Advanced Micro Devices, Inc. (NASDAQ: AMD) — Right Horse, Wrong Price: the Second Source Priced Like the Leader

An independent equity-research note Report date: 2026-06-09 Price at analysis: ~$475–490 | Market cap: ~$775–808B | Enterprise value: ~$800B Shares outstanding: ~1.63B (plus up to 160M OpenAI warrant shares ≈ +9.8% fully-diluted) Fiscal year: ends late December (FY2025 ended 2025-12-27) | CIK: 0000002488

Standing disclaimer: The analytical sections below carry no investment recommendation and no price target. They discuss valuation only as embedded expectations and scenarios. The single deliberate exception is the Claude’s Take block immediately below, which is explicitly labeled and subjective.

⚡ Claude’s Take

This block is the author’s own independent, subjective opinion and general information only — not investment advice. The analytical sections that follow take no position and carry no price target.

Verdict: HOLD / AVOID-here — a genuinely great operator priced as if it were already the winner. Accumulate only on weakness. Conviction: medium. Directional zone: I’d be a buyer in the ~$280–340 band (paying for the EPYC franchise + cheap optionality on Instinct, not full Analyst-Day delivery), neutral in the ~$430–490 fair-but-full zone, and a non-chaser above ~$500. Base-case present value clusters around $430–560, i.e. roughly where it trades — which is precisely the problem.

Tag: “Right horse, wrong price — the second source priced like the leader.”

AMD is the best turnaround in large-cap semis and a real franchise: EPYC has dragged Intel into a genuine duopoly and is gaining unit share on an expanding CPU TAM, and Instinct makes AMD the only credible merchant #2 in AI accelerators as the world begs for an alternative to NVIDIA. I don’t doubt the business. What I won’t do is pay for it here. At ~21x EV/sales — above NVIDIA’s price-to-sales while earning roughly a quarter of NVIDIA’s operating margin — and at the 99.7th percentile of AMD’s own 10-year price-to-sales history, the stock embeds substantial delivery of management’s >$20-EPS, >35%-margin 2028-ish model with no margin of safety. The load-bearing assumption is that Instinct eventually earns NVIDIA-like economics; management’s own commentary says the opposite today (MI450 gross margin is below corporate average, and the CPU/embedded mix is what defends the blended-margin target). So the engine driving the multiple is currently a margin drag. Add a textbook late-stage capital cycle (NVDA, AMD, AVGO and all three memory makers building into “insatiable” demand simultaneously), a ~10% warrant overhang handed to OpenAI, and ~0.41% insider ownership with zero open-market buys, and the risk/reward is left-skewed: the base case ≈ today’s price three years out, the bull needs AMD to beat an already-aggressive plan, and the bear only needs the AI cycle to behave like every prior cycle.

What flips me bullish: a printed quarter showing Instinct gross margin converging toward corporate average at scale — proof the AI franchise earns its multiple rather than promising to. What flips me bearish (faster): a hyperscaler capex-digestion quarter or an OpenAI/Meta anchor slip — at this multiple there is no floor to catch it. Great company; I want it ~30–40% cheaper.

1. Executive Summary

Advanced Micro Devices is a fabless designer of x86 server and client CPUs, data-center GPUs/AI accelerators, gaming and semi-custom silicon, and — via the 2022 Xilinx and Pensando acquisitions — FPGAs, adaptive SoCs, and DPUs. Over three years it has transformed from “the company finally beating Intel” into “the only credible #2 to NVIDIA in AI compute.” FY2025 revenue grew +34% to $34.64B, with Q1-FY2026 accelerating to +38% YoY ($10.25B). Data Center is now 48% of revenue ($16.6B), built on a profitable, share-gaining EPYC server-CPU franchise plus a Instinct GPU line that went from ~$0 to multi-billions inside two years.

