninterrupted through 2027. The multiple is the risk.

Analyst’s Take

Networking is best understood as two sub-segments with opposite risk profiles. Switching is the higher-quality half: a multi-vendor market, but margin-rich (Arista runs 40%+ operating margins) and only moderately supply-constrained. Optics is the higher-growth, higher-risk half, governed by a simple physics — each GPU-generation step (Blackwell→Rubin) roughly doubles the optical interconnect per accelerator, so demand for EML lasers and transceivers compounds faster than the industry can add indium-phosphide (InP) fab capacity; components are sold out through 2027, which is why names like Lumentum have re-rated by four-figure percentages. The structural-growth story inside optics is silicon photonics and co-packaged optics (CPO): integrating the optics onto silicon and moving them into the switch or GPU package cuts interconnect power by roughly 40%, and it is the only credible architecture for feeding million-GPU clusters before optics power draw and transceiver failure rates become the limiter. Yole sizes silicon photonics at roughly $2.6B today rising toward ~$22B by 2034 (~26% CAGR)[12] — a steeper trajectory than the broader transceiver market, because SiPho/CPO is migrating from a niche to the default high-end architecture rather than simply growing with port counts. The offsetting risk is cyclical and specific: optics multiples (Lumentum near ~50× forward earnings) now discount the long-term-agreement-backed boom running uninterrupted through 2027, and NVIDIA’s $2B equity stakes in both Lumentum and Coherent show that supply — not just demand — is being deliberately engineered, which both validates the cycle and seeds the capacity that historically ends it. Optics has repeatedly over-built into a demand step and then de-rated hard once lead times normalized; the current runway looks long because interconnect content rises every GPU generation, but that eventual normalization is the risk the multiples are not paying for. The other two sub-segments are quieter, lower-beta corners. Fiber (Corning) is a pick-and-shovel on every new campus — steady volume and a re-rating Optical Communications segment, though group margins are diluted by its non-AI display and auto glass, so the AI signal is the segment mix, not the headline. AI-RAN (AI compute placed inside telecom base stations) is a longer-dated option that rides carrier capex cycles rather than the GPU cycle, so it trades on telecom multiples and is best treated as optionality on a >$200B-by-2030 TAM rather than a near-term earnings driver.

ReadOn fundamentals, switching is the higher-quality sub-segment — margin-rich and less exposed to a single supply cycle — while optics offers the sharper growth but at multiples that already discount an uninterrupted boom and carry genuine cycle risk. Silicon photonics/CPO is the structural-growth option within optics, and fiber is the lower-beta way to hold the same buildout.

Key risk

Optics cycle risk. Lumentum trades near ~50× forward earnings (and +1,474% LTM); the optics complement to every GPU upgrade is real, but the multiples assume the LTA-backed boom runs uninterrupted through 2027.

Bottleneck VERY HIGH (optics) / MEDIUM (switching)

  • EML lasers and InP transceivers are sold out through 2027 — each GPU-generation upgrade (Blackwell→Rubin) roughly doubles the optical interconnect needed, so demand compounds faster than capacity.
  • Switching is less constrained (multi-vendor), but the underlying Tomahawk/Jericho ASIC silicon from Broadcom is allocated.
  • Resolution: capacity is being added through 2027 under long-term agreements (and funded partly by NVIDIA's equity investments in Lumentum and Coherent), but it ramps slowly because InP fab capacity has long lead times.

Market Size

Share of the $725B capex~$45–60B of the $725B (6–8%, the networking bucket), versus a researched ~$60B 2026E networking TAM[13] — close, because networking is almost entirely an AI-buildout spend today.
  • ~$60B in 2026E overall.[13] DC Ethernet switching ~$60–65B; optical transceivers ~$22B growing 25%+.
  • Silicon photonics is a distinct, faster-growing slice inside optics, not a synonym for it: Yole sizes it at ~$2.6B in 2025 rising to ~$22B by 2034 (~26% CAGR), versus the broader transceiver market’s ~25%/yr off a far larger base.[12] The steeper curve reflects an architecture shift — SiPho and co-packaged optics moving from niche to the default high-end interconnect — rather than simple unit growth, so its share of total optics spend rises over the decade.
  • Optical components have been the highest-returning sub-segment of the entire chain in 2025–26 (Lumentum +1,474% LTM).[14]
  • AI-RAN is a longer-dated option: cumulative TAM >$200B by 2030.[15]

Value Added & Margins

  • VERY HIGH for optics specialists and switching leaders. Arista ~43% operating margin; Broadcom networking ~50%; Lumentum ~22% and rising on AI mix.
  • The optics names carry the highest cycle risk: extraordinary returns but extreme multiples (Lumentum ~50x forward earnings), all premised on the LTA-backed boom running uninterrupted.
  • Fiber (Corning) is lower-margin at the group level (~diversified) but the Optical Communications segment is re-rating as the AI mix climbs.
  • AI-RAN (Nokia, Ericsson) is the lowest-margin, longest-dated corner — carrier-grade hardware economics (mid-single-digit to low-teens operating margins), monetized on telecom capex cycles, so it adds optionality rather than near-term value capture.
DC Ethernet switching share [13]
Arista19%hyperscaler design wins
NVIDIA (Ethernet)15.2%Spectrum-X
Cisco14%diversified incumbent
Others51.8%white-box + Juniper/HPE etc.
L6 sub-segments — 2026E annual market size ($B) [13]
Switching$62BOptical transceivers$22B↳ Silicon photonics$2.6BOptical fiber$8BAI-RAN$3B
Switching$62BDC Ethernet — the largest slice
Optical transceivers$22Bgrowing 25%+/yr
↳ Silicon photonics$2.6Ba carve-out INSIDE optics; ~26% CAGR to ~$22B by 2034
Optical fiber$8BCorning Optical Communications
AI-RAN$3Bearly stage; >$200B cumulative TAM by 2030
Silicon photonics share — 2025 (~$2.6B → ~$22B by 2034) [12]
Intel21%#1, integrated SiPho
Cisco / Acacia17%
Broadcom (CPO)14%
Lumentum10%
NVIDIA (CPO)8%emerging
Others30%Marvell, Ayar Labs, etc.
Optical transceiver makers — 2025 (~$22B TAM) [12]
Coherent25%InP leader
Innolight (China)24%#1 in modules
Eoptolink13%China challenger
Lumentum12%50–60% EML share
Others26%Fabrinet, Marvell, AAOI
RAN / AI-RAN market share — 2025 [15]
Huawei30%excluded from West
Ericsson24%no NVIDIA tie
Nokia18%NVIDIA $1B / AI-RAN tie
Samsung7%
ZTE + others21%
Optical fiber — Corning's position [1]
Corning30%#1 global; Meta $6B deal
Others70%Prysmian, Sumitomo, Furukawa
Growth vs forward valuation — Networking (12 names)[7][8]
3x4x5x6x8x10x12x15x20x25x30x0%20%40%60%80%2-yr forward revenue CAGREV / fwd sales (NTM, log)ALABAAOIAVGOANETMRVLCRDOLITENVDACOHRGLWCSCONOK

Valuation readNetworking

The richest dispersion in the chain. Astera Labs is priced for perfection — ~24× forward sales, far above the layer line, on ~35% growth — the purest connectivity premium. At the other end, Cisco (~7×/5%) and Corning (~8×/17%) are the value anchors, and Lumentum (~12× on the layer's fastest ~67% CAGR) and Coherent (~8×/34%) screen cheap-for-growth among the optics names. The switching/custom-silicon overlaps from L4 (Broadcom, Marvell, NVIDIA) sit mid-pack. The optics sub-group rewards a closer look — high growth at reasonable multiples — but is gated by InP/EML substrate supply.

L6 · NetworkingWatch: operating margin (40%+ = winner). P/E on design-win momentum.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
AVGOBroadcomAVGO AI semiconductor FY25 $20B (+65%); networking (Tomahawk/Jericho switch ASICs) is a major component. AI switch backlog >$10B; Tomahawk 6 = 102 Tbps. AI semis ≈ 31% of AVGO's $64B FY25 total.~31% (AI semis)$63.9BFYE Oct ’25$159Bconsensus+149%41%$1,984.8B81.4×22.9×0.932%29.1×24.9×55.1×77.8×AI switch backlog >$10B
ANETArista Networks~Pure DC networking (>92% DC-focused). FY25 revenue $9.0B; ~19% DC switch share. Meta + Microsoft are anchor customers.~70%+ (cloud)$9.0BFYE Dec ’25$14Bconsensus+59%43%$195.6B53.4×34.9×2.025%20.1×14.5×44.0×44.8×Deferred revenue + purchase commitments rising
CSCOCisco SystemsDC switching is a modest slice of Cisco's ~$56B revenue; 14% DC switch share, 29.8% total Ethernet. AI orders from webscale crossed $1B+ run-rate.~10% (est.)$56.7BFYE Jul ’25$69Bconsensus+21%22%$470.6B39.7×25.0×1.637%7.7×9.6×28.4×50.7×
NVDANVIDIA pure-playDC networking (Spectrum-X Ethernet + InfiniBand + NVLink) sits inside NVIDIA's $193.7B DC revenue; networking +142% YoY in FY26. ~15% DC Ethernet share.~10% (networking)$215.9BFYE Jan ’26$546Bconsensus+153%60%$5,097.2B32.2×16.6×0.7214%20.1×32.7×31.2×110.0×
L6 · OpticsWatch: operating + gross margin. Mind the multiple vs LTA coverage (cycle risk).
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
MRVLMarvell TechnologyData center ~75% of Marvell's $8.2B FY26 (Jan) revenue; optical DSPs lead the market. Also custom-ASIC (see L4).~75% (data center)$8.2BFYE Jan ’26$15Bconsensus+84%16%51%$175.7B65.6×36.3×2.4106%21.4×11.9×70.1×121.8×
LITELumentum HoldingsDatacom/AI optics now the growth core. FY25 (Jun) $1.6B understates run-rate — TTM revenue +90%, GM ~41%, op margin ~22%. 50–60% global EML share; OCS leader.~70% (est.)$1.6BFYE Jun ’25$6Bconsensus+237%-12%28%$69.6B158.4×49.4×0.628.0×75.8×128.6×745.9×LTAs locked through 2027
COHRCoherent CorpDatacom transceivers a fast-growing share of Coherent's $5.8B FY25 (Jun); InP laser capacity sold out through 2027. TTM GM ~37%.~40% (datacom, est.)$5.8BFYE Jun ’25$9Bconsensus+63%9%35%$74.8B182.0×47.2×0.911.3×7.0×57.8×negInP sold out through 2027
AAOIApplied Optoelectronics800G/1.6T datacom transceivers for hyperscalers; explosive ramp (stock +441% YTD). Smallest, highest-beta optics name.~85% (est.)$456MFYE Dec ’25-11%30%$14.7B38.5×0.829.1×13.2×neg
L6 · FiberWatch: the optical-segment gross margin; the group multiple understates the AI segment.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
GLWCorningOptical Communications ~$5B+ of Corning's ~$14–15B total revenue; Q1'26 optical comm $1.8B (+36%). Enterprise/AI fiber driving the mix up.~35% (Optical, est.)$15.6BFYE Dec ’25$23Bconsensus+45%36%$163.6B91.4×45.0×1.5139%10.0×13.8×45.4×267.3×Multi-year fiber commitments (Meta + others)
L6 · InterconnectWatch: operating margin; growth-adjusted multiples (small, fast-growing).
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
ALABAstera Labs pure-playPure-play PCIe/CXL/Ethernet retimers + fabric ICs for AI racks. ~Pure AI play.100% (pure-play)$853MFYE Dec ’2520%76%$55.1B218.8×76.5×144%55.1×36.8×229.2×229.7×
CRDOCredo Technology pure-playActive Electrical Cables + SerDes for hyperscaler racks; ~pure AI play, hyperscaler-direct.100% (pure-play)$437MFYE Apr ’259%65%$40.1B119.6×39.5×412%37.6×21.8×113.0×233.1×
L6 · AI-RANWatch: operating margin. Telecom multiples — treat as optionality, not a near-term driver.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
NOKNokiaAI & Cloud (data-center networking + cloud) a portion of Nokia's ~€19B revenue; Q1'26 AI & Cloud +49% YoY.~10% (est.)$23.1BFYE Dec ’254%$87.6B98.1×32.2×1.63.8×3.6×30.4×47.2×
ERICEricsson#1 RAN ex-China but no NVIDIA partnership; founding AI-RAN Alliance member. Flat 2026 RAN outlook.~5% (est.)$25.5BFYE Dec ’2513%$43.5B16.1×19.3×1.9-78%0.4×13.7×

Notes, deals & disclosures

The NVIDIA–Lumentum / Coherent investments ($2B each, March 2026) — what they actually are. These are not simple supply contracts. NVIDIA is making an equity investment of $2B in each company (≈$4B total) to fund their build-out of US-based optics manufacturing (new fabs), and the deals carry a multibillion-dollar NVIDIA purchase commitment plus future capacity-access rights — all on a non-exclusive basis. In other words, NVIDIA is using its balance sheet to lock up future optics supply and accelerate domestic capacity, the same playbook as its Nokia ($1B/2.9%), CoreWeave ($2B) and memory-adjacent stakes. For Lumentum and Coherent, it is validation, capital, and a guaranteed anchor customer at once — which is why both stocks re-rated sharply.

