The AI supply chain
Chips + power → data centers → cloud & applications
Strong industry demand does not guarantee investment returns.
Related terms: HBM, DRAM and more →
01Chip inputs & manufacturingChip designs become products through materials, equipment, wafer fabrication and packaging.Collapse
Business models and earnings drivers
Companies earn revenue from design tools, equipment, materials and manufacturing services. Process complexity affects investment; yields and capacity utilization affect profitability.
Related US listings: ASML · TSM · AMAT · CDNS
View related companies →Key indicators and sources
Initial research · 2026-09-10
Manufacturing is expanding; a universal shortage is not established.
Investment read-through (interpretation): Investigate suppliers that monetize manufacturing complexity; planned capacity must become profitable deliveries.
- TSMC Q2 2026 revenue was US$40.20bn, up 33.7% year over year. Advanced nodes (7nm and below) were 77% of wafer revenue; this is not AI-only revenue. [1]
- ASML reported Q2 sales of €9.326bn. Its plan for 30% more low-NA EUV capacity in 2027 is a capacity plan, not completed output. [2]
Candidates for deeper research: TSMC: advanced manufacturing; ASML: lithography. Investment value depends on earnings, price and the selected security.
Principal risk: Customer deferrals, export restrictions or excess new capacity could weaken returns.
Next check: Compare delivered systems and margins with expansion plans; obtain specific packaging capacity and lead-time evidence.
02Compute, memory & storageProcessors perform computation; memory and storage support data access.Collapse
Business models and earnings drivers
Revenue depends on unit sales, pricing and product mix. Demand growth must be assessed alongside costs and margins.
Related US listings: NVDA · AMD · MU · SNDK
View related companies →SRAM vs DRAM vs HBM and NAND: memory guide →
Key indicators and sources
Initial research · 2026-09-10
Memory tightness is supplier-reported; compute and storage demand is realized.
Investment read-through (interpretation): Memory pricing durability merits closer research. Separate price/mix gains from unit growth, and data-center sales from AI-only sales.
- NVIDIA Q2 FY2027 Data Center revenue was US$89.0bn, up 117% year over year. Growth is evidence of sales, not proof of a GPU-wide shortage. [1]
- In June 24 remarks, Micron said DRAM/NAND demand exceeded supply and expected tightness beyond 2027. That is management’s forecast, not an independently established outcome. [2]
- Sandisk Q4 FY2026 Datacenter revenue was US$2.977bn, up 103% sequentially. AI-only contribution is not isolated. Its ticker is SNDK. [3]
Candidates for deeper research: NVDA and MU illustrate compute and memory exposure. SNDK reports data-center business, which is not a measure of AI-only revenue. This section covers operating exposure rather than buy-price estimates.
Principal risk: Added supply, demand pulled forward or weaker prices could reverse unusually strong memory profits.
Next check: Track price versus bit growth, HBM qualification and new-fab output; normalize margins before projecting returns.
Sources, dates & coverage
03Servers & connectivityServers and networks connect chips into coordinated computing clusters.Collapse
Business models and earnings drivers
Cluster construction drives demand for servers and interconnects. Acceptance, revenue recognition and cash collection determine how orders translate into results.
Related US listings: DELL · SMCI · ANET · COHR
View related companies →Key indicators and sources
Initial research · 2026-09-10
Networking and optics are growing; scarcity needs more evidence.
Investment read-through (interpretation): Look for profitable shipment conversion as clusters scale. Product launches and aggregate growth do not prove pricing power.
- Broadcom Q3 FY2026 AI semiconductor revenue was US$16.7bn, up 221% year over year. It combines custom accelerators and networking. [1]
- Arista Q2 2026 revenue was US$3.0357bn, up 37.7% year over year. This is total company revenue, not an AI-only measure. [2]
- Coherent Q4 FY2026 revenue was US$2.05bn, up 34% on a reported basis. Management identifies AI optical connectivity and is expanding capacity; totals include other markets. [3]
Candidates for deeper research: AVGO: AI sales exposure established, accelerator/network split needed. ANET and COHR: relevant products, but AI earnings attribution still needs work.
Principal risk: Customer concentration, architecture changes and capacity additions could pressure supplier margins.
Next check: Verify customer acceptance, volume shipments and product margins, then compare valuation with sustainable earnings.
04Power, grid & coolingPower supply, grid connections and cooling capacity jointly affect data-center deployment.Collapse
Business models and earnings drivers
Engineering firms provide construction services; equipment firms supply electrical and cooling systems. Project delays, labor and material costs affect profit.
