How to Analyze Semiconductor Stocks: Margins, CapEx & FCF

Semiconductor stocks cannot be analyzed with one margin multiple or one quarter of free cash flow. Memory, foundry, fabless, intellectual-property, and hybrid companies convert demand into revenue and cash in different ways. The practical solution is to identify the business model, bridge revenue and gross margin, separate reported from normalized free cash flow, and test returns through a full cycle.

This guide provides a reproducible framework. It complements the AI Infrastructure Research Hub, the general free cash flow analysis guide, and the ROIC versus WACC guide. Those pages explain the wider ecosystem and general formulas; this article focuses on the adjustments that matter specifically for semiconductor businesses.

Written and reviewed by: Shingo Harada, Founder and Author, Stock Metric Lab Data cutoff and last reviewed: September 18, 2026

Key takeaways

Start with the economic model. A memory producer, a foundry, a fabless designer, an IP licensor, and a mixed hardware-software company should not be compared as if they owned the same assets or faced the same cycle.
Decompose revenue into volume, average selling price, and product mix. Revenue growth without that bridge can conceal falling unit economics or a temporary price spike.
Build a gross-margin bridge using pricing, mix, cost per unit, utilization, yield, start-up costs, and inventory charges. A high margin at peak utilization is not automatically a sustainable margin.
Calculate reported cash free cash flow first. Then construct normalized free cash flow with explicit adjustments rather than relabeling a favorable quarter as normal.
Treat maintenance, recurring technology-transition, and expansion CapEx separately. When companies do not disclose the split, present a range and label it as an estimate.
For fabless companies, low property and equipment spending does not eliminate capital commitments. Inventory, prepayments, supply agreements, R&D, stock compensation, and dilution still matter.
Compare ROIC and incremental ROIC through a cycle, then use scenario valuation or a Reverse DCF Calculator to test what the share price requires.

The one-minute semiconductor analysis workflow

Classify the company as memory, foundry, fabless/platform, IP/licensing, or hybrid.
Collect at least five years of quarterly filings when available, or enough history to cover an expansion and contraction.
Bridge revenue using volume, average selling price, and mix rather than relying on total growth alone.
Reconcile gross-margin movement to price, cost, utilization, yield, start-up expense, and product mix.
Calculate reported cash FCF from the cash-flow statement and identify working-capital timing effects.
Estimate maintenance, technology-transition, and expansion investment as separate ranges.
Normalize operating profit, taxes, depreciation, CapEx, and operating working capital.
Test through-cycle ROIC, incremental ROIC, and balance-sheet resilience.
Value downside, base, and upside cases without annualizing a peak quarter.
Record what evidence would disconfirm the thesis before making a judgment.

Step 1: identify the business model

The first analytical error is treating every semiconductor company as a chip manufacturer. The value chain contains distinct economic models.

Model	Revenue engine	Main cycle indicators	Capital that can be overlooked	Useful site examples
Memory	Bits shipped x ASP x product mix	DRAM/NAND pricing, inventory, utilization, cost per bit, HBM mix	Fabs, equipment, technology transitions, inventory write-downs	SK hynix and Micron
Foundry	Wafer capacity x utilization x revenue per wafer	Node mix, utilization, yield, advanced packaging, overseas-fab dilution	Fabs, equipment, construction in progress, start-up costs	TSMC
Fabless/platform	Units or systems x price plus software or services	Product transitions, customer concentration, allocation, lead times	R&D, inventory, prepayments, supply commitments, SBC	NVIDIA and AMD
IP/licensing	License fees plus royalties tied to customer shipments	Royalty rate, chip value, end-market mix, design wins	R&D, SBC, contract assets, customer concentration	Arm
Hybrid	Semiconductor products plus software or infrastructure	Segment mix, custom-silicon programs, renewals, acquisition integration	R&D, inventory, debt, acquired intangibles, restructuring	Broadcom

This classification controls the rest of the model. High cash conversion at an IP company and heavy reinvestment at a foundry can both be economically rational. The relevant question is whether each company earns enough on the capital it actually needs to defend and expand its competitive position.

Step 2: build a revenue bridge

A useful starting identity is:

Revenue change is driven by volume, average selling price, and product mix.

These effects interact, so a company may disclose ranges rather than precise contributions. Preserve those ranges. Do not manufacture false precision by forcing every movement into an exact number.