Reading the optics financials. Lumentum's and Coherent's fiscal-year figures (FY ends June) predate the optics inflection and understate the run-rate — Lumentum's FY25 revenue of $1.6B sits against TTM revenue growth of +90% and a recovered ~41% gross margin. Use TTM, not the stale FY, for these names. The segment context also matters: Coherent and Corning are diversified (industrial lasers; display/auto glass), so the AI lever is the datacom/optical-communications segment, not the whole company; Arista, Astera and Credo are the cleaner pure-plays.

How much of each company's revenue is actually this layer. Broadcom's networking sits inside its $20B FY25 AI-semiconductor revenue (≈31% of its $64B total); Cisco's DC switching is a small slice of a ~$56B diversified company; Corning's Optical Communications is ~$5B+ of ~$14–15B; NVIDIA's networking is inside its $193.7B Data Center line (but grew +142%). The pure-plays — Arista, AAOI, Astera, Credo — are where a dollar of revenue is almost entirely this layer.

Glossary — 6 terms
Ethernet switch — Routes traffic between servers. Arista + Cisco lead; inside training clusters, NVIDIA InfiniBand competes with Ethernet.
Optical transceiver — Module converting electrical signals to light for fiber. AI clusters moved 800G→1.6T. Coherent + Lumentum lead.
EML (Externally Modulated Laser) — Laser inside high-speed transceivers; enables 200G/lane → 1.6T modules. Lumentum 50–60% share.
OCS (Optical Circuit Switch) — Routes light without converting to electrons — big power saving. Google deployed at scale in TPU pods; Lumentum leads.
SiPho / CPO — Silicon photonics integrates optics onto silicon; co-packaged optics puts them in the switch/GPU package, cutting power ~40%.
AI-RAN — AI-empowered Radio Access Network — AI inference compute inside telecom base stations (NVIDIA + Nokia, Oct'25).

Layer 5 — Servers & Power Supplies

Layer ProfileL5 · SERVERS

What this layer does

Server makers physically assemble GPUs, CPUs, memory, networking and power supplies into rack systems. Distinct from chip makers (L4) — servers integrate the chips someone else designed.

Who pays whom

Hyperscalers (L8) increasingly buy direct from ODMs (Foxconn et al.) and bypass branded OEMs. ODMs buy GPUs from L4, HBM from L3, switches from L6, PSUs from Delta, and assemble. Branded OEMs (Dell, Supermicro) serve enterprises and non-hyperscaler clouds.

Key Metrics to Track

The key variable is gross-margin trajectory, not revenue growth — revenue will look explosive regardless. Watch ODM share gains vs OEMs, working-capital intensity (huge GPU inventory), and customer concentration. Delta (PSUs) is the higher-quality way to play the layer because high-density racks need its efficient power supplies.

Margin that mattersGross margin is the entire thesis: Foxconn's 5–7% gross on AI servers tells you the assembly layer has almost no pricing power.
Valuation lens that mattersEV/EBITDA and P/E look optically cheap and should — the thesis here is revenue beta, not margin. Anchor on ROIC and working-capital intensity (huge GPU inventory) rather than the multiple; a low P/E on a 5–7% gross-margin box-assembler is not 'value'.

Analyst’s Take

Servers are the structural value trap of the chain. Revenue looks spectacular — Foxconn’s AI-server line is compounding rapidly — but at 5–7% gross margins almost none of it reaches the bottom line. The economics are dominated by the chip (40–45% of the bill of materials), ODMs undercut the branded OEMs from below, and hyperscalers deliberately dual-source to keep pricing power near zero. The useful way to read the layer is as a real-time gauge of buildout volume rather than a profit pool: its revenue tracks the GPU cycle almost one-for-one, so it signals the pace of deployment, but it is a top-line indicator, not an earnings story — and margins compress, not expand, as volume scales. The one pocket of better quality is the power-supply specialists, whose high-efficiency PSUs are increasingly content-rich as rack density climbs and which earn more than box assembly.

ReadRead L5 as a top-line proxy for GPU-cycle deployment volume, not as a margin or earnings story: its revenue mirrors the buildout almost one-for-one, but 5–7% gross margins mean little of it becomes profit. The relatively higher-quality niche is the power-supply specialists. The metric that matters is gross-margin trajectory — which compresses as volume scales — not revenue growth.

Key risk

Margin commoditization. Assembly has almost no pricing power (Foxconn ~5–7% gross on AI servers) and hyperscalers dual-source ODMs; revenue scales spectacularly with the GPU cycle, but very little drops to the bottom line.

Bottleneck MEDIUM

  • Server assembly is not the structural bottleneck — it is gated by what flows into it (L4 GPUs, L3 HBM, L2 CoWoS). Once those clear, racks can be built.
  • Foxconn is already assembling GB200/GB300 at scale; multiple ODMs are qualified, so there is no single-vendor choke point here.
  • The real risk in this layer is commercial, not physical: margin compression as volume scales and hyperscalers dual-source.

Market Size

Share of the $725B capex~$60–70B of the $725B (8–10%, the ‘servers, CPUs, NICs, PSUs’ bucket), against a $444B total server market in 2025[2] — most of that TAM is non-AI servers; the AI slice is what the capex funds.
  • $444B server market in 2025 (+80% YoY); GPU-embedded servers are >50% of revenue.[2]
  • ODM Direct (hyperscaler-direct) is 59.4% of the market and rising as hyperscalers cut out branded OEMs.[2]
  • 2026E ~$650–700B as Rubin-class racks (higher ASP per rack) ship.[2]
  • One swing factor for the host-CPU portion of this bucket: agentic workloads are lifting the CPU-to-GPU ratio per server — the full analysis sits in the GPU-vs-CPU box in L4 (where chip economics live), and the takeaway for L5 is simply a modest tailwind to server CPU content.

Value Added & Margins

  • VERY LOW — the structurally worst layer for pricing power. Foxconn earns 5–7% gross margin on AI servers — almost no pricing power.
  • Branded OEMs are squeezed from both sides: ODMs undercut them below, and GPU costs (40–45% of the BOM, the bill of materials) dominate above. Dell's group GM is ~22% but its AI-server margin is materially lower; Supermicro runs ~9.7% gross and falling; HPE's server GM dropped to ~6%.
  • It is a volume/working-capital game: revenue scales spectacularly with the GPU cycle, but very little of it drops to the bottom line. Own this layer for revenue beta, not for margins.
AI-server ODM share — 2025 [2]
Foxconn (Hon Hai)40%GB200/GB300 at scale
Quanta18%
Wistron / Wiwynn14%
Inventec9%
Others19%incl. SMCI ODM
Branded AI-server OEM share — 2025 [2]
Dell35%~$25B AI-server run-rate
Supermicro22%closest pure-play
HPE16%
Lenovo12%
Others15%
Server power-supply (PSU) share [1]
Delta Electronics48%#1 for AI-server PSUs
Lite-On18%
AcBel / others34%
Growth vs forward valuation — Servers (4 names)[7][8]
0.3x0.5x1x2x5x10x0%20%40%2-yr forward revenue CAGREV / fwd sales (NTM, log)2308DELLSMCI2317

Valuation readServers

An intentionally low-multiple, thin-margin layer: Dell (~1.2× forward sales) and Supermicro (~0.6×) trade on box-assembly economics even as AI-server revenue explodes — Dell's surge is already in the trailing year, so its forward CAGR is a modest ~10%. The point of the layer is that scale doesn't convert into a profit pool; read these on AI-server backlog and incremental margin, not the sales multiple. Foxconn (Hon Hai) now plots at ~0.4× forward sales — the thinnest multiple in the entire chain, the ~6%-gross-margin assembly model made visible — while Delta Electronics (~21% growth) is the higher-quality power/thermal play that earns a fuller multiple.

L5 · ODMWatch: gross margin (5–7% = no pricing power). Read ROIC and working capital, not the P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
2317.TWFoxconn (Hon Hai)AI servers a fast-growing slice of Foxconn's ~$220B revenue; cloud & networking ~40% of group. Lead NVIDIA GB200/GB300 rack ODM (part of 59.4% combined ODM share). Razor-thin ~6% gross margin.~40% (cloud/AI)$258.1BFYE Dec ’256%$117.8B19.7×12.3×1.118%0.4×2.1×9.7×99.4×
L5 · OEMWatch: gross-margin trajectory. A low P/E here is not 'value'.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
DELLDell TechnologiesAI server $25B+ FY26 run-rate within Dell's $113.5B FY26 (Jan) revenue (ISG segment); group GM ~22%, but AI-server margin is lower than traditional server.~22% (AI servers, est.)$113.5BFYE Jan ’26$155Bconsensus+37%20%7%$199.0B35.3×20.5×1.357%1.8×-81.3×19.0×30.1×AI server backlog ~$14B
SMCISuper Micro Computer pure-play~Pure-play AI server maker; ~9.5% server share. FY (Jun). Q4 revenue +134% YoY but GM only ~9.7% and falling.~90% (AI servers)$22.0BFYE Jun ’25$52Bconsensus+135%11%6%$22.8B20.0×11.8×0.9326%0.7×3.0×19.1×neg
L5 · PSUWatch: gross/operating margin (the quality niche). P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
2308.TWDelta ElectronicsDC power (PSUs) + cooling + power management a growing share of Delta's ~$13–14B revenue; #1 AI-server PSU supplier.~25% (DC power, est.)$17.7BFYE Dec ’2534%15%$208.5B109.4×39.1×3.7101%11.0×21.9×48.1×201.3×

Notes, deals & disclosures

Foxconn's revenue dwarfs everyone — but read it carefully. Foxconn (Hon Hai) is a ~$220B-revenue company, far larger than Dell or Supermicro, and its cloud/networking + AI-server business is a fast-growing ~40% of that. But at ~6% gross margin, scale doesn't translate into the profit pool you'd expect — which is exactly the layer's lesson. Dell's AI server backlog is ~$14B and its AI-server run-rate is $25B+, but that sits inside a $113.5B company at group margins that mask the thin AI-server economics. Supermicro is the closest pure-play and the highest revenue beta, but also the thinnest margins and a governance overhang. Delta Electronics is the quieter, higher-quality name: PSUs and power management carry better margins than box assembly.

Glossary — 3 terms
ODM — Original Design Manufacturer — designs AND builds hardware on contract (Foxconn, Quanta). Hyperscalers buy direct; ODMs = 59.4% of the server market.
OEM — Original Equipment Manufacturer — branded makers (Dell, HPE, Supermicro) selling to enterprises, often built by ODMs underneath.
PSU — Power Supply Unit — converts wall AC to the DC voltages servers need. Delta is #1 for AI-server PSUs.

Layer 4 — AI Accelerators (Chips)

Layer ProfileL4 · AI ACCELERATORS

What this layer does

The chips that actually do AI math. Every other layer ultimately exists to feed these with power, cooling, memory and data. Three types do the work — GPUs (programmable, general), ASICs (hard-wired, hyperscaler-custom), and a niche of wafer-scale — plus the CPUs that host them and the power ICs that feed them.

Who pays whom

Hyperscalers (L8), neoclouds and L5 ODMs buy GPUs and ASICs from this layer. NVIDIA/AMD pay TSMC (L2) for manufacturing + CoWoS packaging and SK Hynix/Micron (L3) for HBM. Hyperscalers co-design ASICs with Broadcom/Marvell.

Key Metrics to Track

NVIDIA is the master variable. Bull case: continued GPU dominance (Rubin late-2026, post-Rubin through 2028, $200B+/yr DC revenue). Bear case: ASIC substitution accelerates and/or a demand air-pocket hits the most over-earning name in tech. Track DC revenue + guidance, gross-margin trajectory, CoWoS/HBM supply, and ASIC ramp (Broadcom/Marvell backlog) as the share-shift tell.

Margin that mattersGross margin is the pricing-power tell. NVIDIA's ~73% vs the rest of the layer is the single clearest signal of who holds the monopoly.
Valuation lens that mattersNVIDIA on forward P/E + PEG looks deceptively cheap (~17× forward on triple-digit growth → a sub-0.5 PEG). The real debate is not the multiple but earnings durability: NVIDIA captures ~⅓ of all AI capex as profit, and the bear case is that ASIC substitution or a demand air-pocket de-rates the most over-earning name in tech. Watch gross-margin trajectory as the tell.