The business is mixed-quality, not uniformly excellent. The genuine moat lives in one segment: EPYC, protected by the x86 duopoly (only AMD and Intel hold the cross-licenses), real switching costs, and scale R&D — and it is gaining unit share (FY2025 Client +51% on +31% units and +15% ASP) on a CPU TAM management says doubled to >$120B. Embedded (Xilinx/Versal) is a real switching-cost moat at 36% segment margin but cyclically depressed. Client/Gaming is a cost-advantaged but commodity, no-captivity business. And the segment driving the entire equity narrative — Instinct data-center GPU — has no durable moat: NVIDIA’s CUDA is a deep demand-side captivity moat (≈90%+ share, ~60% operating margins) that AMD’s open-source ROCm narrows but does not neutralize, and the merchant pie is itself being carved out by hyperscaler ASICs (Broadcom/Marvell co-design). The financial tell is direct: if AMD’s AI silicon had a moat, it would price like NVIDIA’s; it cannot, which is why Instinct is dilutive to gross margin today.

Financially, scale genuinely improves the economics — GAAP operating margin went 1.8% (FY2023) → 10.7% (FY2025) → 14.4% (Q1-26), FCF reached $5.5B on an asset-light model, and the balance sheet is net-cash (~$8B). But two honest qualifications govern: (1) non-GAAP flatters economics by adding back ~$1.64B/yr of recurring, dilutive SBC (the ~$2.25B/yr Xilinx intangible amortization add-back is more defensible); and (2) headline ROE ~8% / ROIC ~6% are depressed by $41.8B of Xilinx goodwill+intangibles (54% of the balance sheet) — strip that and operating ROIC is ~23%, but the headline reflects the price paid (in peak-2022 stock) for Xilinx. Capital allocation is competent but not exemplary: shrewd strategy (Xilinx’s franchise, ZT’s design team bought and ZT manufacturing sold to Sanmina) repeatedly funded with equity issued at peaks (Xilinx all-stock; the OpenAI penny-warrant for up to ~10% of the company), and a buyback that does not even offset SBC (shares rose in FY2025).

On valuation, AMD is the most expensive name in the AI-silicon complex relative to the profit it currently earns — ~21x EV/sales and ~99x EV/EBITDA, with a forward P/E spanning ~37x (FY2027 consensus EPS ~$13) to ~65x (FY2026 ~$7.4). A reverse-DCF shows the current price ≈ the present value of management delivering its own Analyst-Day plan near the early-to-middle of its window — meeting expectations earns a mediocre return; missing them is punished severely, with no valuation floor. This is not a crowded short (2.75% of float); the variant-perception debate is not direction but degree and price — specifically whether the AI franchise earns the NVIDIA-like profitability the price already assumes. The institutional verdict (no recommendation): a real, improving franchise whose durable economics sit almost entirely in EPYC, wrapped in a price that has already underwritten the unproven part.

2. Business Overview

What AMD does. Advanced Micro Devices is a fabless designer of high-performance and adaptive computing silicon: server and client x86 CPUs, data-center GPUs/AI accelerators, gaming GPUs, semi-custom console SoCs, and — via the 2022 Xilinx and Pensando acquisitions — FPGAs, adaptive SoCs, DPUs, and AI NICs. AMD designs the silicon and the surrounding software stack; it manufactures nothing. All leading-edge wafers are fabricated by TSMC, and assembly/test/packaging (ATMP) is outsourced to third parties (FY2025 10-K, Item 1A Risk Factors, filed 2026-02-04). The investment-relevant consequence is that AMD’s competitive position is a design and software position layered on top of a shared foundry: when AMD leads Intel, it is substantially because both AMD and Intel buy capacity from the same TSMC processes and AMD has executed the architecture and node transitions better — not because AMD owns a manufacturing asset Intel cannot rent.

