Advanced Micro Devices (AMD)

By both methods, the value today (DCF $287 · Multiples $304) is below the market price ($594).

Today — expensive, no margin of safety: at $594 trades ~95.4% above its value discounted to today (~$304); the expected return does not even reach the risk-free rate (4.5%).

At 5 years — Sobrevalorado: the expected total return is negative — the price already discounts a demanding scenario that, if not met, results in a loss.

The bridge: the return at 5 years falls below the risk-free rate (4.5%) — which is why there is not even a discount to today's value. To require a 15% annual return, it would need to be bought at ~$188.

Educational / informational. Does not constitute investment advice.

The inverse approach: instead of projecting growth to obtain the value, the market price ($594) is taken as given and it solves for what annual owner-earnings growth would need to hold for 5 years for the present value —at the method's rate (4.5%, no-growth terminal)— to equal that price. It is the disconfirmation test: the expectations the price already pays for, contrasted against the method's projection.

The price discounts growth (53.8%/year) above our base case (32.0%/year) → it is priced for a demanding scenario and leaves little cushion against a slowdown.

That growth implies ~$48.6 bn of owner earnings in year 5 (vs ~$22.6 bn of our base case). It recalculates if the DCF assumptions are edited.

Year-by-year projection of the selected scenario. From each year, two versions of the flow are derived: growth FCF (operating flow − total capex, the cash surplus) and maintenance FCF (the owner earnings: what the business yields if it only sustains its capacity). The flow is returned almost in full (dividend + buyback) or redeployed into the operation, so that EV stays roughly flat and multiples compress because the metric grows, not because of cash accumulation. The valuation is done on Utilidad operativa neta (EV/NOPAT). In edit mode, revenue, margins, capex, and exit multiples can be adjusted.

Rolling grid: year 0 is the real TTM (revenue US$41.3bn at 27 June 2026, XBRL). AMD competes with Intel in server CPUs (EPYC) and with Nvidia in AI GPUs (Instinct, OpenAI and Meta agreements for up to 12 GW). It is valued on economic EV/NOPAT: after-tax operating income adds back the amortization of Xilinx intangibles (~$2.3bn, non-cash) but expenses SBC ($1.6bn, a real cost — 20% of FCF, §3), unlike the non-GAAP figure the market uses (which adds it back). Economic NOPAT of ~$5.65bn (TTM) → ~$25.6bn by 2031 on an aggressive AI ramp (revenue ×3 to ~$126bn, economic margin from ~21% to ~24%). Exit multiple 24× (base; bear 18× / bull 28×): the top of the semis band (§4) — today's entry multiple is unsustainable as growth decelerates. At 24× after-tax operating income, the 5-year value falls below the price → even a spectacular ramp does not justify the price: a case of concatenated optimism (aggressive growth × premium multiple). Net cash of $9bn in the bridge; no dividend; a net diluter (SBC + warrant). The bull case models the warrant dilution (share path to 1,850M). Update of the base year to the TTM ended 27 June 2026 (revenue of US$41.3 billion). GAAP operating income for that window (US$6.49 billion, a 15.7% margin) is not directly comparable with the 17.7% economic margin of the previous model, which adds back the non-cash amortization of Xilinx intangibles (approximately US$2.3 billion a year). Recalculated on that same basis, the economic margin of the current rolling window is 21.2%: a sustained improvement and not a deterioration. The four quarters that make up the window show the same elevated level, between 19.8% and 22.9%, with no one-off charge to explain it, which rules out single-quarter noise. The margin trajectory is re-anchored to this new starting point while keeping each scenario's long-term target (24% in the base case, 20% in the bear case and 27% in the bull case) and the same proportion of the remaining gap closed each year as in the previous curve. In the bear case, this implies a gradual compression of the current margin toward that long-term target, rather than starting from the same level as the base case.

Today's elevated multiple is the price of growth: if the business grows, the entry point cheapens on its own going forward (the metric grows while EV stays roughly flat). The exit multiple at 3 years is higher than the terminal at 5 years —at 3 years there is more growth still ahead—, so the value curve shows whether value creation is concentrated in the early or the later years. The required return is applied to the base scenario.

They are valued separately, with their own rationale, and are not incorporated into the base or the verdict (they are excess return). When assigning them value — in Editmode —, the total with optionalities updates live, without moving the base.