Analyst’s Take

Accelerators are the economic center of gravity of the entire chain: NVIDIA alone converts on the order of a third of all AI capex into profit, on ~73% gross margins and ~$120B of net income — the single largest profit pool in technology — which is why much of this report is effectively a study of the businesses that orbit this layer. The fundamental tension is ASIC substitution. Every hyperscaler gigawatt running on a custom TPU, Trainium or MTIA is merchant-GPU demand displaced, and Broadcom’s custom-silicon revenue alone has reached $20B. The layer’s economics hold as long as total accelerator demand keeps outrunning combined GPU-plus-ASIC supply, so merchant GPU volume can still rise even as ASIC share grows — the point at which that ceases to be true is the point at which the layer’s profit pool comes under real pressure.

ReadMerchant GPUs and custom ASICs are not an either/or on fundamentals: the GPU leader wins if demand stays general-purpose and time-to-market-sensitive, the ASIC designers win if workload-specific share shifts, and both have been growing because total demand still exceeds combined supply. The single variable to monitor is the GPU-versus-ASIC share split.

Key risk

Hyperscaler ASIC substitution. Google TPU, AWS Trainium ($20B+ run-rate), Meta MTIA and Microsoft Maia are all in production — every ASIC gigawatt is a gigawatt NVIDIA doesn't sell at merchant margins. The deeper risk is a de-rate of the most over-earning name in tech if demand air-pockets.

Bottleneck EXTREME (the bottleneck)

  • NVIDIA Blackwell has been sold out 12+ months ahead and Rubin is already booked — this is the binding constraint of the whole cycle.
  • Crucially, the limit is not chip design — it is the supply of L3 HBM and L2 CoWoS advanced packaging. As those scale through 2026–27, the bottleneck eases.
  • ASIC substitution is the structural relief valve: every TPU/Trainium/MTIA GW deployed is demand NVIDIA doesn't have to supply — but total demand is still outrunning combined GPU + ASIC capacity.

Market Size

Share of the $725B capexThe big one: ~$290–325B of the $725B (40–45%, the largest single bucket), against a ~$300B 2026E accelerator TAM[2] — they essentially match, because the accelerator is the AI capex.
  • ~$300B in 2026E;[2] BofA models a $1.2T accelerator TAM by 2030.[10]
  • NVIDIA FY26 Data Center revenue $193.7B (89.7% of its total); Broadcom AI semiconductors $20B FY25 (+65%); AMD data center $16.6B (+32%).[1]
  • Custom ASICs are the fastest-growing share — Broadcom alone signals multiple 10-GW custom programs through 2030.[1]

Value Added & Margins

  • EXTREME — the highest in the chain. NVIDIA: ~73% gross margin, ~60–65% operating margin, ~$120B net income (FY26).
  • NVIDIA single-handedly captures on the order of a third of all AI capex as profit — the single largest profit pool in AI, and the reason the whole supply chain orbits it.
  • AMD earns lower margins and is CoWoS-gated; Broadcom's custom-AI runs ~50% operating margin; Cerebras is not yet profitable. The accelerator is where pricing power (the highest in the chain, for NVIDIA) and the bottleneck coincide.
Merchant AI GPU share [1]
NVIDIA88%86–90%; ~73% gross margin
AMD6%Instinct MI300/MI450
Others6%Intel Gaudi, startups
Custom AI ASIC (XPU) share — 2025 [1]
Broadcom70%Google TPU, Meta MTIA, OpenAI
Marvell13%AWS Trainium, MS Maia
Alchip / GUC / others17%
Data-center CPU share — 2025 [1]
Intel (Xeon)55%losing share
AMD (EPYC)39%gaining fast
Arm (Graviton etc.)6%hyperscaler in-house
Compute-silicon market size — GPU vs ASIC vs CPU ($B, history & forecast) [1]
$73B2023$142B2024$238B2025$337B2026E
Merchant GPU$260BNVIDIA ~90% + AMD
Custom ASIC (XPU)$45BBroadcom / Marvell-built
Data-center CPU$32BIntel / AMD / Arm host chips
GPU power-delivery (voltage regulation) — 2025 [1]
Monolithic Power (MPWR)55%GPU VR leader
Infineon20%
Renesas / others25%
Growth vs forward valuation — Accelerators (6 names)[7][8]
8x10x12x15x20x0%20%40%60%2-yr forward revenue CAGREV / fwd sales (NTM, log)MPWRAVGOMRVLAMDINTCNVDA

Valuation readAccelerators

Within accelerators the regression actually slopes the right way — higher growth earns a higher multiple. NVIDIA screens cheapest-for-growth (~9.5× forward sales on a ~30% 2-yr CAGR) despite being the layer's anchor, while AMD, Broadcom and Marvell cluster richer (~11–13×) on faster accelerator/ASIC growth. The two names rich for their growth are Monolithic Power (~18× on ~19%, the analog-content premium) and Intel (~10× on ~10%, a turnaround multiple growth has yet to match). The lesson: the merchant-GPU leader is not the expensive name here once you weight for growth.

GPU vs CPU vs ASIC — what each does, and can one replace another?

These are the three chip types that do computation in an AI data center. They are complementary, not interchangeable: a modern AI server uses all three. The investment thesis, the bottleneck, and the profit pool, however, sit overwhelmingly on the accelerator (GPU + ASIC) — which is why this layer is named for accelerators, not CPUs.

CPU — Central Processing Unit
The generalist / orchestrator

A few powerful cores (8–128) optimized for sequential, branch-heavy, latency-sensitive work: running the operating system, orchestrating the job, moving data to the accelerators, handling storage and networking. Analogy: a few experts solving hard problems one step at a time. Every AI server still needs one — in NVIDIA's GB200, the Grace CPU is the 'G' paired with the Blackwell GPU. Examples: Intel Xeon, AMD EPYC, NVIDIA Grace (Arm), AWS Graviton.

Value: Necessary but lower-value: smaller share of the bill of materials, a competitive 3-way market (Intel/AMD/Arm) → thinner margins. Historically ‘not the AI thesis’ — though agentic workloads are now lifting host-CPU content per node (see the ratio note below).

GPU — Graphics Processing Unit
The engine of AI

Thousands of simpler cores optimized for massively parallel math (matrix multiplication). Training and running neural networks is almost entirely matrix multiply across billions of parameters — embarrassingly parallel — so GPUs are 10–100x more efficient than CPUs at it. Analogy: thousands of students each doing simple arithmetic at once. Crucially, GPUs are programmable: the same chip runs any model architecture. Examples: NVIDIA Blackwell/Rubin, AMD Instinct.

Value: Highest value in the chain. ~40–45% of every capex dollar; NVIDIA's ~86–90% share + 73% gross margin = the single largest profit pool in AI.

ASIC — Application-Specific IC
The specialist

A chip hard-wired for one workload (e.g. transformer inference). For that one task it is even more power- and cost-efficient than a GPU — typically 30–50% cheaper per workload — but it cannot flexibly run new model types. Hyperscalers design ASICs for their own huge, stable workloads. Examples: Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia (designed with Broadcom / Marvell, built by TSMC).

Value: The hyperscalers' lever against NVIDIA pricing. Counted as cloud revenue (L8) and as Broadcom/Marvell silicon (L4). The key substitution battle is GPU vs ASIC — not GPU vs CPU.

Can one replace another? CPU vs GPU: no — they are partners, not substitutes. A CPU can technically run AI but is so much slower at the parallel math that it is economically unviable for training or large-scale inference; a GPU is poor at the serial control logic a CPU handles. They sit side-by-side in every node — the CPU orchestrates, the GPU computes. The real substitution battle is GPU vs ASIC, because both do the AI math. ASICs win on cost/efficiency for a fixed, high-volume workload; GPUs win on flexibility, time-to-market, and running any model. A hyperscaler with one enormous stable workload (Google's ranking on TPU) builds an ASIC; everyone else — and anyone needing to run the newest models — buys GPUs. That is why merchant GPU demand keeps rising even as ASIC volumes grow.
The agentic shift in the CPU:GPU ratio — an emerging and under-priced change. Chatbot- and training-era AI servers were built around a few host CPUs feeding many GPUs: a classic HGX box pairs 2 CPUs with 8 GPUs (≈1:4), and even NVIDIA’s Grace-Blackwell racks run ~1 Grace CPU per 2 Blackwell GPUs (1:2). A growing view in the research — pushed hardest by Arm and the CPU vendors — is that agentic AI moves the working ratio toward ~1:1. The reason is architectural: an agent is not one large matrix-multiply but a long, branch-heavy loop — planning, tool and API calls, retrieval (RAG), code execution, validation and guardrails, memory and state, and orchestration across many concurrent sessions. That is serial, latency-sensitive control work — i.e. CPU work — and it scales with the number of agent steps rather than with model size, so as inference shifts from one-shot answers to multi-step agentic workflows, CPU cycles per GPU rise.

My read: the direction is right, but 1:1 is the bullish end of the range — treat it as a trajectory, not a settled number, and note that it does not dethrone the GPU, which still performs the model compute and holds the bill-of-materials and the margin pool. What it changes is the floor under host-CPU and server content: the CPU moves from a rounding-error afterthought toward a structurally larger, faster-growing line.

Market-size implication — and does it fit the current estimate? The capex allocation in §2.1 (the ~$60–70B ‘servers, CPUs, NICs, PSUs’ bucket) and this report’s framing of the CPU as ‘lower-value, not the AI thesis’ are both calibrated to the chatbot-era 1:4–8 ratio. If the shift toward 1:1 plays out, CPU content per AI node — and merchant data-center CPU TAM (Intel DCAI, AMD EPYC, Arm/Ampere, plus NVIDIA’s own Grace) — biases above those numbers; the current model most likely under-weights the CPU line at the margin rather than over-counting it. The caveat is who captures it: a rising share of the ‘CPU’ in AI racks is the GPU vendor’s own Arm part (Grace), so some of the uplift accrues to NVIDIA/Arm rather than to merchant x86 (the standard Intel/AMD server CPUs sold on the open market, versus a cloud’s own custom Arm chips) — the agentic shift is bullish for CPU content broadly, but it is as much an Arm-share story as an Intel/AMD re-rating.
Use cases
  • CPU: Web/app servers, databases, job orchestration, the host in every AI node, light inference. Agentic workloads (tool-calling, RAG, multi-step orchestration) are CPU-heavy and are raising CPU content per node.
  • GPU: Training frontier models; large-scale + flexible inference; the default for anyone without the scale to design a custom ASIC.
  • ASIC: A hyperscaler's own massive, stable workload — Google search/ads ranking (TPU), Amazon's Anthropic training (Trainium).
L4 · GPUWatch: gross margin (the monopoly tell, ~73%). P/E and PEG across the cycle.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
NVDANVIDIA pure-playData Center $193.7B = 89.7% of NVIDIA's $215.9B FY26 (Jan) revenue (91.5% in Q4). 86–90% AI accelerator share. 73% gross margin.~90% (Data Center)$215.9BFYE Jan ’26$546Bconsensus+153%71%60%$5,097.2B32.2×16.6×0.7214%20.1×32.7×31.2×110.0×Blackwell sold out 12+ months; Rubin booked
AMDAdvanced Micro DevicesData Center segment $16.6B FY25 (+32%) = 48% of AMD's $34.6B total revenue (Instinct GPU + EPYC CPU; Instinct not broken out). ~5–7% AI accelerator share.~48% (Data Center)$34.6BFYE Dec ’25$76Bconsensus+119%50%11%$806.2B163.7×38.2×1.291%21.5×12.5×109.4×112.4×
INTCIntelGaudi 3 ~1–2% AI accelerator share; immaterial to Intel's revenue. Story is foundry (L2), not merchant AI chips.~3% (Gaudi, est.)$52.9BFYE Dec ’25$65Bconsensus+23%35%-0%$596.4B77.1×1.411.1×5.4×45.6×neg
L4 · ASICWatch: operating margin on custom silicon. EV/EBITDA.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
AVGOBroadcomAI semiconductor $20B FY25 (+65%; >half of semi revenue; 31% of $64B total). Custom XPUs for Google TPU, Meta MTIA + others. ~50% AI op margin.~31% (AI semis)$63.9BFYE Oct ’25$159Bconsensus+149%68%41%$1,984.8B81.4×22.9×0.932%29.1×24.9×55.1×77.8×Multi-year XPU + AI switch backlog
MRVLMarvell TechnologyData center ~75% of Marvell's $8.2B FY26 revenue; AI ~50% of DC. AWS Trainium + Microsoft Maia custom silicon.~75% (Data Center)$8.2BFYE Jan ’26$15Bconsensus+84%51%16%$175.7B65.6×36.3×2.4106%21.4×11.9×70.1×121.8×
L4 · Wafer-scaleWatch: gross margin (pre-profit). EV/sales — speculative.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
CBRSCerebras Systems pure-playPure-play wafer-scale inference. FY25 revenue ~$510M (+76%); IPO'd May'26 (+68% day 1). Extreme valuation (P/E ~580x, P/S ~100x); not yet GAAP-profitable group-wide.100% (pure-play)$510MFYE Dec ’2539%$55.9B620.3×109.5×-25.3×negOpenAI multi-year compute
L4 · Power ICWatch: gross/operating margin. P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
MPWRMonolithic Power SystemsEnterprise/compute (incl. GPU power delivery) a major and growing share of MPWR revenue; critical voltage regulation for every GPU rack.~30% (est.)$2.8BFYE Dec ’25$5Bconsensus+62%55%26%$79.5B115.9×53.7×2.240%26.9×21.6×93.7×161.7×

Notes, deals & disclosures

How much of each company's revenue is this layer. NVIDIA is ~90% Data Center — effectively a pure AI-accelerator play at $5T+ market cap. AMD's data center segment is $16.6B, only ~48% of the company (the rest is client/gaming/embedded), and Instinct GPUs aren't broken out separately from EPYC CPUs. Broadcom's AI silicon ($20B) is ~31% of a $64B company that is also VMware software and broad-line networking. Marvell's data center is ~75% of an $8.2B company. Monolithic Power and Cerebras round out the layer — MPWR as the GPU power-delivery pick-and-shovel, Cerebras as the speculative wafer-scale pure-play (P/E ~580x, not yet group-profitable).