Segment structure (changed in FY2025). Beginning Q1 2025 AMD collapsed to three reportable segments — Data Center; Client and Gaming (now combined, though Client and Gaming revenue is still disclosed separately); and Embedded — with prior periods retrospectively recast (10-K, Note 4 Segment Reporting). FY2025 results:

Total revenue grew +34% YoY (from $25,785M FY2024). Note the gap between segment operating income ($7.70B summed across the three segments) and consolidated GAAP operating income ($3.69B): the $4.0B “All Other” drag is overwhelmingly acquisition-intangible amortization ($2,254M — the Xilinx/Pensando purchase-accounting tail) and stock-based compensation ($1,638M) (10-K, Note 4). Any honest read of AMD’s economics must hold this in mind — the segments look like a high-margin franchise; the consolidated GAAP business converts only ~11% of revenue to operating profit, and reported ROE is ~8%.

Product families and how it makes money. Data Center is now the engine (48% of revenue, doubling off a $6.5B FY2023 base to $16.6B). It contains two distinct businesses: EPYC server CPUs (Genoa/Turin, with Venice on the roadmap) — a profitable, share-gaining, recurring franchise — and Instinct data-center GPUs (MI300X, MI325X, the MI350 series launched June 2025, and the forthcoming MI400/MI450 + “Helios” rack systems and MI500). The GPU line is the high-growth, low-visibility, fiercely contested part. Client is Ryzen desktop/notebook CPUs and APUs ($10.6B, +51% YoY — genuine share gains plus AI-PC mix). Gaming is Radeon discrete GPUs plus semi-custom console SoCs for Sony (PlayStation) and Microsoft (Xbox); it has shrunk from $6.2B (FY2023) to $3.9B as the console cycle ages. Embedded is the Xilinx FPGA/Versal adaptive SoC plus embedded CPU/APU and Pensando DPU portfolio — the smallest segment ($3.45B, still declining from a $5.3B FY2023 peak) but the highest-margin at 36% and the most moat-like (long design-in lifecycles).

Recurring vs. cyclical. AMD has no subscription revenue; “recurring” here means franchise persistence, not contracted ARR. EPYC and Ryzen are durable, multi-generation share-capture stories. Gaming console SoCs are explicitly cyclical (locked to 6–7-year console generations — note the 18%-of-revenue console customer in FY2023, now diluted below 10%). Embedded is late-cyclical (inventory-correcting since 2024). Instinct GPU is the opposite of recurring — it is lumpy, deal-driven hyperscaler capex riding the AI build-out, with revenue concentrated in a handful of large customers and large quarterly swings.

Customer and end-market concentration. Encouragingly, no single customer was ≥10% of consolidated revenue in FY2025 or FY2024 (one Client+Gaming console customer was 18% in FY2023) (10-K, Note 4). But the Data Center GPU business is concentrated in a small set of hyperscalers and AI labs — management cites 7 of the top 10 AI companies using Instinct and names OpenAI, Meta, Oracle, and the U.S. Department of Energy as MI400/MI450 anchors (Analyst Day, 2025-11-11). Geographically, FY2025 billings were 33% U.S. and 22% China (incl. Hong Kong) — material given export controls: the MI308 restriction triggered ~$800M of inventory and related charges in Q2 2025 before AMD secured some U.S. export licenses (10-K, Item 1A). China is thus both a demand source and a policy risk that has already cost real money.

Capital actions shaping the model. AMD acquired ZT Systems for $4.4B (closed 2025-03-31) to add hyperscale rack/systems design, then divested ZT’s manufacturing to Sanmina — keeping the design talent, shedding the low-margin contract-manufacturing assets (10-K, Note 5; Analyst Day). This signals AMD wants to sell rack-scale “AI factory” solutions (Helios) without becoming a systems integrator.

Verdict on the business. A genuinely improved, more diversified franchise than three years ago, with a real CPU franchise (Data Center + Client) now producing 80% of revenue and a credible #2 position in the AI accelerator gold rush. But it is a mixed-quality portfolio: a strong, scale-advantaged CPU business; a structurally moat-light commodity Client/Gaming business; a high-margin but cyclically-shrinking Embedded business; and a high-growth, low-moat, hyperscaler-concentrated GPU business whose economics are unproven at scale. The ~11% GAAP operating margin and ~8% ROE — depressed by the Xilinx amortization tail — say the franchise quality has not yet shown up in consolidated returns.