The verdict, the base CAGR, and the margin of safety are always calculated on the base; optionalities do not alter them (with optionalities at $0 they do not move).

Each card fixes a required annual return and answers: if the business is worth $379 in 5 years, what is the maximum that can be paid today to obtain that return? Since it now trades at $594, the margin of safety is how much cheaper the market is than that maximum. The three thresholds: 4% covers inflation (the floor), 10% is the long-term average return, and 15% is the level of a great investment.

With a target price of $379 in 5 years and a required return of 4.5% annually, the maximum to pay today is $304. Against the current market price ($594), the margin of safety is -95.4% (trades above the maximum → a premium is paid) and the total return at that price would be -8.6% annually.

• Economic EV/NOPAT (SBC expensed). It is valued on economic EV/NOPAT (adds back Xilinx amortization, expenses SBC). The non-GAAP figure the market uses adds SBC back → misleads the multiple lower.
• Extreme valuation. At ~131× net operating profit — the AI story fully priced in and then some.
• Concatenated optimism. The price discounts aggressive growth (revenue ×3) AND a premium multiple — two stacked forms of optimism. Even the favorable scenario barely justifies the price; the growth is overpaid for.
• The DCF does not close the price. The DCF (second lens, on economic earnings) gives a value well below the price: at ~131× net operating profit, no reasonable projection justifies the quote.
• No margin of safety. At ~$594 the base-case return (~-9%) is negative: the growth, however real, is already paid for. The warrant (~20% dilution) subtracts further. It is selling the euphoria, not buying value.
• No dividend. AMD does not pay a dividend (a growth business) → the return depends 100% on appreciation, which at this valuation is negative in the base case.

Each margin over sales, year by year: historical (solid line) → projection (dotted).

• ✓ ROIC exceeds the cost of capital (~10%)
• ✓ High incremental ROIC (compounding)
• ✓ CFROIC backs up the ROIC (95%, cash vs. accruals)
• ✓ Healthy balance sheet (low corporate debt)
• ✓ Durable competitive moat (multiple advantages)

• Strong, predictable cash flow. FCF ~$8.4bn TTM, net cash $9bn — a solid balance sheet. But reported FCF adds back SBC; owner-FCF net of SBC (~$6.8bn) is the real metric (§3).
• Return on capital. High ROIC on tangible capital (chip design is capital-light), but diluted by Xilinx goodwill ($25bn). Economic profitability rises with AI scale.
• The SBC trap. SBC ($1.6bn, 20% of FCF — material against the flow, §3) is a real cost that non-GAAP adds back. Valued on EV/economic NOPAT (SBC expensed); the multiple on non-GAAP earnings is misleading (on economic earnings it is ~131×).
• GAAP distorted by Xilinx. Xilinx's intangible amortization (~$2.3bn/year, non-cash) depresses GAAP (EPS $2.65 vs. non-GAAP $4.17); it is legitimately added back (it does not consume cash), but the capital of the acquisition still has to be charged.
• Creates economic value. Yes — the return on capital exceeds the cost; the business creates value. But at this valuation, the value is captured by the seller, not the buyer: the price already discounts the future value.
• Customer concentration. Low today (no customer ≥10%), but the OpenAI/Meta deals concentrate the AI future in a handful of hyperscalers — a growing dependence on those deals materializing.

Annual levels from the official filing (10-K); the % over each bar is the year-over-year (YoY) growth vs the prior year.

AMD earns by designing chips (manufactured by TSMC): CPUs (EPYC server, Ryzen consumer), AI GPUs (Instinct), and FPGAs (Xilinx). The engine is Data Center (server CPUs + the AI ramp); Instinct's $ figure is not reported discretely (embedded). GAAP earnings are distorted by Xilinx amortization (~$2.3bn, non-cash); non-GAAP adds it back but also adds back SBC ($1.6bn, a real cost, §3) → it is valued on economic EV/NOPAT (SBC expensed). It is a net diluter (SBC + OpenAI/Meta warrants), not a repurchaser.

The % are the annual (year-over-year) growth: each year —historical and projected— vs the prior one; the TTM (trailing 12m) vs the TTM of a year ago, to avoid overlapping windows. The historicals are exact figures from the official filings; the projected years come from the year-by-year model (the intermediate years without their own series are interpolated between the anchors). The projected columns (+1y…+5y) are 12-month windows counted from the TTM close (27-jun-2026): the projection starts from the most recently reported data, not the fiscal year. The projected base is realistic and unbiased — the risk discount is applied at the end, via the required return. The rationale for each metric is in the (i).