The AMD–OpenAI warrant is the structure to understand. OpenAI's 6-GW MI450 commitment comes with a warrant for up to 160M AMD shares (~10% of the company) at $0.01, vesting in tranches tied to deployment milestones and AMD's share price (up to $600). It is simultaneously a customer contract and a massive equity alignment — and a template for how compute demand is being financed with equity rather than cash across this cycle.

Glossary — 5 terms
GPU — Graphics Processing Unit — thousands of cores for parallel math; programmable, runs any model. NVIDIA, AMD.
ASIC — Application-Specific IC — hard-wired for one workload; 30–50% cheaper per workload but inflexible. Google TPU, AWS Trainium, Meta MTIA, Microsoft Maia.
CPU — Central Processing Unit — few powerful cores for serial/control logic; hosts and orchestrates the GPUs. Xeon, EPYC, Grace, Graviton.
Blackwell / Rubin — NVIDIA GPU generations. Blackwell (B200/GB200) shipped 2025; Rubin late-2026, the basis of OpenAI's 10-GW deal. Each ~doubles performance.
Wafer-scale — Using an entire silicon wafer as one chip — only Cerebras. WSE-3 has 4T transistors / 900k cores; inference-focused.

Layer 3 — Memory + Storage

Layer ProfileL3 · MEMORY & STORAGE

What this layer does

Memory and storage hold the data AI chips work on. The closer to the GPU, the faster and more expensive: HBM sits on the GPU package; DRAM is system memory; NAND/HDD are bulk storage for the ‘data lakes’ that feed inference.

Who pays whom

Chip makers (L4) buy HBM to package next to GPUs; server makers (L5) buy DRAM and storage. Memory makers buy capital equipment from L1 (Lam etch, ASML, Hanmi/BESI bonders).

Key Metrics to Track

HBM: SK Hynix is the purest play and clearest NVIDIA derivative — track HBM3E shipments, HBM4 qualification (sole vs dual-source), and contract pricing. DRAM/NAND: watch the bit-supply/demand balance and contract-price direction — that, not revenue, is what turns the cycle. For SNDK/WDC/STX, use TTM not stale FY figures (Sandisk's FY straddles the WDC spin-off).

Margin that mattersEBITDA margin reads the cycle best — memory is capital-intensive and cyclical, so D&A distorts net income; watch EBITDA alongside the (huge) operating-margin swings.
Valuation lens that mattersThis is the most misread layer. Do NOT anchor on forward P/E — memory is deeply cyclical, so a low forward P/E (SK Hynix ~5×, Micron ~8×) signals peak earnings the market expects to mean-revert, not cheapness; the multiple is lowest exactly when earnings are about to roll. Value it through the cycle on P/B and EV/EBITDA against mid-cycle earnings, and let contract-price direction (not the P/E) tell you where the cycle is.

Analyst’s Take

Memory is the most interesting valuation puzzle in the chain. The fundamentals are monopoly-like right now — SK Hynix posted 80%+ operating margins at the cycle peak and is sold out of HBM into 2027 — yet the market hands these names mid-single-digit forward P/Es. That tension is the thesis: the market is correctly pricing in mean reversion, because memory has always been cyclical and today’s shortage becomes tomorrow’s glut. My read is that this cycle has an unusually long runway (HBM is structurally tied to every GPU, and HBM4 is the next leg), so the down-cycle is further out than the multiples imply — but I never forget which way this industry eventually breaks.

ReadWithin memory, the Korean HBM leader is the purest expression of the HBM-to-NVIDIA linkage and the US name is the domestic pure-play; both carry the same cyclical signature, so the layer should be sized for cyclicality. The mid-single-digit forward multiples price the eventual glut rather than offering a free lunch — for this layer the analytical question is the timing of the down-cycle, not whether it arrives.

Key risk

Memory is cyclical. Today's sold-out, peak-margin conditions (SK Hynix 80%+ operating margin; NAND contract prices +246% in 2025) have always mean-reverted. The low forward P/Es (5–8×) are the market pricing in the eventual down-cycle, not cheapness.

Bottleneck EXTREME (tied with L4)

  • SK Hynix DRAM, NAND and HBM are all sold out through 2026 — memory is co-binding with accelerators because every GPU needs HBM and the HBM supply is concentrated in three players.
  • Resolution comes from all three makers' capex and the HBM4 ramp (2026–27); Samsung/Micron HBM4 qualifications would add second sources and ease the choke.
  • But memory is cyclical: today's shortage is tomorrow's glut. The tightness is real through 2027 per SK Hynix, but the layer has always mean-reverted.

Market Size

Share of the $725B capex~$70–85B of the $725B (10–12%, HBM + system memory), against a ~$300B+ total memory TAM[3] — AI memory is the fast-growing slice of a much larger commodity market.
  • Memory total ~$300B+ in 2026E. HBM: $35B (2025) → $58B (2026) → $100B (2028).[3]
  • Data centers consume >50% of industry DRAM + NAND for the first time. NAND contract prices rose +246% in 2025.[3]
  • SK Hynix FY25 revenue hit ₩97.1T (~$64.5B, +~50%) with HBM revenue more than doubling.[1]

Value Added & Margins

  • HIGH and rising — but priced for a peak. SK Hynix posted ~80%+ operating margins at the cycle peak; Micron's HBM mix pushed group gross margin sharply higher.
  • The market is paying low forward multiples (SK Hynix ~5x, Micron ~8x forward P/E) precisely because these are peak-cycle earnings the market expects to normalize.
  • NAND/HDD margins (Sandisk hit ~56% TTM gross on a +246% price move) are almost certainly an unsustainable peak — long-cyclicals briefly priced like monopolies.
HBM market share — Q3 2025 [3]
SK Hynix57%HBM3E leader; ~70% of NVIDIA Rubin HBM4
Samsung22%HBM4 qualification in progress
Micron21%only US-listed pure memory maker
DRAM market share — 2025 (~$180B) [3]
SK Hynix36%took #1 in 2025
Samsung34%
Micron25%
Others5%
NAND flash share — 2025 (~$70B) [3]
Samsung33%
SK / Solidigm20%
Kioxia19%
Sandisk14%pure-play
Micron11%
Others3%
HDD market share — 2025 (~$20B) [3]
Seagate43%
Western Digital37%
Toshiba20%
Memory & storage market size by segment ($B, history & forecast) [3]
$105B2023$167B2024$255B2025$317B2026E
DRAM (ex-HBM)$150Bcommodity DRAM
HBM$58B→ ~$100B by 2028
NAND$85B+246% contract price in '25
HDD$24Bnearline for data lakes
Growth vs forward valuation — Memory (6 names)[7][8]
4x5x6x8x10x12x0%20%40%60%2-yr forward revenue CAGREV / fwd sales (NTM, log)STXWDCSNDKMU000660005930

Valuation readMemory

A cyclical layer enjoying an up-cycle, so read the multiples with care. SanDisk (~52% CAGR) and Micron (~26%) screen cheapest on EV/forward-sales (~5.5–5.8×) precisely because the denominator is supercharged by a memory-pricing upswing — peak forward sales make the multiple look low. The HDD names (Western Digital ~10×/32%, Seagate ~12×/31%) carry richer multiples on the AI mass-storage re-rate. The Korean DRAM majors now plot too — SK Hynix (~7×/23%) and Samsung (~4×/18%, the conglomerate discount) — the purest HBM-cycle exposure. The discipline is to apply a mid-cycle, not peak, revenue denominator before calling anything cheap; on through-cycle numbers these screens compress materially.

L3 · HBMWatch: EBITDA margin (reads the cycle) + gross. Through-cycle P/B and forward P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYEBITDA MarginGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
000660.KSSK HynixFY25 revenue ₩97.1T (~$64.5B, +~50% YoY); HBM revenue more than doubled in 2025. 57% HBM share; ~70% of HBM4 for NVIDIA Rubin; 34% DRAM (took #1). Net profit doubled.~100% (memory)$68.0BFYE Dec ’2567%60%$1,114.5B5.9×2.9260%9.5×11.2×33.4×Sold out through 2026; tight through 2027
005930.KSSamsung ElectronicsMemory is a portion of Samsung's ~₩300T+ revenue (also smartphones, displays, foundry). 22% HBM, 33% DRAM. HBM4 sampling/qualifying.~25% (memory in conglom, est.)$233.5BFYE Dec ’2529%39%$1,411.2B5.8×0.2492%5.2×13.3×29.8×
MUMicron TechnologyFY25 (Aug) $37.4B; HBM ramp pushing group GM up sharply. 21% HBM, 26% DRAM; data center >50% of revenue. Sold out 2026.100% (memory)$37.4BFYE Aug ’2549%40%$1,030.4B43.0×8.7×0.3756%17.7×14.2×27.3×356.1×Sold out 2026
L3 · StorageWatch: gross margin (peak, mean-reverting). Through-cycle P/B; use TTM, not stale FY.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
SNDKSandisk pure-playSpun from WDC Feb'25; FY25 (Jun) $7.4B distorted by spin (net loss, neg. EBITDA margin). TTM tells the real story: revenue +251%, GM ~56% as NAND contract prices +246% in 2025. HBF (high-bandwidth flash) AI thesis emerging.100% (NAND pure-play)$7.4BFYE Jun ’2530%$232.9B53.8×9.0×17.7×22.7×41.2×103.1×Sold out 2026
WDCWestern DigitalPost-Sandisk-spin, ~pure HDD + platforms. Nearline HDD is the cheap mass-storage tier for AI 'data lakes'. Sold out 2026.~100% (storage)$9.5BFYE Jun ’25$18Bconsensus+85%39%$183.5B31.9×30.2×0.5483%15.6×25.5×45.7×88.4×Sold out 2026
STXSeagate TechnologyHDD oligopoly (only 3 players); Mozaic HAMR high-capacity drives for AI mass storage.100% (HDD)$9.1BFYE Jun ’25$16Bconsensus+78%35%$194.7B82.1×32.8×0.6108%17.7×413.7×54.8×120.9×Sold out 2026

Notes, deals & disclosures

Pure-play vs conglomerate matters enormously here. SK Hynix is the cleanest HBM exposure — memory is essentially the whole company, HBM revenue more than doubled in 2025, and it supplies ~70% of HBM4 for NVIDIA's Rubin. Micron is the only US-listed pure memory maker. Samsung, by contrast, buries its memory upside inside a conglomerate (handsets, displays, foundry), so the HBM boom is diluted. On storage, the SanDisk spin-off from Western Digital (Feb'25) is why SanDisk's FY25 shows a net loss — the fiscal year straddles the separation; the TTM figures (+251% revenue) reflect the real NAND up-cycle. The whole layer is the textbook cyclical-priced-like-monopoly debate: spectacular current margins, low forward multiples.

Is HBM still cyclical, or has the structure changed? The central debate in this layer. The bull case for a structural break rests on four real differences from commodity DRAM. (1) HBM is sold on long-term agreements with pricing fixed up to a year ahead, not on the spot market — so the violent spot-price swings that historically drove the cycle are dampened. (2) It is co-designed and qualified per customer (each HBM stack is spec'd into a specific GPU), which creates switching costs and makes it behave more like a spec-in component than a fungible commodity. (3) Supply is gated by advanced packaging — TSV stacking and hybrid bonders (Hanmi, BESI), not just wafer capacity — so adding HBM supply is slower and harder to over-build than vanilla DRAM. (4) HBM content per GPU keeps rising (HBM3E → HBM4 → HBM4E, more stacks and taller stacks each generation) and HBM consumes roughly 3× the wafer area per bit of standard DRAM, so HBM growth actually tightens commodity DRAM supply rather than competing with it — both can stay firm at once.