3. Industry Dynamics

The semiconductor industry and where the profit pool now sits. The logic-semiconductor value chain is a stack of tollbooths, and AMD occupies the most contested rung of it. At the top, EDA/IP licensors (Synopsys, Cadence, Arm) and, uniquely, the two manufacturing monopolists — ASML (sole EUV supplier) and TSMC (the dominant leading-edge foundry) — capture the most durable economics. Fabless designers like AMD sit in the middle: they own the architecture and the brand but rent both the lithography and the fabrication. The global semiconductor market reached ~$772B in 2025 (+22%) and WSTS forecasts ~$975B in 2026 (+25%), with Logic guided +37% — almost entirely AI-driven (industry data (WSTS)). The single profit pool that matters for AMD’s equity story is the data-center AI accelerator market, and the uncomfortable structural fact is that AMD is the clear number two in it. Merchant data-center GPU is ~90%+ NVIDIA by units (industry estimates, 2026-06-09); AMD’s Instinct line (MI300/MI325/MI350, now MI450/Helios) is the only credible merchant alternative, but it competes for the residual. Worse, a third pool — custom ASICs (Google TPU, Amazon Trainium, Meta MTIA, Microsoft Maia), enabled by Broadcom and Marvell — is growing fast and is captured by neither NVIDIA nor AMD; AMD management itself concedes ASICs are ~20–25% of the accelerator TAM, GPUs the rest (FACT — Investor Day, 2025-11-11).

Pressure-testing AMD’s TAM. At its November 2025 Analyst Day, AMD raised its framing from a prior “>$500B data-center accelerator TAM by 2028” to a “>$1 trillion silicon TAM by 2030,” explicitly broadened to fold in data-center CPUs and scale-up/scale-out networking alongside accelerators, and explicitly not a “solution” TAM (FACT — Investor Day). Against that, management targets >35% revenue CAGR, >35% operating margin, >$20 EPS, an >80% AI-revenue CAGR over three-to-five years, and “double-digit” data-center-AI share (FACT — Investor Day). INTERPRETATION: this number is internally coherent but self-serving and should be treated as a hypothesis, not evidence. It is far broader than the merchant-GPU pool AMD actually wins in — it counts CPUs, networking, and the full accelerator TAM including the hyperscaler ASIC share AMD does not capture. A mid-teens share of $1T implies ~$120–150B of AMD revenue, versus $16.6B of data-center revenue today — a ~8–9× leap that rests on the AI capital cycle persisting and on AMD closing the software/systems gap to NVIDIA. The peer reports show every incumbent reaching for the same demand base: NVIDIA cites a $3–4T AI-infrastructure TAM by decade-end (flagged ASSUMPTION in peer analysis), TSMC guides ~25% revenue CAGR 2024–29 — all three sit on a single foundation, ~$1T of hyperscaler capex in calendar 2026 (industry estimates). When the entire complex underwrites its forecasts off one customer cohort’s discretionary spend, the TAM numbers are correlated, not independent.

The x86 duopoly and ARM’s encroachment. AMD’s good business — and the source of its credibility — is the x86 CPU duopoly with Intel. AMD has taken server share for five consecutive EPYC generations (Genoa → Turin → Venice), and management claims ~40% revenue share in the most attractive hyperscale segments (FACT — Investor Day); FY2025 Client revenue rose 51% to $10.6B on a 31% unit increase and a 15% ASP increase (FACT — 10-K MD&A), evidence of genuine pricing power against a weakened Intel. AMD frames the server CPU TAM growing high-teens and adding ~$30B (doubling) by 2030 (FACT — Investor Day). The structural threat here is not Intel but ARM: AWS Graviton, NVIDIA Grace, Ampere, and Arm-based PC silicon all chip at the x86 unit TAM. AMD’s own 10-K flags that competitors “pursuing alternative computing architectures, such as Arm… could grow the Arm ecosystem and increase competition in consumer, commercial and data center, reducing demand for our products” (FACT — 10-K, Item 1A). The x86 duopoly is real and presently lucrative, but it is a shrinking-share-of-a-contested-pie business over a long horizon.