• Real, not nominal, growth. Revenue +34% is real (Data Center volume + server CPUs + the AI ramp); not inflationary. Growth is large and real — the debate is the valuation, not the growth.
• Where the engine comes from. Data Center (+32% in 2025, +107% in Q2'26): EPYC server CPUs (structural) + the Instinct AI GPU ramp (explosive, OpenAI/Meta 12 GW). Client (Ryzen) grows with PC share. The engine is server + AI.
• Dependence on AI execution. The growth that justifies the valuation depends on executing the AI roadmap (MI400/Helios) and on ROCm closing the gap with CUDA — flawless execution that the price already discounts.
• Realistic growth. The base case projects revenue tripling to ~$126bn by 2031 (an aggressive AI ramp) — even so, net operating profit at a disciplined multiple does not justify the price. The growth is real but overpaid.
• Reinvestment runway. Wide: research (~$8.8bn) feeds the CPU/GPU/FPGA roadmap; the AI market grows into the hundreds of billions. The runway is real; the return depends on capturing profitable share.
• Dilution (SBC + warrants). AMD is a net diluter (SBC $1.6bn > buybacks $1.3bn), and the OpenAI/Meta warrants (320M shares, ~20%) dilute if they vest — a headwind to per-share value that growth has to overcome.

• Relevant in 10 years / recurrence. Yes — CPUs, GPUs, and FPGAs are compute infrastructure with structural, growing demand (AI, data center). The risk is not relevance but competitive position (challenger to NVDA/Intel) and the cycle.
• Differentiated product. Yes in CPUs (chiplets) and embedded (Xilinx); in AI GPUs it is competitive on hardware but is chasing NVDA's software (CUDA vs. ROCm). Strong differentiation in CPUs, a challenger's position in AI.
• Share / position. A competitor to Intel in server CPUs and to Nvidia —the discrete GPU share leader per the 10-K— in AI GPUs, with the OpenAI and Meta deals. A rising challenger position, not a leader's.
• Pricing power. Limited in AI (it competes with Nvidia); stronger in server CPUs (high-value EPYC). Pricing power in AI is that of the challenger, not the leader.
• Operating leverage. High: the gross margin rises (50%→53%) with the AI/server mix, and the operating margin leverages with scale. Research (~$8.8bn) is the fixed cost being leveraged.
• Behavior in a recession / cycle. Cyclical: semiconductors have sharp cycle downturns (AMD had losses in prior cycles); an oversupply of AI capacity or a cooling in hyperscaler capex would hit hard. Cyclicality is the vulnerability at this valuation.

Not all debt is equal: only the structural needs refinancing; the rest is operational (self-liquidating).

Structural debt is what is exposed to the contraction phase of the cycle; operational debt (leases, matched funding) self-liquidates with the business.

• ✓Leverage (net debt / EBITDA)Net cash $9.9 bn
• ✓Interest coverage (EBIT / interest)44.3x
• –Liquidity (current ratio)no data
• ✓Cash quality (CFROIC vs ROIC)CFROIC backs 95% of ROIC
• !Value creation (ROIC − 10% bar)+4pp
• ✓Malinvestment test (capex vs incremental ROIC)Productive investment (1.4x)
• !Float / working capitalConsumes cash $0.5 bn (positive WC)
• !Dilution (SBC % of revenue + shares)SBC 4.6% of revenue

A traffic-light interpreted by the method (not generic): float (negative WC) adds up, capex is judged by incremental ROIC (malinvestment test), and a lender is not subjected to corporate solvency. The (i) shows the derivation of each number.

• Net cash. Net cash $9bn (growing), minimal debt ($3.2bn) — solid balance sheet, no solvency risk. A capital-light chip designer with ample headroom.
• Cash generation. FCF ~$8.4bn TTM, but it adds back SBC; owner-FCF net of SBC (~$6.8bn) is the shareholder's real cash. Strong, but not as strong as reported FCF suggests.
• Net diluter. AMD is a net diluter (SBC $1.6bn > buybacks $1.3bn; diluted shares grew in each of the last 3 years) — per-share value faces the headwind of dilution, not the tailwind of buybacks.
• Warrant overhang. The OpenAI/Meta warrants (320M shares, ~20% of shares outstanding, at $0.01) are a large dilution overhang if the deals materialize and vest — to be modeled once the conditions are met.
• Semiconductor cyclicality. The risk is the cycle (AMD had losses in prior cycle downturns), not solvency; the net-cash balance sheet protects it from a crisis, but revenue and margin are cyclical.