For memory, separate DRAM and NAND and look for both bit shipments and ASP. For a foundry, use wafer shipments or capacity, utilization, and revenue per wafer when disclosed; node and customer mix may be more informative than wafer count alone. For fabless platforms, separate data-center, client, gaming, embedded, networking, and software revenue rather than treating all growth as one demand signal.

Ask three questions:

Did customers buy more physical output?
Did price per unit rise or fall?
Did the company sell a richer or poorer product mix?

A price-led upturn with flat bits has different durability from a volume-led expansion supported by new applications. Conversely, volume growth can destroy value if it requires price concessions or low-return capacity.

Step 3: construct a gross-margin bridge

Gross margin is where semiconductor cycle, technology, and manufacturing economics meet. Reconcile the change with the following drivers:

Selling prices and contractual price resets.
Product, node, customer, and geographic mix.
Cost per bit, die, wafer, package, system, or license.
Factory utilization and fixed-cost absorption.
Yield improvement or deterioration.
New-node and overseas-fab start-up costs.
Inventory write-downs, excess purchase obligations, and later reversals.
Subsidies or incentives, with the company’s accounting treatment stated.

Label the evidence. A stated company margin is a reported fact. Subtracting two disclosed margins is a calculation. Assigning half of the change to better mix without a company bridge is an estimate. Concluding that a margin is sustainable is an analytical judgment.

Avoid assuming that a higher margin is always structurally better. A memory producer can reach exceptional margins when supply is tight, while a foundry can face temporary dilution during a valuable new-node ramp. The analyst must distinguish temporary cycle effects from durable cost, technology, and mix advantages.

Step 4: reconcile cash conversion

Begin with a transparent reported measure:

Simple cash FCF = cash from operating activities – cash purchases of property and equipment

If the cash-flow line also includes intangible assets, say so. If the company defines its own non-GAAP FCF, reproduce its definition before using it. The FCF Yield Calculator can compare a selected cash-flow measure with enterprise or equity value, but the chosen numerator must remain consistent.

Then inspect the bridge from earnings to operating cash flow:

Accounts receivable and customer-financing terms.
Inventory build, write-downs, and supplier deposits.
Accounts payable and accrued capital expenditure.
Customer advances or deferred revenue.
Tax payments, withholding, and one-time settlements.
Stock-based compensation and the resulting dilution.
Restructuring, acquisition, and integration cash costs.

Working capital can make peak-cycle cash generation look stronger or weaker than the underlying economics. A fabless company can also shift financing into supply commitments and prepayments instead of owned factories. Those contractual claims are not current CapEx and should not be mechanically deducted from FCF, but they do affect downside resilience and future cash needs.

Step 5: separate maintenance, technology-transition, and expansion CapEx

For a semiconductor manufacturer, simply calling all CapEx either maintenance or growth is too crude. Use three buckets:

Maintenance CapEx: spending needed to keep existing productive capacity safe and operational.
Recurring technology-transition CapEx: spending needed to migrate processes, improve yields, and remain competitive even if total capacity does not grow.
Expansion CapEx: spending intended to add economically useful capacity or enter a new product area.

Companies rarely provide a complete split. Use disclosed capacity plans, construction in progress, depreciation, equipment ages, node ramps, subsidy schedules, and management commentary to construct a range. A depreciation proxy can be a starting point, but it is not a fact: historical-cost depreciation may understate replacement cost, while accelerated depreciation can temporarily overstate it.

For subsidies, show both gross cash investment and a clearly labeled net-incentive sensitivity. Cash receipts may not match the period or asset that generated the related expenditure. Never present subsidy-net CapEx as if it were standardized GAAP CapEx.

Definitions and formulas for normalized FCF and ROIC

Reported cash FCF answers what the cash-flow statement produced in the selected period. Normalized FCF estimates repeatable cash generation after removing cycle and timing distortions.

Normalized FCFF = normalized EBIT x (1 – normalized cash tax rate) + normalized D&A – normalized cash CapEx – normalized change in operating working capital

Normalize each input independently. Do not start with a target FCF margin and reverse-engineer convenient assumptions.