My read on the cyclicality debate: these structural supports are genuine, so this cycle should be longer and shallower than past memory cycles — but ‘no longer cyclical’ overstates it. It is still DRAM, made in the same fabs by the same three players; if AI accelerator demand stalls, HBM capacity and the wafers behind it flood back toward commodity DRAM and the glut mechanism reasserts. The LTA pricing power also erodes the moment a credible second source qualifies — Samsung and Micron passing HBM4 qual is the single most important thing to watch, because dual-sourcing is what historically ends memory pricing power. So the honest framing is a structurally dampened cycle with a long runway (HBM4 in 2026–27), not the abolition of the cycle. The variables that decide it: HBM4 capacity additions vs GPU unit growth, whether LTA pricing holds through the next qualification wave, and how much HBM keeps cannibalizing commodity DRAM wafers.

Glossary — 5 terms
HBM — High Bandwidth Memory — stacked DRAM placed right next to the GPU on-package; every AI accelerator needs it. SK Hynix 57%.
DRAM — Standard system memory; only three makers globally (SK Hynix, Samsung, Micron). HBM is a specialized form of DRAM.
NAND — Flash inside SSDs; persistent storage. Sandisk is the largest pure-play.
HDD — Spinning-disk storage; a 3-player oligopoly (Seagate, WD, Toshiba). Cheap mass storage for AI data lakes.
HBF — High Bandwidth Flash — Sandisk's new NAND-organized-like-HBM architecture for inference; still emerging.

Layer 2 — Foundry & Advanced Packaging

Layer ProfileL2 · FOUNDRY & PACKAGING

What this layer does

Foundries physically manufacture the chips that fabless companies (NVIDIA, AMD, Broadcom) design. Manufacturing happens on a wafer — a thin, round slice of ultra-pure silicon (typically 300mm across) on which hundreds of identical chips are patterned at once, then cut apart; the wafer is the fundamental unit a fab processes and prices. TSMC manufactures essentially all leading-edge AI silicon and holds the CoWoS advanced-packaging monopoly. OSATs (outsourced semiconductor assembly & test firms, e.g. ASE, Amkor) handle overflow and non-AI packaging.

Who pays whom

Chip designers (L4) pay foundries to manufacture; foundries pay equipment makers (L1: ASML for EUV, AMAT/Lam for deposition/etch) and materials suppliers (Hoya mask blanks, Shin-Etsu wafers).

Key Metrics to Track

TSMC is the highest-quality pick-and-shovels name in the chain — track monthly revenue, HPC-mix %, gross margin, capex, and CoWoS capacity commentary. The bear case is single-country (Taiwan) concentration, not competition. OSATs: advanced-packaging revenue mix and capex.

Margin that mattersGross margin is the monopoly tell — TSMC's record ~66% gross is pricing power earned by being the only option at the leading edge.
Valuation lens that mattersTSMC on forward P/E + EV/EBITDA. The special feature is a persistent geopolitical (Taiwan) discount on the multiple relative to the quality of the monopoly — you are paid to hold single-country concentration risk. OSATs trade on ordinary P/E + EV/EBITDA.

Analyst’s Take

Foundry sits at the structural choke point of the chain: TSMC is the only manufacturer that can produce leading-edge silicon and CoWoS-package it at volume, which lets it ration capacity while NVIDIA, AMD and Broadcom queue. Its quality is a function of irreplaceability rather than headline margin — it earns less per dollar of revenue than NVIDIA, but it faces no comparable competitive threat, because Samsung and Intel cannot absorb meaningful leading-edge AI volume. The bear case is therefore not competitive displacement but a single concentration: Taiwan. That geopolitical tail risk sits under the entire stack and is the one factor the foundry’s own execution cannot neutralize.

ReadThe foundry’s economics rest on irreplaceability rather than pricing aggression, and it screens reasonably on ~20× earnings. The risk to underwrite here is geopolitical concentration in Taiwan, not competitive displacement.

Key risk

Taiwan / single-country concentration. TSMC manufactures essentially all leading-edge AI silicon and holds the CoWoS monopoly, and Hoya supplies ~75% of EUV mask blanks from Japan — single-country, single-supplier tail risks that sit under the entire chain.

Bottleneck EXTREME (advanced packaging)

  • TSMC CoWoS is sold out through 2027 despite tripling capacity (35K→130K wafers/month); it is the single biggest physical constraint on Blackwell/Rubin output.
  • TSMC is expanding toward ~170K wafers/month by 2027; AMD and Broadcom get rationed allocation behind NVIDIA.
  • There is no competitive substitute at the leading edge — Samsung and Intel cannot absorb meaningful AI volume, so TSMC is a true single point of dependence (and a Taiwan tail risk).

Market Size

Share of the $725B capexNo direct line in the $725B — TSMC is paid out of the accelerator bucket (it is the manufacturing cost inside each GPU/ASIC). Its own revenue ($121.2B FY25 → ~$160B 2026E)[1] is the better gauge.
  • Pure-play foundry ~$175B (2025);[3] TSMC alone $121.2B FY25 → ~$160B 2026E.[1]
  • HPC (incl. AI) rose to 61% of TSMC’s revenue in Q1’26, up from 51% a year earlier — AI is now the majority of the mix.[1]
  • CoWoS capacity 35K (2024) → ~130K (2026E) → ~170K (2027) wafers/month, still over-subscribed.[1]

Value Added & Margins

  • EXTREME. TSMC posted a record ~66% gross margin in Q1'26 and ~49–51% operating margin; FY25 net income ~$54B.
  • This is the highest-quality monopoly in the chain (pricing power among the highest anywhere): the barrier to replicate leading-edge + CoWoS is decades and hundreds of billions.
  • OSATs (ASE ~$18B, Amkor ~$7B) earn ordinary packaging margins but are leveraged to advanced-packaging capacity growth and CoWoS overflow.
Pure-play foundry market share — 2025 [3]
TSMC69.9%sole leading-edge + CoWoS
Samsung7.2%trailing at leading edge
SMIC5.3%China; mature nodes
Others17.6%UMC, GlobalFoundries, etc.
OSAT / advanced-packaging share — 2025 [3]
ASE (incl. SPIL)30%#1 OSAT
Amkor15%#2; US/Arizona
JCET12%China
Powertech (PTI)9%
TongFu / others34%
TSMC revenue by platform — Q1 2026 (HPC ~61%) [1]
HPC (AI + CPU/GPU/networking)61%AI is the majority, but not all, of HPC
Smartphone25%Apple A-series + modems — mostly non-AI
IoT5%edge / consumer
Automotive5%
DCE / other4%
Growth vs forward valuation — Foundry (6 names)[7][8]
2x3x4x5x6x8x10x12x15x0%20%2-yr forward revenue CAGREV / fwd sales (NTM, log)TSM688981INTC005930ASXAMKR

Valuation readFoundry

A monopoly layer where the leader earns the premium. TSMC plots at ~14× forward sales on ~16% growth — the foundry monopoly's pricing power — while ASE (~3.6×/18%) and Amkor (~2×/9%) are the cheap OSAT packagers and SMIC (~13% growth) carries a policy-driven China premium. Intel (~10×/~10%) reflects a turnaround multiple ahead of its growth. The read: pay up only for the node leader; the back-end packagers are structurally lower-margin, lower-multiple businesses despite solid advanced-packaging growth.

TSMC capacity & AI-mix trajectory

202420252026E2027E
CoWoS capacity (k wafers/mo)35~75~130~170
HPC (incl. AI) % of revenue51%~58%61%+~65%
Revenue (US$B)$90$121~$160~$200
Gross margin~53%~59%~66%mid-60s%

CoWoS advanced-packaging capacity has tripled yet remains sold out through 2027; AI/HPC is now the majority of TSMC's revenue mix. Figures are curated estimates from TSMC guidance and disclosures.[1]

L2 · FoundryWatch: gross margin (the monopoly tell, ~66%). ~20× earnings.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
TSMTaiwan Semiconductor (TSMC)HPC (incl. AI) = 61% of TSMC's Q1'26 revenue (up from 51% a year ago); HPC +20% QoQ. FY25 $121.2B; 69.9% pure-play foundry share; sole CoWoS advanced packaging.~61% (HPC/AI)$121.3BFYE Dec ’25$160Bguide/curated+32%60%51%$2,176.9B35.8×21.5×1.358%16.7×93.0×95.0×CoWoS sold out through 2027
005930.KSSamsung ElectronicsFoundry is a small part of Samsung; 7.2% share; 3nm yield improving, 2nm + HBM4 in focus.~30% (foundry, est.)$233.5BFYE Dec ’2539%13%$1,411.2B5.8×0.2492%5.2×13.3×29.8×
688981.SSSMIC5.3% pure-play foundry share; restricted to 7nm/5nm-class by US export controls (no EUV). China's leading-edge champion.~10% (mature, est.)$9.3BFYE Dec ’2521%11%$172.9B239.7×121.7×20.60%18.0×8.3×28.2×neg
INTCIntel18A ramping; sub-scale at leading edge vs TSMC. US strategic asset (10% US government stake).~15% (foundry, est.)$52.9BFYE Dec ’25$65Bconsensus+23%35%-0%$596.4B77.1×1.411.1×5.4×45.6×neg
L2 · OSATWatch: operating margin (commodity packaging). P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
ASXASE Technology HoldingAdvanced packaging a growing share of ASE's ~$18B revenue; #1 OSAT; gets CoWoS overflow from TSMC.~30% (adv pkg, est.)$20.6BFYE Dec ’2518%8%$85.3B59.8×24.5×5.088%4.0×7.7×neg
AMKRAmkor Technology#2 OSAT (~$7B revenue); Arizona facility co-located near TSMC's US fabs for advanced packaging.~35% (adv pkg, est.)$6.7BFYE Dec ’25$8Bconsensus+26%14%7%$17.8B41.4×29.4×0.8286%2.5×4.0×15.1×neg

Notes, deals & disclosures

TSMC is ~70% of the world's foundry and the only place leading-edge AI silicon gets made — but note the reporting nuance: its ADR (TSM) trades in USD while it reports financials in New Taiwan dollars, which is why Yahoo's headline P/S for it is broken (we recompute it at 16.4x from USD market cap ÷ USD revenue). HPC/AI is now 61% of revenue and climbing. The OSATs (ASE, Amkor) are the secondary way to play advanced packaging — Amkor's Arizona facility sits next to TSMC's US fabs and is a US-onshoring beneficiary, but neither earns TSMC-like margins.

Glossary — 4 terms
Foundry — A factory that manufactures chips for fabless designers. TSMC dominates leading-edge (69.9% pure-play share).
Node (3nm/5nm) — Feature size; smaller = more performance/lower power. Blackwell is 4nm-class; AI leading edge is 3nm in 2026, 2nm next.
CoWoS — Chip-on-Wafer-on-Substrate — TSMC's advanced packaging that stacks HBM next to the GPU. Effectively a TSMC monopoly and the single biggest constraint on NVIDIA output.
OSAT — Outsourced Semiconductor Assembly & Test — packaging specialists (ASE, Amkor) that take CoWoS overflow + most non-AI packaging.

Layer 1 — EDA + Equipment + Materials

Layer ProfileL1 · EDA & EQUIPMENT

What this layer does

The deepest layer — everything upstream of the foundry. Three sub-categories: (a) EDA software where chips are designed (Synopsys, Cadence) + CPU IP (Arm); (b) wafer-fab equipment (ASML, AMAT, Lam, TEL, KLA); (c) specialty materials and niche tools (Hoya mask blanks, Hanmi/BESI HBM bonders, Lasertec, Shin-Etsu wafers).

Who pays whom

Chip designers pay Synopsys/Cadence/Arm to design; foundries (L2) and memory makers (L3) pay equipment makers for WFE and buy materials. This layer sells the tools and inputs that make every chip above it possible.

Key Metrics to Track

ASML is the most defensible monopoly in the chain — the bear case is China exposure (~20–30% of revenue) under tightening export controls, not competition. EDA: recurring revenue + AI-design-tool adoption. Memory-equipment names (Lam, Hanmi, BESI, Advantest) are the cleanest HBM-capex derivatives. Watch order backlogs and China-revenue mix.

Margin that mattersGross margin captures the software/IP economics — EDA runs ~95% incremental gross; for the equipment makers, gross margin shows pricing power on irreplaceable tools.
Valuation lens that mattersEDA (Synopsys, Cadence) earns a premium forward P/E on ~95%-incremental-margin recurring software — value it like SaaS. Equipment makers (ASML, AMAT, Lam, KLA) on P/E + EV/EBITDA against order backlog and book-to-bill; the swing factor on the multiple is China-revenue exposure under export controls.