Who owns the shares — the alignment and whether there is a controlling shareholder.

• Track record (Lisa Su). Solid: has led AMD since 2014 and brought it to compete with Intel and Nvidia — the credibility anchor of the bull case.
• Capital allocation. Heavy reinvestment (~$8.8bn of research, appropriate for a chip designer) + modest buybacks (do not offset SBC). The OpenAI deal (warrant) is the bet; no dividend (growth business).
• Alignment / skin in the game. Professional management (insiders <1%); Su holds ~$2.2bn in stock (material in dollar terms). Compensation is tied to relative return versus the S&P 500 — performance-based alignment.
• The warrant structure. The OpenAI/Meta warrants (320M shares at $0.01) align anchor customers with the stock in exchange for a 12 GW commitment — creative, but a ~20% dilution overhang if they vest; one to watch.

• It is not cheap — it's the opposite: AMD trades at ~131× net operating profit. The AI story (the OpenAI/Meta deals, the Instinct ramp) is fully priced in and then some.
• Non-GAAP understates the multiple: non-GAAP earnings add SBC ($1.6bn) back, which is a real cost. On economic earnings (SBC expensed) the multiple is ~131×.
• This is a case of euphoria, not a perception-reality discrepancy: the market correctly sees a great business winning in AI, but pays a price that discounts flawless execution — with no missing buyers or motivated sellers, there is no bargain, only a good business that is expensive.

AMD is not cheap: it's a great business (server CPUs and AI GPUs) at a euphoria price (~131× net operating profit). The perception-reality gap that creates a value opportunity is entirely absent — the market perceives the quality and pays handsomely for it. It's the mirror image of a bargain: recognized quality, overpaid.

The annual return (CAGR at 5 years) in each scenario, with the total period return below — the margin of safety made visual: upside range wide, downside range narrow.

Even in the bear scenario, the return holds at -25%/year (-76% total): the margin of safety protects the downside. The bull (+1%/year, +6% total) exceeds it comfortably — a favorable asymmetry, with a narrow downside range and a wide upside range.

• NVDA's software moat (CUDA). If ROCm fails to close the gap with CUDA, AMD remains the second-choice AI GPU supplier at lower prices/margins — the central disconfirmation test of the AI case that justifies the valuation.
• Extreme valuation (~131×). At ~131× net operating profit , any execution stumble or slowdown in the AI ramp is punished hard.
• Semiconductor cyclicality. An oversupply of AI capacity or a cooling in hyperscaler capex would hit revenue and margin right when the price discounts perpetual growth.
• Export controls to China. China (~22% of 2025 revenue by billing location) is a regulatory overhang: the US government expects to receive 15% of licensed MI308 sales to China (no regulation published), and a tightening would close off the market.
• Dilution (SBC + warrants). SBC ($1.6bn) dilutes continuously, and the OpenAI/Meta warrants (320M shares, ~20%) would dilute further if they vest — a headwind to per-share value.

Each thinker's analytical framework applied to our data.

• Buffett / Graham Quality + margin of safety

  A narrow moat and ROIC 14% above the 10% bar, but the price sits above value (no margin) → great business, expensive.

• Peter Lynch Growth at a reasonable price (GARP)

  A hyper-growth growing 32% at a multiple/growth of 4.1 → expensive for its growth.

• Joel Greenblatt Cheap and high-return (Magic Formula)

  Earnings yield 1% (EBIT/EV) + ROIC 14% → falls outside the Magic Formula.

• Howard Marks Perception vs reality + cycle

  The price discounts 54% vs our 32%: priced for perfection, no favorable mispricing.

• Seth Klarman Capital protection (bear scenario)

  Bear-scenario floor -25%/yr over 5y (material loss) → risk of permanent capital loss.

• Pat Dorsey Moat strength (Five Rules)

  A narrow moat, stable; sources: intangibles, switching costs, efficient scale, cost advantage → partially passes the Five Rules.

• Aswath Damodaran Expectations implied by the price

  The price demands 54% growth, far above our 32% — heroic assumptions.