Foundry and HBM dependence — the binding constraints. AMD is fabless and rents the most strategically scarce capacity on earth. It relies on TSMC for all wafers at 7nm and below (every microprocessor and GPU), GlobalFoundries above 7nm, and UMC/Samsung for some ICs (FACT — 10-K). MI455 is built on TSMC 2nm + 3nm with 3D chiplet packaging and HBM4 (FACT — CES 2026, 2026-01-05). This concentrates three exposures: (1) cost/supply — TSMC holds >90% of the leading edge with no price discounting (N2 wafers ~$30,000), so AMD has limited leverage over its largest input cost; (2) packaging — TSMC’s CoWoS advanced-packaging is the chokehold on every AI accelerator, AMD’s included; and (3) geopolitics — leading-edge capacity sits overwhelmingly on Taiwan, within missile range of the PRC; the 10-K explicitly warns that “geopolitical changes between China and Taiwan could disrupt the operations of our Taiwan-based third-party wafer foundries” and notes the realized April 2024 Taiwan earthquake (FACT — 10-K). Layered on top is the HBM constraint: high-bandwidth memory is the binding limiter on AI-GPU output industry-wide, supplied by a tight oligopoly — SK Hynix ~62%, Micron ~21%, Samsung the balance (industry HBM market data). HBM scarcity protects pricing today but is being competed away as all three suppliers add capacity into the peak (Micron capex >$25B in FY26 vs $8.4B in FY24).

Regulation — a real, quantified revenue headwind. US export controls are not a theoretical risk for AMD; they have already cost cash. China (incl. Hong Kong) was $7,751M of FY2025 revenue — 22.4% of the total and AMD’s second-largest market (FACT — 10-K Note 4). The April 2025 BIS license requirement on advanced semiconductors to D5 countries directly hit the Instinct MI308: AMD recorded ~$800M of inventory and related charges in Q2 FY2025, ~$440M net of which compressed full-year data-center gross margin (FACT — 10-K MD&A). AMD was granted some licenses, and in August 2025 US officials signalled an expectation that the government would receive 15% of MI308-China revenue (FACT — 10-K). The 10-K is blunt that controls “make it easier for our China-based competitors to develop and sell their own solutions” and may “encourage customers in China… to pursue alternatives to U.S. semiconductors” (FACT — 10-K). CHIPS Act incentives marginally offset the geographic risk by subsidizing US fab buildout, but AMD, as a fabless designer, captures little of that directly. INTERPRETATION: China is a structural overhang — partly an AI-accelerator licensing lottery, partly ordinary client/embedded/console demand — and the precise AI-China split is an open question for Financials.

Cyclicality and the capital cycle (Marathon lens). AMD’s segments straddle two different cyclical phases. Gaming (semi-custom console SoCs, late in the PlayStation/Xbox cycle), Client (PC replacement cycle), and Embedded (FPGA/industrial/auto — revenue fell 3% in FY2025 on “mixed” end-market demand; FACT — 10-K) are the ordinary, mature cyclicals. Data Center is in an investment super-cycle — and that is the heart of the risk. Applying Marathon’s capital-returns framework: the highest returns in technology (NVIDIA at ~75% gross / ~60% operating margin) are drawing in the most capital in technology, from every direction at once — AMD itself, every hyperscaler’s in-house ASIC program, Broadcom/Marvell, Chinese domestic suppliers, and the capital markets financing the buildout. Synchronized, pro-cyclical capacity additions (the three HBM makers all expanding into the peak; TSMC’s CoWoS ramp) are the textbook signal that the next down-leg is being financed now. The asset-growth anomaly warns that this much capital flooding in at peak returns reliably precedes mean reversion. The single most important industry risk is therefore not a demand collapse but a capital-cycle digestion — a pause or air-pocket in ~$1T of hyperscaler capex meeting newly-arrived competitive and in-house capacity. AMD, as the newer, lower-share data-center ramp dependent on a handful of hyperscaler, sovereign, and OpenAI-type anchor customers, would likely see allocation and orders cut faster than the leader’s in such a digestion.