Analyst’s Take

This layer hosts the most durable monopolies in technology — the bedrock under everything above it. ASML is the purest expression: the sole maker of EUV (extreme-ultraviolet) lithography, with a High-NA (high-numerical-aperture, the next-generation EUV) machine costing $380M and no alternative supplier anywhere. EDA (Synopsys, Cadence) is almost as strong — a ~65%-share duopoly with software economics (~95% incremental gross margins) embedded in every chip designed. The reason these are not crowded trades is that the moats are measured in decades of accumulated know-how, so they look expensive on near-term earnings while being close to undisruptable. The one structural overhang is China: export controls cap the addressable market at the margin.

ReadThese are the most irreplaceable inputs in the chain — sole-source EUV lithography and an EDA duopoly with software-like margins — whose moats are measured in decades rather than quarters; they screen ‘expensive’ precisely because they are close to undisruptable. The niche Asian tool-makers are the higher-beta way to hold HBM-capex growth. The variable to track is China revenue mix under export controls.

Key risk

Single-supplier concentration + China. The most defensible positions are also sole-source (ASML for EUV, Hoya for EUV mask blanks, Lasertec for EUV mask inspection); the swing risk is China exposure (~20–30% of ASML revenue) under tightening export controls.

Bottleneck HIGH but distributed

  • EUV machines (ASML High-NA) carry 18-month lead times; HBM thermal-compression bonders (Hanmi, BESI) are sold out — distinct choke points rather than one binding constraint.
  • Because the constraints are spread across many specialized suppliers, the layer eases gradually as each expands capacity (ASML targets ~$71B revenue by 2030).
  • The deeper risk is concentration: single suppliers for irreplaceable inputs (ASML for EUV, Hoya for ~75% of EUV mask blanks, Lasertec for EUV mask inspection).

Market Size

Share of the $725B capexMostly indirect — L1 is funded by L2/L3 capex, not the hyperscaler $725B. The one direct touch is the ‘software, EDA, security’ bucket (~$25–35B, 3–5%), of which EDA is a slice; WFE TAM is ~$120B.[3]
  • Total $140B+ (2026E): WFE ~$120B (2025), EDA ~$21B, photoresist + silicon wafers ~$15B.[3]
  • HBM TC bonders are growing 50%+/year off a small base; every sub-segment is AI-leveraged.[12]
  • Software (EDA) is tiny in dollars (~1% of the chain) but carries ~95% incremental gross margins.[1]

Value Added & Margins

  • EXTREME at the inputs — the most durable monopolies in the entire chain. ASML: ~51% gross / ~35% operating margin as the sole EUV maker (pricing power the highest in the chain).
  • Synopsys/Cadence are a ~65%-share EDA duopoly with software economics (~35% operating margins, ~95% incremental gross). KLA holds ~50% of metrology. Hoya earns 30%+ EBIT on EUV blanks.
  • These businesses are defensible because the bar to replicate them is decades of accumulated know-how, not capital alone — and many (ASML, Lasertec, Hoya, Hanmi) are effectively sole-source.
WFE equipment share — 2025 (~$120B) [3]
ASML22%sole EUV
Applied Materials18%broadest
Tokyo Electron13%
Lam Research11%etch/memory
KLA7%metrology
Others29%
EDA software share — 2025 (~$21B) [1]
Synopsys32%
Cadence30%
Siemens EDA14%
Others24%Ansys (now SNPS), etc.
Specialty / niche tool leaders — share of their own segment [1]
Lasertec100%Hoya75%Hanmi71%Advantest55%BESI45%Shin-Etsu30%
Lasertec100%EUV mask inspection — effective monopoly
Hoya75%EUV mask blanks
Hanmi71%HBM TC bonders
Advantest55%AI / HBM test (ATE) — vs Teradyne
BESI45%hybrid bonding — Micron HBM4 sole-source
Shin-Etsu30%silicon wafers — #1 of ~5 makers
Growth vs forward valuation — EDA & equipment (16 names)[7][8]
4x5x6x8x10x12x15x20x25x30x40x60x0%20%40%2-yr forward revenue CAGREV / fwd sales (NTM, log)ARM042700BESIKLACCDNS68576146ASML6920LRCXSNPSAMAT80357741Q4063

Valuation readEDA & equipment

A textbook ‘moat, not growth’ layer. The EDA duopoly (Synopsys ~10×, Cadence ~15× forward sales) commands the highest multiples on the most modest growth (~10–12%) — the premium is the near-100% recurring, mission-critical software franchise, not the growth rate. Wafer-fab equipment (Applied Materials ~9×/5%, Lam ~13×/16%, KLA ~16×/14%) has re-rated on AI capex; KLA looks richest for its growth, AMAT the most muted. The big foreign monopolies now plot too: ASML (~14×/12%) anchors the lithography premium, with Tokyo Electron, Advantest and Disco mid-pack, while Hanmi and BESI (~20–25% growth, HBM/hybrid-bonding tooling) sit higher. Arm is the extreme outlier — ~54× forward sales — the IP-licensing model priced for perfection. The near-flat regression is the point: multiple here is driven by durability and visibility, not near-term growth.

L1 · EDAWatch: gross margin (~95% incremental, software economics). P/E and PEG.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
SNPSSynopsysEDA + IP + (now Ansys) simulation; ~35% EDA share; every AI chip is designed in it. ~95% incremental gross margins (software). Broadly AI-exposed; ~35% op margin.~40% (AI-levered, est.)$7.1BFYE Oct ’25$11Bconsensus+51%77%13%$100.9B80.8×30.9×3.5-82%12.6×3.3×68.8×32.6×Multi-year recurring + backlog
CDNSCadence Design Systems~30% EDA share; gen-AI placement/routing tools. Same software economics as Synopsys.~40% (est.)$5.3BFYE Dec ’25$7Bconsensus+32%86%31%$102.5B86.8×39.6×3.623%18.5×15.6×53.0×68.4×Multi-year recurring + backlog
L1 · IPWatch: operating margin (royalty model). High P/E vs growth.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
ARMArm HoldingsRoyalties + licensing on virtually every AI device; Arm cores in NVIDIA Grace, AWS Graviton, Google Axion, mobile. FY (Mar).~35% (est.)$4.9BTTM98%30%$327.5B366.5×100.8×2.846%66.6×39.0×293.9×424.1×Royalty backlog from shipped designs
L1 · WFEWatch: gross/operating margin (pricing power on irreplaceable tools). P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
ASMLASML HoldingSole EUV maker; FY25 ~$38B revenue; 51% GM. China ~20–30% of revenue (export-control risk). High-NA EUV $380M/unit, 18-month lead times.~50% (AI-levered, est.)$38.0BFYE Dec ’2553%35%$616.4B53.3×33.6×2.419%15.7×1414.2×42.2×64.3×Multi-year EUV order backlog; targets ~$71B rev by 2030
AMATApplied MaterialsBroadest WFE portfolio (~18% WFE share). AI-leverage via logic + DRAM/HBM capex; not separately broken out.~45% (est.)$28.4BFYE Oct ’25$42Bconsensus+47%49%30%$355.6B42.1×27.7×1.534%12.3×14.8×38.8×117.0×
LRCXLam ResearchEtch leader; #1 memory equipment — directly levered to HBM/DRAM/NAND capex surge. ~11% WFE share.~55% (HBM/mem, est.)$18.4BFYE Jun ’25$30Bconsensus+65%49%32%$398.6B60.3×40.2×1.941%18.4×37.6×51.3×91.6×
KLACKLA Corporation~50% metrology/inspection share (~7% of total WFE); yield-critical as nodes shrink. ~50% segment GM.~45% (est.)$12.2BFYE Jun ’25$17Bconsensus+39%61%41%$258.0B55.8×39.6×2.112%19.7×44.0×45.1×89.3×
8035.TTokyo Electron~13% WFE share; coat/develop near-monopoly + etch; EUV-adjacent process tools. Reports JPY.~45% (est.)$15.3BTTM45%29%$150.4B42.0×41.0×2.050%9.8×11.5×32.4×104.7×
L1 · TestWatch: operating margin (AI/HBM test mix). P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
6857.TAdvantestDominant HBM + AI-SoC test (ATE); directly levered to HBM stack counts + complex AI chips. Reports JPY.~70% (HBM/AI test, est.)$4.9BFYE Mar ’2557%29%$123.4B52.8×114.1×2.1220%17.4×24.7×36.9×81.7×
L1 · NicheWatch: gross margin (monopoly niches). P/E — illiquid, high-multiple.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
6920.TLasertec pure-play~Monopoly in EUV photomask actinic defect inspection; pure-play to EUV adoption. Reports JPY.~80% (EUV, est.)$1.6BFYE Jun ’2559%$23.3B42.9×38.0×19%14.8×16.5×28.9×67.2×
6146.TDisco CorporationBack-end dicing/grinding niche monopolist; essential to advanced packaging (CoWoS) volume. Reports JPY.~50% (est.)$2.5BFYE Mar ’2571%$46.0B54.1×49.7×2.116.8×12.5×35.1×
042700.KSHanmi Semiconductor pure-play~71% global HBM TC-bonder share; pure-play to HBM stacking. Reports KRW. Tiny base (~$0.8B FY+1 est), huge growth.~90% (HBM bonders)$404MFYE Dec ’2558%$21.2B57.8×1.7-65%63.0×153.9×369.1×TC bonders growing 50%+/yr
BESI.ASBE Semiconductor (BESI)Hybrid bonding leader (next-gen HBM4/advanced packaging); Micron HBM4 sole-source TC bonder. Reports EUR.~60% (adv pkg, est.)$688MFYE Dec ’2563%$25.5B145.5×46.6×2.364%34.6×47.9×107.3×144.5×
L1 · MaterialsWatch: gross margin. P/E (steady, share-of-segment).
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
7741.THoya CorporationPhotomask blanks are a portion of Hoya's revenue (also eyeglass lenses + medical/endoscopes); 60%+ photomask, 75%+ EUV blanks, only validated High-NA blank supplier. Reports JPY.~30% (mask blanks of conglom, est.)$5.9BTTM55%47%$57.0B36.6×39.8×2.68%9.6×8.9×23.1×52.0×
4063.TShin-Etsu Chemical#1 silicon wafer (~33% share) + leading photoresist; broad chemicals conglomerate (PVC too). Reports JPY.~35% (semis materials, est.)$16.1BTTM34%21%$86.2B29.3×22.5×2.0-6%5.3×3.1×14.2×60.0×
QQnity ElectronicsNov'25 DuPont spinoff; ~$4.7B revenue in CMP slurries, photoresists, advanced-packaging materials.~40% (est.)$4.8BFYE Dec ’25$6Bconsensus+23%46%21%$33.4B51.3×33.5×2.8-22%6.7×4.7×26.5×47.5×

Notes, deals & disclosures

This layer is where the durable monopolies live, but most names are diversified — so segment context matters. Hoya, for instance, is a 75%-share EUV-mask-blank supplier, but mask blanks are only part of a company that also makes eyeglass lenses and medical endoscopes; Shin-Etsu's #1 silicon-wafer position sits inside a broad chemicals group. The cleanest pure-plays on AI/HBM capacity are the small Asian specialists — Hanmi (71% of HBM bonders), Lasertec (EUV mask-inspection monopoly), Advantest (HBM/AI-SoC test) — which is also why they are the most cyclical. ASML, Synopsys and Cadence are the high-quality, more-liquid ways to own the layer; their margins (and multiples) reflect genuine monopoly/duopoly economics. Note Synopsys closed its $35B Ansys acquisition in Jul'25, adding simulation to design.

Glossary — 6 terms
EDA — Electronic Design Automation — software to design chips. Synopsys + Cadence duopoly (~65% combined). Every AI chip is designed in it.
WFE — Wafer Fab Equipment — the machines inside a fab. Big 5: ASML (litho), AMAT (broadest), Lam (etch), TEL (coater/etch), KLA (metrology).
EUV — Extreme Ultraviolet Lithography — prints sub-7nm features. Only ASML makes it; High-NA machines cost $380M each. The most defensible position in the chain.
Metrology — Measuring/inspecting chips at each step to catch defects. KLA ~50% share — yield-critical as nodes shrink.
TC bonder — Thermal-Compression bonder — stacks and bonds HBM dies. Hanmi ~71% share; BESI is Micron's HBM4 sole-source and leads hybrid bonding.
Mask blank — The quartz 'stencil' input for patterning chips. Hoya supplies 60%+ of photomask, 75%+ of EUV blanks.

Layer 0 — Power & Cooling

Layer ProfileL0 · POWER & COOLING

What this layer does

The foundational input layer — every other layer needs electricity, and heat must be removed. As rack density climbed from ~10 kW to ~130 kW (Blackwell) toward ~1 MW (Rubin), the entire power-and-cooling chain became a hard constraint.