For asset-heavy companies, calculate through-cycle ROIC:

ROIC = normalized NOPAT / average invested capital

Also examine incremental ROIC over several years:

Incremental ROIC is approximately change in normalized NOPAT / change in invested capital

Invested capital should use a consistent treatment of leases, goodwill, excess cash, construction in progress, and capitalized R&D. The ROIC versus WACC guide explains the general definitions and estimation limits. In semiconductors, multi-year measurement is essential because new fabs, nodes, and product platforms may incur cash spending long before utilization and profit mature.

A hypothetical normalization example

Assume a foundry reports $10.0 billion of revenue, $3.0 billion of EBIT, $2.0 billion of D&A, $4.0 billion of cash CapEx, and a $0.4 billion working-capital release.

Reported operating cash flow is not automatically normalized because the working-capital release may reverse.
Suppose the analyst estimates normalized EBIT of $2.5 billion, a 15% cash tax rate, normalized D&A of $1.9 billion, recurring cash CapEx of $2.8 billion, and no normalized working-capital benefit.
Normalized FCFF would be $2.5bn x (1 – 15%) + $1.9bn – $2.8bn = $1.225bn.

Every input after the word “suppose” is an estimate, not a reported fact. A better model would show a range for margin, tax, CapEx, and utilization and explain what operational evidence supports each case.

Worked examples from 2026 company disclosures

The examples below illustrate method, not cross-company ranking. Periods, currencies, accounting standards, and FCF definitions differ. All calculated values are Stock Metric Lab calculations unless identified as company-defined.

Micron: why a peak memory quarter needs two CapEx views

Reported facts: For the nine months ended May 28, 2026, Micron reported revenue of $78.959 billion, operating cash flow of $45.702 billion, cash purchases of property, plant and equipment of $19.602 billion, and government-incentive proceeds of $2.989 billion. Its fiscal third-quarter GAAP gross margin was 84.6%. The company said DRAM revenue rose 67%, driven largely by ASP in the low-60% range, while NAND revenue rose 99%, with ASP in the mid-80% range.

Calculation: Gross-basis simple cash FCF was $45.702bn – $19.602bn = $26.100bn, or 33.1% of nine-month revenue. A separate subsidy-net sensitivity was $45.702bn – ($19.602bn – $2.989bn) = $29.089bn, or 36.8%.

Interpretation: The two answers should not be blended. The subsidy-net figure is a non-GAAP sensitivity, and incentive receipts may not match the timing of the funded assets. Micron also reported a $19.953 billion operating-cash-flow use from accounts receivable. Do not annualize either nine-month FCF result or treat the 84.6% gross margin as a normal-cycle assumption.

TSMC: current FCF can coexist with heavy technology investment

Reported facts: TSMC reported second-quarter 2026 revenue of NT$1,270.38 billion, gross margin of 67.7%, operating cash flow of NT$783.365 billion, capital expenditures of NT$496.002 billion, and depreciation and amortization of NT$198.538 billion. The company displayed FCF of NT$287.36 billion.

Calculation: NT$783.365bn – NT$496.002bn = NT$287.363bn of simple cash FCF, a 22.6% margin. CapEx was 2.50 times depreciation and amortization.

Interpretation: CapEx above depreciation does not prove overinvestment; it may reflect leading-edge demand and technology migration. It does show why a durable return analysis needs future utilization, yield, node mix, overseas-fab economics, and incremental ROIC. TSMC’s comments about future N2 and overseas-fab margin dilution are management outlook, not reported historical results.

NVIDIA and AMD: asset-light does not mean commitment-light

Reported facts: For the first half of fiscal 2027, NVIDIA reported revenue of $177.837 billion, operating cash flow of $74.421 billion, and $4.434 billion of purchases of property, equipment, and intangible assets. It also disclosed $279 billion of supply and capacity commitments. For the first half of fiscal 2026, AMD reported revenue of $21.789 billion, operating cash flow of $5.321 billion, and $1.197 billion of cash property and equipment purchases.

Calculations: NVIDIA’s simple cash FCF was $69.987 billion, or 39.4% of revenue. The $4.434 billion cash-investment line was 2.5% of revenue, but it includes intangibles and is not directly comparable with a PP&E-only line. AMD’s simple cash FCF was $4.124 billion, or 18.9% of revenue. AMD’s $4.925 billion of first-half R&D equaled 22.6% of revenue.