Switching costs, barriers, and scale. At the industry level the barriers are formidable but asymmetric. The manufacturing barrier (leading-edge fabs, EUV, CoWoS) protects TSMC and ASML, not AMD — AMD pays the toll. AMD’s own barriers are narrower: x86 instruction-set compatibility and a 50-year IP estate (a genuine duopoly moat vs Intel), the cost and yield risk of a multi-quarter re-tape-out (which deters customers from casually switching CPU vendors), and, in AI, an emerging full-stack systems capability (Helios rack-scale, ROCm software). But in AI the decisive barrier is software and scale-up networking, where NVIDIA’s CUDA captivity and NVLink fabric remain ahead — AMD’s UALink/Ultra-Ethernet answer is credible but unproven at frontier scale. Greenwald’s share-stability test gives a split verdict: AMD is gaining share in server CPU (consistent with a real, widening position vs Intel) but is the challenger against entrenched share in AI GPU.

Verdict — a structurally excellent demand environment, of uncertain durability, in the late boom of a capital cycle. The semiconductor industry AMD sells into is, right now, one of the most attractive in the global economy: secular AI demand, near-$1T market approaching, supply chokepoints (TSMC leading edge, CoWoS, HBM) that protect pricing for everyone holding allocation. AMD’s specific position is improving — a genuine x86 duopoly throwing off rising cash, plus a credible, share-gaining #2 franchise in the single largest profit pool in technology. But the industry is squarely in the boom phase of a textbook capital cycle, and AMD is the more cyclically-exposed, lower-share participant in the most contested pool. The $1T-by-2030 TAM is real demand layered with promotional extrapolation; it rests on hyperscaler capex — the most discretionary spend in technology — at what may be a cyclical peak, and it counts a large ASIC slice AMD does not capture. The honest synthesis: a good industry to be levered to on a one-to-two-year view, of genuinely uncertain quality on a five-to-ten-year view, where the dominant risk is not that AI fails but that the enormous competing capital now flooding the accelerator profit pool — from competitors, customers, and states — does exactly what the capital cycle says it does, and compresses the returns that today’s price embeds.

4. Competitive Position

AMD is best understood as four businesses with four different moats (and two with essentially none). Applying Greenwald’s taxonomy (supply/cost advantage; demand/customer captivity; economies of scale + captivity) segment-by-segment is the only honest way to assess durability — a blended “AMD has a moat” claim is wrong.

Server CPU (EPYC) — the real moat: economies of scale + customer captivity, inside an x86 duopoly. This is AMD’s genuine competitive advantage and the one place the share data passes Greenwald’s tests in AMD’s favor. The x86 instruction set is a two-player barrier: only AMD and Intel hold the cross-licenses to build x86 server CPUs, and the installed base of x86-compiled, validated, certified enterprise and cloud software creates demand-side captivity — migrating a hyperscale or enterprise fleet to a different ISA (Arm) requires recompilation, re-validation, and operational re-qualification. Management states EPYC has reached “over 40%” server CPU revenue share and is “the de facto standard in the cloud,” with a stated path to >50% (Analyst Day, 2025-11-11; Q1 2026 call cites record enterprise sell-through, 2026-05-05). Treat the share figure as a hypothesis to validate — it is management-stated revenue share, which overstates unit penetration if EPYC ASPs run above Intel Xeon; it must be reconciled to Mercury Research unit data (OPEN QUESTION).