Related US listings: FIX · EME · VRT · BE
View related companies →Transformers: demand, delivery and earnings →
Key indicators and sources
Initial research · 2026-09-10
Power and cooling show execution friction; market-wide scarcity remains unestablished.
Investment read-through (interpretation): Investigate suppliers that shorten time to usable power or cooling. Installation acceptance and cash collection matter more than announcements.
- Vertiv Q2 2026 sales were US$3.274bn, up 24% (18% organic). It reported minor timing shifts from temporary supply-chain congestion and phased projects. [1]
- Bloom Q2 2026 revenue was US$1.0654bn and operating cash flow US$226.4m. These are company-wide totals, not AI-only results or proof of all sites being commissioned. [2]
- DOE’s July 9 draft-study announcement describes additional transmission needs and congestion concentrated in limited hours. It does not establish constant nationwide scarcity. [3]
Candidates for deeper research: VRT and BE: deeper research on delivery economics. Existing Transformer research remains a separate, nationally inconclusive subtopic.
Principal risk: Permitting, interconnection, fuel costs, installation delays and competing supply can undermine the benefit.
Next check: Track accepted installations, energized capacity and cash conversion; reconstruct BE’s 2025 case separately using information available then.
05Data-center deployment & operationData centers provide space, colocation services or computing resources.Collapse
Business models and earnings drivers
Revenue depends on commissioned capacity, utilization and service pricing. Financing and operating costs affect earnings and cash flow.
Related US listings: DLR · EQIX · NBIS · IREN
View related companies →Key indicators and sources
Initial research · 2026-09-10
Contracts are growing; delivery and financing determine equity value.
Investment read-through (interpretation): Prioritize conversion of commitments into usable capacity and recurring cash flow. Backlog, annual rent and power capacity are not interchangeable.
- CoreWeave Q2 2026 revenue was US$2.575bn, alongside a US$626m net loss and US$640m net interest expense. Active power was 1.5GW versus about 3.7GW contracted. [1]
- Digital Realty’s Q2 new leases had a nine-month weighted-average contractual commencement lag. Signed-not-commenced backlog was US$1.4bn annualized GAAP base rent at its ownership share. [2]
- Equinix Q2 2026 revenue rose 16%, including one-time xScale fees; monthly recurring revenue grew 11%. Headline and recurring growth differ. [3]
Candidates for deeper research: CRWV: AI capacity, with financing scrutiny first. DLR/EQIX: lease and recurring-revenue pathways, with AI tenant share unisolated.
Principal risk: Delays, debt refinancing, dilution or equipment obsolescence may absorb revenue growth.
Next check: Reconcile project start dates, active capacity, debt maturities and cash after expansion; do not rank operators by backlog alone.
06Cloud, models & applicationsCloud services, models and applications create downstream demand for computing resources.Collapse
Business models and earnings drivers
Paying demand and customer retention support revenue; inference costs and infrastructure investment affect business returns.
Related US listings: MSFT · AMZN · GOOGL · META
View related companies →Key indicators and sources
Initial research · 2026-09-10
Cloud demand monetizes, but investment consumes substantial cash.
Investment read-through (interpretation): Test returns after capacity and replacement spending, not usage growth alone. Whole-company cash flow does not isolate AI returns.
- Microsoft FY2026 Q4 Cloud revenue was US$59.3bn, up 27%. Whole-company quarterly free cash flow was US$19.6bn; Cloud includes non-AI products. [1]
- Amazon Q2 2026 AWS revenue was US$42.2bn, up 37%. Whole-company trailing-12-month free cash flow was negative US$7.6bn, including retail and other operations. [2]
- Alphabet Q2 2026 Cloud revenue was US$24.768bn. Companywide quarterly free cash flow was negative US$5.855bn: US$39.069bn operating cash flow less US$44.924bn PP&E purchases. [3]
Candidates for deeper research: MSFT, AMZN and Alphabet (GOOGL/GOOG): monetization research candidates. Cloud totals are broader than AI; compare cash-flow periods consistently.
Principal risk: Falling inference prices, customer optimization and high replacement spending could weaken returns despite usage growth.
Next check: Track paid adoption, cloud margins and cash after investment together; analyze incremental AI economics rather than total-company earnings alone.
Research framework and coverage
Demand changes → order intake → earnings and cash conversion → valuation conditions → counterevidence and risks