Interpretation: NVIDIA’s $279 billion commitment is not current CapEx and should not be subtracted mechanically from FCF. It does reveal a different form of capital exposure. AMD’s second-quarter gross-margin improvement to 54% from 40% included the absence of a prior-year export-control inventory and related charge as well as product mix. For both, test inventory, receivables, prepayments, commitments, R&D efficiency, SBC, and fully diluted FCF per share.

Broadcom: segment mix changes the meaning of consolidated FCF

Reported facts: Broadcom reported fiscal third-quarter 2026 revenue of $29.591 billion, operating cash flow of $14.197 billion, and company-defined non-GAAP FCF of $13.665 billion. Semiconductor revenue was $20.839 billion and infrastructure-software revenue was $8.752 billion. GAAP gross margin was 69%.

Calculations: The reported company-defined FCF margin was 46.2%. The difference between operating cash flow and company-defined FCF implies about $0.5 billion of capital expenditure, but that amount is derived rather than a separately cited cash-flow line here.

Interpretation: Consolidated FCF and gross margin include a software business, acquisition-related amortization, and different revenue timing. They should not be compared directly with a pure-play semiconductor company. Analyze semiconductor and software economics separately, then reconcile debt, acquisition costs, and corporate cash flows.

SK hynix: know when not to calculate FCF

Reported facts: SK hynix’s preliminary second-quarter 2026 release reported revenue of KRW79.3187 trillion, operating profit of KRW60.5426 trillion, operating margin of 76%, cash of KRW88.0 trillion, and debt of KRW18.6 trillion. It described rising DRAM and NAND prices and a richer mix of HBM, AI-server DRAM, and enterprise SSD products.

Limitation: The release did not provide a full cash-flow and capital-expenditure table. Therefore it supports analysis of pricing, mix, balance-sheet position, and capacity discipline, but not a reproducible FCF calculation. The release was preliminary and had not completed independent-auditor review. The SK hynix analysis provides the broader ADR and company context.

Applying the framework across the AI semiconductor cluster

Use the same workflow, but emphasize different evidence for each company:

NVIDIA: peak gross margin, customer concentration, inventory and purchase obligations, supply commitments, networking and software mix, and platform durability.
AMD: data-center mix, product transitions, foundry dependence, inventory, R&D productivity, SBC, dilution, and incremental ROIC.
Arm: royalty economics, license timing, end-market units, royalty rates, R&D, SBC, customer concentration, and controlled-company considerations.
Broadcom: custom silicon, networking, software renewals, acquisition amortization, leverage, and segment cash conversion.
TSMC: utilization, yield, advanced-node and packaging mix, overseas-fab dilution, technology-transition spending, and return on new capacity.
SK hynix: HBM mix, DRAM and NAND pricing, inventory, cost per bit, capacity discipline, and through-cycle normalized FCF.

The NVIDIA versus AMD versus Broadcom comparison shows how apparently similar AI exposure can produce different economics. The AI CapEx evaluation framework examines the customer side: supplier growth is more durable when customer investment also earns acceptable returns.

Through-cycle valuation and reverse DCF

A semiconductor valuation should not rely on one peak or trough input. Build at least three cases.

Input	Downside question	Base-case question	Upside question
Revenue	What happens if price, utilization, or product ramps disappoint?	What does a mid-cycle volume, price, and mix path imply?	What evidence supports durable share or category growth?
Gross margin	How far could price, utilization, yield, or start-up costs compress margin?	Which cycle effects normalize and which advantages remain?	Can mix and cost gains persist after supply responds?
CapEx and working capital	How much cash is tied up during a downturn or ramp?	What recurring technology investment is required?	Can utilization grow faster than capacity spending?
Dilution and balance sheet	Does SBC, debt, or a commitment become more burdensome?	What is fully diluted FCF per share?	Does incremental ROIC justify reinvestment?

Use the DCF valuation guide for explicit cash-flow valuation. Use the Reverse DCF Calculator to infer the growth and margin path embedded in the market value. A reverse DCF is a expectations test, not proof that the market price is correct or incorrect.