Run the Greenwald share-stability test: AMD’s server share rose from mid-single-digits (2018) to ~40%+ (2025). A >5-point swing over the window technically signals an absence of barriers protecting the incumbent — which is exactly right: AMD has been breaching Intel’s eroded position via TSMC’s process lead and Zen execution, not defending a settled fortress. The forward question is whether AMD can now defend 40%+ the way Intel failed to. The advantage is real but rented from TSMC: if Intel’s foundry (18A/14A) closes the process gap, AMD’s cost/performance edge compresses. Verdict: a durable-enough moat (ISA duopoly + switching costs + scale R&D), but it is execution- and foundry-dependent, not structural in the way a network effect would be. This is AMD’s best business and the clearest tie between moat and financial outcome — EPYC carries Data Center’s 21.7% segment margin.

Data-center GPU (Instinct) — no durable moat; the CUDA gap is the defining weakness. Be direct: NVIDIA’s CUDA is a deep demand-side captivity moat, and AMD does not have an equivalent. Fifteen-plus years of CUDA libraries, framework optimization, and a developer base that codes to CUDA by default create switching costs that AMD’s ROCm — open-source, improving, “battle-tested” per management (Analyst Day) — narrows but does not neutralize. ROCm is a fast-follower, not a lock-in: it lowers the cost of trying AMD, but it does not make leaving AMD painful, which is what an actual moat requires. The share data confirms the gap. Industry market-share estimates puts NVIDIA at ~90%+ of merchant data-center GPU/accelerator units, stable-to-rising through Blackwell despite years of funded AMD competition — by Greenwald’s stability test, a formidable NVIDIA barrier. AMD is the credible #2, winning price-sensitive and second-source allocation (the largest MI450 deployments management cites are for inference, the more price-elastic, less CUDA-locked workload — Q1 2026 call). AMD also lacks an NVLink-class scale-up fabric (it is building open Ultra-Accelerator-Link/Helios in response). Management’s >80% AI-revenue CAGR over 3–5 years and “double-digit” DC-AI share target (Analyst Day) is a forward bet on MI450/Helios ramping in 2026–27 — unproven at scale and structurally capped by the CUDA moat it does not own. Verdict: no durable competitive advantage in merchant AI GPUs today; AMD competes on price, openness, and TCO as the second source. The financial tell: if AMD had a moat here, it would price like NVIDIA (~75% GM); it cannot, which is why Instinct dilutes, not lifts, data-center gross margin.

Custom-silicon / ASIC threat — squeezes both AMD and NVIDIA. The merchant-GPU market AMD is fighting for is itself being undercut by hyperscaler in-house silicon — Google TPU, Amazon Trainium/Inferentia, Meta MTIA, Microsoft Maia — designed with Broadcom and Marvell as merchant-ASIC partners (industry analysis of custom-silicon vendors; NVDA report). The 10-K explicitly flags that “some of our customers are internally developing their own data center microprocessor and accelerator products which could impact the available market” (10-K, Item 1, Competition). Custom ASICs were ~28% of AI-chip units in 2026 and growing faster than merchant GPUs (industry estimates). For AMD this is a double bind: it is the challenger trying to take merchant share while the merchant pie itself is being carved out by the very hyperscalers that are its target customers. AMD has no captive-ASIC franchise to offset this (the ZT/Sanmina move was about systems, not custom silicon).

Client / Gaming — cost/process advantage, but commodity end-markets with no captivity. Ryzen’s strength is a TSMC node lead over Intel plus chiplet cost structure — a real supply/cost advantage, but in a market with no demand-side captivity: PC buyers and OEMs switch on price/performance every cycle, and NVIDIA is the discrete-GPU share leader (10-K, Item 1) while Arm-based PC entrants (Qualcomm, Apple-driven expectations) pressure x86 itself. Client grew 51% in FY2025 on share gains and AI-PC mix, but this is share taken on merit each generation, not share protected by a moat. Gaming is worse: console SoCs are a low-margin, design-win-locked but cyclically-declining business (Gaming revenue down 37% since FY2023), and Radeon is a distant #2 to NVIDIA. Verdict: weak/no moat. The financial tell is the 19.6% combined segment margin and the cyclicality — a moated franchise would not see Gaming revenue halve over a console cycle.