Who pays whom

Hyperscalers + DC operators pay electricity producers via PPAs (Constellation, Vistra, Talen, NextEra), pay electrical-equipment makers (Schneider, Eaton, Vertiv, GE Vernova) for switchgear/transformers/turbines, and pay cooling vendors (Vertiv, Schneider, Daikin) for thermal management. Fuel-cell and SMR makers (Bloom, BWXT) sell behind-the-meter power.

Key Metrics to Track

Size everything in gigawatts (GW) — power, not dollars, is the scarce unit here. Scale reference: 1 GW is roughly the output of one large nuclear reactor and powers ~750,000–1,000,000 US homes; a single frontier AI campus now draws ~0.5–2 GW, versus ~30–50 MW for a typical data center a decade ago. When a company touts an ‘x GW pipeline,’ that is the cumulative capacity it has under contract or in development — a 10 GW pipeline is ~ten reactors’ worth, multi-year and not yet built, so read it as a backlog/option, not current output. Market sizing: US data centers draw an estimated ~35–40 GW of average load today (~4–5% of all US electricity), and most forecasts have that reaching ~80–100+ GW by 2030 — roughly a tripling, and the single fastest-growing load on a US grid that is only ~1,200 GW of peak capacity growing ~1%/yr, which is why data centers are the swing demand. Per hyperscaler, commitments now run to tens of GW (the Stargate/OpenAI buildout alone targets ~10 GW). What to track: equipment makers — orders, backlog, book-to-bill, lead-time commentary (the pricing-power tell); utilities — contracted PPAs and behind-the-meter co-location GW signed; fuel cells/SMRs — backlog and commercial milestones. Because the layer is supply-constrained, GW-of-backlog growth matters more than current revenue.

Margin that mattersOperating margin is the read for the equipment franchises (Vertiv ~18%); for the utilities, lean on EBITDA and contracted-return visibility instead.
Valuation lens that mattersEquipment franchises (Vertiv, Eaton, GE Vernova) on EV/EBITDA + P/E against backlog and book-to-bill — backlog growth matters more than current earnings in a supply-constrained layer. Utilities (Constellation, Vistra, Talen) on EV/EBITDA + contracted-PPA visibility and rate-base growth, not headline P/E.

Analyst’s Take

The binding constraint of the 2026–27 buildout is shifting from silicon to electrons. Grid-interconnect queues run 7–10 years and transformer lead times exceed 100 weeks, so the scarce input is increasingly the ability to power and cool a building rather than to buy the chips that fill it. The fundamental consequence is that lead times have become pricing power: GE Vernova’s turbines are sold out for years and Vertiv’s orders rose 252% YoY. That makes the equipment makers and on-site-generation suppliers re-rate on backlog and order growth well before it appears in revenue — a different and earlier signal than the regulated utilities, whose returns are rate-capped regardless of how scarce power becomes.

ReadPower is shaping up as the next binding bottleneck of the cycle. On fundamentals, electrical equipment with multi-year backlogs and behind-the-meter generation optionality shows clearer pricing power than rate-capped regulated utilities; the leading indicators are orders and book-to-bill, not current revenue.

Key risk

Power is the binding constraint. Grid-interconnect queues run 7–10 years and transformer/switchgear lead times exceed 100 weeks[16] — power, not silicon, is increasingly the limiter of 2026–27, with US data-center power demand compounding ~15–20%/yr.

Bottleneck EXTREME (the next bottleneck)

  • Grid-interconnect queues run 7–10 years in key US markets; transformer and switchgear lead times exceed 100 weeks — power, not silicon, is increasingly the binding constraint of 2026–27.
  • Resolution is slow and multi-pronged: behind-the-meter generation (Bloom SOFC, on-site gas), nuclear restarts (Three Mile Island), SMRs (BWXT BANR), and new gas turbines (GE Vernova, sold out for years).
  • Power demand is growing 10–100x faster than the HBM/CoWoS choke points it sits beneath — McKinsey models US DC power demand compounding 15–20%/yr through 2028.

Market Size

Share of the $725B capex~$82–105B of the $725B (power & electrical ~$60–75B + cooling ~$22–30B), against a researched ~$80B AI-relevant power-and-cooling TAM[4] — and the fastest-growing line in the funnel.
  • AI-relevant power & cooling ~$80B in 2026E; McKinsey’s $5.2T global DC build by 2030 routes ~$1.3T (25%) to ‘energizers’ (utility + electrical).[4]
  • Generation equipment (gas turbines, grid) and electrical gear (switchgear, transformers, UPS) have multi-year backlogs and the longest lead times in the chain.[1]
  • Cooling is a smaller but fast-growing slice as liquid/direct-to-chip cooling becomes mandatory above ~100 kW/rack.[1]
  • In gigawatts (the unit that actually binds): US data-center load ~35–40 GW today (~4–5% of US electricity) → ~80–100+ GW by 2030 on most forecasts — a ~3× rise and the fastest-growing demand on the US grid.[4] A single frontier campus draws ~0.5–2 GW; per-hyperscaler pipelines run to tens of GW.[1]

Value Added & Margins

  • MEDIUM-HIGH and rising — lead times have become pricing power. Vertiv ~16–22% operating margin (Q4'25 orders +252% YoY); Schneider ~17–18%; Constellation ~30% (utility-style).
  • Equipment makers (Vertiv, Schneider, Eaton, GE Vernova) capture the most because their order books are multi-year and backlogged; utilities earn steadier regulated/contracted returns.
  • Behind-the-meter power (Bloom Energy — GAAP-profitable since Q1'26, 2026 guide $3.4–3.8B +80%) re-rated hardest because it bypasses the grid queue entirely.
Growth vs forward valuation — Power & cooling (15 names)[7][8]
1.5x2x3x4x5x6x8x10x12x15x20x0%20%40%60%2-yr forward revenue CAGREV / fwd sales (NTM, log)CCJBENEEVRTFCELGEVTLNETNBWXTTTSUCEGJCIVST6367

Valuation readPower & cooling

The widest valuation dispersion in the chain. The fuel-cell optionality names sit richest for their growth — Bloom Energy at ~13× forward sales on a ~40% CAGR — while the merchant-power IPPs (Vistra ~3×, Constellation ~3.6×, Talen ~5×) look cheap on EV/sales because their value sits in contracted PPAs and rate base, not revenue growth; judge those on EV/EBITDA and PPA coverage, where EV/sales understates them. Among the electrical/thermal names Vertiv (~7× / 17%) carries the thermal-leadership premium, with Eaton and GE Vernova more moderate, and NextEra screens rich for its growth (~8.6× / 9%) versus the unregulated IPPs. Net: a mid-multiple layer except the fuel-cell call options, and the utilities need a different denominator than sales.

Sub-segments

Electricity generation
compare like-for-like by fleet type. Nuclear (the AI-favored 24/7 carbon-free baseload): Constellation ~33 GW is the #1 US operator, then Vistra ~6.4 GW nuclear (its ~44 GW headline is total fleet incl. gas/coal), Talen ~8 GW (incl. the AWS-tied Susquehanna), with Southern, Duke and Public Service Enterprise as large regulated owners. Renewables-led: NextEra ~70 GW (mostly wind/solar). Equipment for new build: GE Vernova (turbines + grid); fuel: Cameco (uranium).
Fuel cells
Bloom Energy is the clear leader in on-site solid-oxide (SOFC) for data centers (the large majority of deployed DC fuel-cell capacity, e.g. the Oracle 2.8 GW program); Ceres licenses the SOFC tech, Doosan is a Ceres licensee, and FuelCell Energy (molten-carbonate) is a small niche. Unlike grid power, this is a concentrated, Bloom-dominated sub-segment.
Electrical equipment
Schneider (#1 switchgear/UPS), Vertiv (power + cooling), Eaton (switchgear/transformers).
Cooling
Vertiv + Schneider (tied #1), Johnson Controls, Trane, Daikin (liquid/DDC).
L0 · Cooling+PowerWatch: operating margin. Orders/book-to-bill, not P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
VRTVertiv Holdings~Pure DC infrastructure. FY25 $10.2B (+~30%); Q4'25 orders +252% YoY. Liebert power + thermal; 360AI reference designs with NVIDIA.~75% (DC, est.)$10.2BFYE Dec ’25$18Bconsensus+74%19%$123.2B80.6×36.3×1.6136%11.4×31.2×52.5×62.7×Record backlog; book-to-bill >1
L0 · ElectricalWatch: operating margin (lead-time pricing power). P/E + backlog.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
SU.PASchneider ElectricData centers ~16% of Schneider's $46.7B revenue (~$7.6B); #1 in switchgear and UPS, also cooling (Motivair immersion). Reports EUR.~16% (DC of total)$46.7BFYE Dec ’2517%$176.1B33.7×23.4×2.0-6%3.8×6.2×21.3×38.0×Multi-year electrical backlog
ETNEaton CorporationData center ~$8B and rising of Eaton's electrical revenue; switchgear + transformers with multi-year backlogs.~15% (DC, est.)$27.4BFYE Dec ’25$35Bconsensus+29%19%$156.8B39.4×25.7×3.0-9%5.5×8.0×28.0×59.2×Multi-year electrical backlog (transformers 100+ wk lead)
L0 · GenerationWatch: EBITDA margin + contracted PPAs. EV/EBITDA, not P/E.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYEBITDA MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
GEVGE VernovaGas turbines + HVDC grid equipment; DC orders $2.4B in Q1'26. Spun from GE Apr'24.~30% (DC-driven, est.)$38.1BFYE Dec ’25$52Bconsensus+36%6%4%$281.0B30.5×42.7×1.81816%7.1×20.2×82.8×30.2×Multi-year turbine + grid backlog
CEGConstellation EnergyPure power utility — AI is a demand driver, not a reported segment. 33 GW nuclear (#1 US); ~30% op margin (utility-style).~10% (AI PPAs, est.)$25.5BFYE Dec ’25$34Bconsensus+32%23%16%$105.0B25.2×21.4×3.71091%3.5×3.1×16.4×negLong-dated PPAs
VSTVistra44 GW nuclear + gas; AI is demand driver, not a segment. Comanche Peak nuclear.~10% (est.)$17.7BFYE Dec ’25$25Bconsensus+43%28%12%$54.1B26.9×14.6×0.52.8×20.7×11.4×113.4×Long-dated PPAs
TLNTalen Energy10 GW merchant (Susquehanna nuclear); AI is demand driver.~20% (AI co-lo, est.)$2.6BFYE Dec ’25$5Bconsensus+78%19%1%$17.5B11.2×5.4×16.0×35.9×12.6×AWS co-location PPA
NEENextEra Energy70 GW (#1 US renewables); AI is demand driver. Most Google/Meta renewable PPAs.~5% (est.)$27.4BFYE Dec ’25$35Bconsensus+26%59%29%$181.8B22.1×19.8×1.9160%6.5×3.3×20.9×neg~300 GW interconnection pipeline
L0 · Fuel CellsWatch: gross-margin trajectory toward profitability. EV/sales (early-stage).
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYGross MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
BEBloom Energy2026 guide $3.4–3.8B (+80%); reached GAAP profit in Q1'26. Behind-the-meter SOFC power, deployable in ~90 days.~60% (DC/AI, est.)$2.0BFYE Dec ’25$6Bconsensus+216%29%4%$85.2B69.3×1.634.8×93.2×373.5×321.0×Multi-GW pipeline + Oracle/AEP commitments
CWR.LCeres Power Holdings pure-playUK SOFC IP licensor (royalty model); £45M FY26 contracted. Doosan + Weichai are mass-production licensees. Reports GBP.~40% (est.)$44MFYE Dec ’2570%-135%-1012.0×10.2×£45M contracted FY26
336260.KSDoosan Fuel Cell pure-playCeres licensee; mass production from Jul'25; first 9 MW commercial Dec'25. Pure-play fuel cell. Reports KRW.~30% (est.)$318MFYE Dec ’25-14%-23%$5.6B-1299.8×16.0×negEarly commercial pipeline
FCELFuelCell Energy pure-playQ1'26 revenue $30.5M (+61%); 12.5 MW DC power block (Mar'26). Pure-play, sub-scale, cash-burning.~30% (est.)$158MFYE Oct ’25-17%-77%$1.3B-13.5×0.47.7×1.9×negEarly DC pipeline
L0 · NuclearWatch: EBITDA margin + PPA coverage. EV/EBITDA.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYEBITDA MarginOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
BWXTBWX TechnologiesNaval nuclear base + BANR SMR purpose-built for data centers; backlog $7.3B (+50%).~10% (SMR nascent, est.)$3.2BFYE Dec ’25$4Bconsensus+30%17%10%$18.4B53.7×38.6×1.721%5.5×14.4×43.5×111.2×$7.3B backlog (+50%)
L0 · FuelWatch: operating margin; commodity (uranium) — track the contract price.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
CCJCamecoUranium fuel for the nuclear restart/expansion AI is driving; Westinghouse stake. Reports CAD.indirect (uranium)$2.5BFYE Dec ’2518%$47.2B100.4×56.5×1.988%18.5×9.3×72.7×104.6×Long-term contract book
L0 · CoolingWatch: operating margin. Orders/book-to-bill.
TickerCompanySegment / share of revenueAI rev % of totalRevenue FY25Revenue 2026EΔ % YoYOperating MarginMarket CapP/E (TTM)Fwd P/EPEGEPS Gr. YoYP/SP/BEV/EBITDAP/FCFBacklog / RPO
JCIJohnson ControlsYORK chillers + building systems; DC a growing slice of a large diversified base. Invested in Accelsius (two-phase cooling).~10% (DC cooling, est.)$23.6BFYE Sep ’25$27Bconsensus+14%12%$84.3B42.2×24.2×2.339%3.4×6.2×21.9×29.4×
TTTrane TechnologiesChillers + thermal management; DC a growing slice of HVAC revenue.~10% (DC cooling, est.)$21.3BFYE Dec ’25$25Bconsensus+19%19%$102.1B35.3×27.1×2.0-2%4.7×11.9×24.9×38.4×
6367.TDaikin IndustriesDC cooling a small but fast-growing slice of Daikin's ~¥4.5T (~$30B) HVAC revenue; FY30 North America DC target ~$2B. DDC + Chilldyne liquid cooling (2025). Reports JPY.~5% (DC cooling, est.)$29.8BFYE Mar ’258%$43.0B24.9×23.6×2.92%1.4×2.1×10.2×55.4×