Quarterly review checklist

Record the reporting period, currency, accounting standard, and whether figures are audited, reviewed, or preliminary.
Reconcile revenue growth to volume, price, mix, acquisitions, and currency.
Bridge gross margin with disclosed operational drivers.
Calculate reported cash FCF from the exact cash-flow lines and preserve the source units.
List working-capital changes and separate recurring from timing effects.
Record gross CapEx, subsidies, depreciation, commitments, and capacity plans separately.
Update normalized margin and CapEx ranges only when new evidence changes them.
Review inventory, receivables, payables, prepayments, lead times, and cancellations.
Track R&D, SBC, diluted shares, debt, net cash, and acquisition obligations.
Compare actual results with the prior downside, base, and upside cases.
Write down disconfirming evidence and unresolved data gaps.
Link every material reported figure to an official filing or issuer source.

Common mistakes and limitations

Annualizing a peak quarter or a working-capital release.
Comparing USD, KRW, and NT$ figures without preserving currency and period.
Ranking companies with different accounting standards and FCF definitions.
Calling all CapEx “growth” because revenue is increasing.
Treating depreciation as a precise maintenance-CapEx figure.
Netting subsidies against CapEx without showing the gross basis and timing caveat.
Ignoring R&D, SBC, dilution, inventory, prepayments, and supply commitments at fabless companies.
Subtracting future purchase commitments mechanically from current FCF.
Treating company-defined non-GAAP FCF as standardized.
Comparing a mixed semiconductor-software margin with a pure-play semiconductor margin.
Using management guidance as if it were a reported result.
Confusing an attractive business with an attractive share price.

No framework removes uncertainty. Segment disclosure may be incomplete, maintenance spending is rarely observable, and cycle turning points are difficult to identify in real time. The answer should therefore be a range supported by evidence, not a single precise normalized margin.

Frequently asked questions

How many years should I use to analyze a semiconductor stock?

Use at least five years of quarterly data when available, but the economic goal is to cover a meaningful expansion and contraction. A longer history may be necessary for memory and foundry businesses, while a newly scaled AI product may require scenario analysis because a complete comparable cycle does not yet exist.

Is EBITDA useful for semiconductor companies?

It can help compare operating trends, but it can be misleading when depreciation reflects economically necessary manufacturing investment. EBITDA should not replace a review of CapEx, technology transitions, working capital, and returns on invested capital.

Should R&D be treated as CapEx for a fabless company?

R&D is expensed under U.S. GAAP, but analysts may capitalize part of it for economic ROIC analysis. If you do, state the amortization life and apply the treatment consistently. Do not add capitalized R&D to cash CapEx when calculating reported cash FCF; that would mix accounting analysis with cash-flow reporting.

How should stock-based compensation be treated?

Keep operating cash flow as reported, then examine SBC as a real claim through dilution and repurchases. Presenting both cash FCF and fully diluted FCF per share avoids pretending SBC is either pure cash expense or economically free.

What is the best cycle indicator?

There is no universal indicator. Memory analysis emphasizes ASP, inventory, bits, utilization, and supply discipline. Foundry analysis emphasizes utilization, node mix, yield, capacity, and customer ramps. Fabless analysis emphasizes product transitions, end demand, inventory, commitments, and customer concentration.

Can I compare semiconductor FCF margins directly?

Only after reconciling period, currency, business mix, cash-flow definition, working capital, subsidies, acquisitions, and capital commitments. Even then, FCF margin should be combined with growth durability, ROIC, balance-sheet risk, and valuation.

Primary sources and methodology

The worked examples use official filings and issuer materials available by September 18, 2026. Reported facts are preserved in their stated currency and period. Stock Metric Lab calculations are arithmetic derived from cited inputs; estimates and judgments are labeled separately.

Micron, FY2026 Q3 Form 10-Q and earnings release.
SK hynix, Q2 2026 business-results release.
TSMC, Q2 2026 management report, earnings release, and earnings transcript.
NVIDIA, FY2027 Q2 Form 10-Q.
AMD, FY2026 Q2 Form 10-Q.
Broadcom, FY2026 Q3 Form 10-Q and earnings release.

This article is for educational and informational purposes only. It is not investment, financial, tax, legal, or accounting advice, and it is not a recommendation to buy, sell, or hold any security. Semiconductor financial measures, normalized cash flows, valuation scenarios, and return estimates depend on definitions, source data, and assumptions that may be incomplete, delayed, revised, or incorrect. Verify all inputs against current primary sources and consider your objectives and risk tolerance before making an investment decision.