Notes, deals & disclosures

Why the utilities don't show an 'AI segment.' Constellation, Vistra, Talen and NextEra are power utilities — AI is a demand driver lifting power prices and contracting their output, not a reported business segment, so there is no clean 'AI revenue' line. The signal instead is the contracts: Microsoft's Three Mile Island restart with Constellation, AWS's 960 MW Susquehanna co-location with Talen, Meta's 1,121 MW Clinton deal. The equipment makers are the more direct plays: Vertiv is the cleanest large-cap DC power-and-cooling pure-play; Schneider and Eaton bury fast-growing DC electrical inside diversified electrification giants (DC is ~16% of Schneider); GE Vernova's gas turbines are sold out for years. Bloom Energy and BWXT are the behind-the-meter/SMR optionality — Bloom's Oracle 2.8 GW Project Jupiter and BWXT's $7.3B backlog are the proof points.

Glossary — 5 terms
PPA — Power Purchase Agreement — long-term contract to buy electricity from a producer (Microsoft–Constellation, AWS–Talen 960 MW).
SOFC — Solid-Oxide Fuel Cell — converts gas to electricity at ~60% efficiency, no combustion, deploys in ~90 days. Bloom is the leader; Ceres licenses the tech.
SMR — Small Modular Reactor — 50–300 MW nuclear units, easier to permit. BWXT's BANR is purpose-built for data centers.
UPS / Switchgear — UPS = backup power during outages; switchgear distributes power inside the DC. Schneider is #1 in both.
Behind-the-meter — On-site generation that bypasses the public grid (and its 7–10-yr interconnect queue).

6 · The AI Deal Network

Three patterns define this cycle: (1) model labs signing decade-long compute commitments larger than their current revenue; (2) chipmakers taking equity in their own customers; and (3) ex-bitcoin miners and neoclouds raising debt against 10–15-year hyperscaler leases. The companion network map visualizes the full set of relationships interactively.

Launch Network Map Opens in a new tab — explore deal flows across all 11 layers

OpenAI's deal stack

  • Microsoft — Largest backer and preferred cloud. The Oct'25 PBC restructuring crystallized its stake at ~27% (~$135B) and locked in a $250B Azure purchase commitment; Microsoft keeps IP and model-access rights through 2032, so its Azure growth and OpenAI's compute bill are now two sides of one contract.
  • NVIDIA — A letter of intent (not yet binding) to invest up to $100B as OpenAI builds 10 GW of capacity — the clearest case of a chipmaker funding its own demand. First gigawatt is slated for H2'26 on Vera Rubin; NVIDIA's equity helps pay for the GPUs OpenAI then buys back from it.
  • AMD — A 6 GW commitment to MI450 accelerators, paired with a warrant for up to 160M AMD shares (~10% of the company) struck at $0.01 and vesting against share-price milestones up to $600. In effect OpenAI is paid in AMD equity for adopting a second GPU source — diversifying away from NVIDIA while handing AMD an anchor customer.
  • Oracle (Stargate) — The Stargate buildout: ~$300B of OCI capacity over five years from 2027, anchored at Abilene, Texas with five further sites. It is the single largest compute contract in the stack and the reason Oracle's backlog re-rated so sharply.
  • Amazon — A $38B, seven-year AWS commitment (Nov'25) — the first OpenAI–Amazon deal, signalling that OpenAI is multi-sourcing cloud beyond Microsoft even as Amazon remains Anthropic's primary backer.
  • CoreWeave — $22.4B of neocloud capacity through May 2031, contracted in three 2025 tranches — OpenAI renting from a debt-funded neocloud that in turn buys its GPUs from NVIDIA, a loop within the loop.
  • Broadcom — ~10 GW of OpenAI-designed custom accelerators co-developed with Broadcom through 2030 — OpenAI's hedge against merchant-GPU pricing and a direct echo of the hyperscalers' ASIC strategy.
  • Cerebras — $10B for ~750 MW of wafer-scale inference capacity, delivered via AWS Bedrock — a bet on non-GPU architectures for high-throughput inference.

Anthropic's deal stack

  • Amazon — Anchor investor and primary cloud. Three equity rounds ($8B + $13B + $25B in Apr'26) total $46B+, making Amazon the largest holder; alongside sit a ~$100B, ten-year AWS commitment and Project Rainier, a 1M+ Trainium2 cluster — so Amazon funds Anthropic and Anthropic spends it back on AWS and Amazon's own ASICs.
  • Google — A ~14% equity stake (~$3B) plus tens of billions of dollars of TPU commitments (1M+ Ironwood chips) — a second hyperscaler backing the same lab, which both diversifies Anthropic's compute and pulls it onto Google's custom silicon.
  • Microsoft — A late entrant on the Anthropic side: $5B (up to $10B) of equity and a $30B Azure commitment on Grace Blackwell / Vera Rubin — notable because Microsoft is simultaneously OpenAI's largest backer, underlining how the hyperscalers now spread bets across competing labs.
  • NVIDIA — Up to $10B, announced jointly with Microsoft (Nov'25) — NVIDIA taking a direct stake in the #2 lab as well as the #1, ensuring equity exposure to whichever model layer wins.
  • Broadcom — A multi-gigawatt custom-silicon program confirmed Apr'26 — Anthropic, like OpenAI, designing its own accelerators with Broadcom to cut dependence on merchant GPUs.

Hyperscalers' deal stack

  • Microsoft — ~27% of OpenAI (~$135B) plus the $250B Azure backlog it generates, and separately $5–10B into Anthropic with a $30B Azure deal. The only player with equity in both leading labs; ~$627B total RPO is the contracted demand behind its capex.
  • Amazon — Largest Anthropic holder ($46B+), a ~$100B/10-yr AWS commitment and Project Rainier (Trainium2), plus a $38B AWS deal with OpenAI (Nov'25). Funds the #2 lab and rents it back AWS capacity and in-house ASICs.
  • Alphabet — The only fully integrated lab-and-cloud: it owns the model (Gemini), the accelerator (TPU/Ironwood) and the cloud, and separately holds ~14% of Anthropic with tens of $B of TPU commitments — so it captures economics at every layer it touches.
  • Oracle — The compute landlord: a ~$300B, five-year Stargate contract with OpenAI anchors a $553B RPO despite Oracle having no model or chip of its own — a pure capacity play funded by debt and customer prepayments.
  • Meta — The outlier — no external cloud and no lab equity. ~$140B of captive 2026 capex builds in-house clusters for Llama, monetized indirectly through ad-targeting uplift rather than API revenue.

NVIDIA's deal stack — the "AI banker"

NVIDIA's non-marketable equity grew from $3.4B (Jan'25) to $22.25B (Jan'26) across 67 venture rounds, with $40B+ more in 2026 YTD — it is now both the largest profit pool in the chain and one of its most active strategic investors. The stakes span its own customers and suppliers:

  • OpenAI — Up to $100B LOI tied to 10 GW — funding its single largest customer's buildout.
  • Anthropic — Up to $10B (jointly with Microsoft) — equity in the #2 lab as well as the #1.
  • xAI — $2B+ — compute-for-equity in Musk's lab.
  • Intel — $5B — a strategic stake in a struggling US CPU/foundry rival-turned-partner.
  • CoreWeave — $2B equity — and ~70% of CoreWeave's operating spend flows back to NVIDIA as GPU purchases.
  • Nebius — $2B — European neocloud capacity.
  • Nokia — $1B / 2.9% — an AI-RAN tie-up putting inference into telecom networks.
  • Lumentum / Coherent — $2B each — equity that funds US optics fabs, plus purchase commitments and capacity-access rights; securing optical supply, not just returns.
  • Synopsys — A strategic stake in EDA — locking in the chip-design tooling at the base of the stack.

Circular financing

What it is. ‘Round-tripping’ = a vendor invests cash in a customer; the customer uses that cash to buy the vendor's product; it counts as both an investment and revenue. The biggest loops: NVDA→OpenAI→NVDA (~$100B), MSFT→OpenAI→Azure/NVDA ($250B), AMZN→Anthropic→AWS ($100B), NVDA→CoreWeave→NVDA ($2B equity + ~70% of CoreWeave's opex flowing back to NVIDIA). The practice is disclosed and legal — but three things make it worth watching.

Why it matters. (1) It inflates the apparent independence of demand. A meaningful slice of the sector's ‘growth’ is the same dollars cycling between a handful of balance sheets, so headline backlog and revenue overstate how broad the customer base really is. (2) It concentrates counterparty risk. If one lab's economics deteriorate, the loss propagates back through the vendor that funded it, the cloud that hosts it, and the neocloud that levered against its contract — the links are correlated, not diversified. (3) It is reflexive in both directions. Vendor equity funds customer compute, which books as vendor revenue, which lifts the vendor's valuation and its capacity to invest again — a flywheel that runs just as fast in reverse if demand stalls.

How to read it. The useful discipline is to strip the circular flows out and ask what end-demand (enterprise API spend, app subscriptions, ad uplift) is left underneath. That residual — not the gross contracted figure — is the real size of the market the $725B of capex is chasing, and the gap between the two is the single most important number in the bull/bear debate (see §5).

7 · Sources & References

Click any footnote number in the text to jump here. External links open the primary source where a public URL exists.

  1. [1] Company filings & earnings — SEC 10-K/10-Q, 20-F, and quarterly earnings releases (FY2025 / Q1 2026) for the named companies.
  2. [2] IDC — Worldwide Quarterly Server Tracker and infrastructure forecasts.
  3. [3] TrendForce — HBM / DRAM / NAND market share, supply and contract-pricing data.
  4. [4] McKinsey & Company — ‘The cost of compute’ and global data-center capital outlook to 2030.
  5. [5] Counterpoint Research — Global GenAI / LLM revenue and AI-app market tracker, Q1 2026.
  6. [6] Goldman Sachs Research — AI infrastructure investment forecasts, 2025–2027.
  7. [7] Yahoo Finance (yfinance) — Live valuation and margin snapshot stored in the project’s SQLite database (see snapshot date on each table).
  8. [8] Simply Wall St — Analyst-consensus revenue forecast tables (sourced from S&P Global Market Intelligence) used for the second-forward-year (FY+2) estimates in the §4 growth-vs-valuation map.
  9. [9] Menlo Ventures / Ramp — ‘The State of Generative AI in the Enterprise’ (2025) and Ramp enterprise card-spend data on LLM API share.
  10. [10] BofA Global Research — Hyperscaler capex vs operating-cash-flow analysis.
  11. [11] Synergy Research Group — Cloud infrastructure services market share and quarterly spend.
  12. [12] Yole Group — Silicon photonics and advanced-packaging market forecasts.
  13. [13] Dell’Oro Group — Data-center Ethernet switching and networking market share.
  14. [14] Bloomberg — Market data, deal terms and capex guidance aggregation.
  15. [15] Omdia — AI-RAN and telecom-infrastructure TAM forecasts.
  16. [16] Bernstein Research — Stacy Rasgon / semiconductor and AI-buildout research on the power constraint.

Disclaimer. Information synthesis for educational and research purposes; not investment advice, and the Analyst’s Take sections are the author’s opinion. Valuation/margin data via Yahoo Finance (snapshot 2026-05-28); segment, contract and backlog data from company filings and earnings releases. Figures are estimates and may contain errors — verify against primary sources before acting.