Embedded (Xilinx FPGA / Versal) — a genuine switching-cost moat, cyclically masked. FPGAs and adaptive SoCs are designed into customer hardware over multi-year cycles with 5–10-year product lifecycles, creating real switching costs and the highest segment margin (36%). Competition is fragmented (Altera/Intel, Lattice, Microsemi) and AMD targets >70% adaptive-revenue share (Analyst Day). The catch: revenue has fallen from $5.3B (FY2023) to $3.45B (FY2025) on an inventory correction — the moat is real but the demand is cyclical, so the high margin sits on a shrinking base. Verdict: the most moat-like business after EPYC, but small and cyclically depressed.

Scale economics and the ROIC test. AMD’s R&D (~$8B) is now competitive in absolute terms but is spread across CPU, GPU, FPGA, DPU, and software — versus NVIDIA’s more concentrated GPU/CUDA spend and Intel’s foundry-burdened budget. On Greenwald’s ROIC test, the verdict is sobering: FY2025 ROE ~8% — mediocre for a purported moat business — depressed by the $2.25B/yr Xilinx-intangible amortization and SBC in All Other. Normalized/cash returns are better, but reported consolidated returns do not yet evidence a wide moat. The wide-moat economics live almost entirely in one segment (EPYC server CPU); everywhere else AMD is either moat-light (Client/Gaming), cyclically masked (Embedded), or moatless and price-taking (Instinct GPU). The bull thesis — management’s >35% revenue CAGR, 55–58% gross margin, >35% operating margin, >$20 EPS (Analyst Day) — requires the moatless GPU business to scale into a high-margin franchise it does not yet structurally deserve. That is the central competitive question for the valuation, and on current evidence it is unproven.

5. Growth History and Forward Opportunities

AMD’s revenue history splits cleanly into three eras, and conflating them is the most common analytical error made on this stock. (FACT) Revenue ran $16,434M (FY2021) → $23,601M (FY2022) → $22,680M (FY2023) → $25,785M (FY2024) → $34,639M (FY2025), with Q1-FY2026 at $10,253M, up 38% year-over-year from $7,438M. That looks like a 5-year ~16% CAGR, but the path was anything but linear, and the headline FY2022 jump is largely an accounting illusion.

Organic vs. acquired. The +44% FY2022 step-up was overwhelmingly inorganic: Xilinx closed in February 2022, folding ~$3.5B of Embedded and FPGA revenue into AMD’s base for ~10.5 months. (INTERPRETATION) Strip Xilinx and underlying FY2022 organic growth was high-single/low-double-digit, not 44%. The proof arrives in FY2023, when revenue fell 3.9% to $22,680M — a genuine PC and gaming downcycle, with Client revenue collapsing as the channel destocked. So the FY2021→FY2023 window shows essentially flat organic revenue dressed up by an acquisition. The ZT Systems deal (closed March 31, 2025, $4.4B) is a smaller distortion: AMD divested ZT’s manufacturing arm to Sanmina in late October 2025 and reports it as discontinued operations, so it inflates the balance sheet and FY2025 cash flow ($1,216M discontinued-ops OCF) more than the revenue line.

The real story is FY2024–FY2026: a demand-driven inflection, not financial engineering. FY2025 grew 34.3% organically to $34,639M, and Q1-FY2026 accelerated to +38%. (FACT) Three engines:

• Data Center — $16,635M in FY2025 (+32%), and a record $5,775M in Q1-26 (+57% YoY). This is the MI300/MI350 Instinct GPU ramp going from roughly nothing in FY2023 to a multi-billion-dollar business inside two years, layered on top of EPYC server-CPU share gains. Management guides Data Center to grow >60% annually over the next three-to-five years and the AI franchise to “tens of billions” by 2027.
• Client — $10,640M in FY2025, +51%. (FACT) Critically, the 10-K attributes this to a 31% increase in unit shipments and a 15% increase in ASP — volume and price both rising, a textbook share-gain-with-pricing-power signature, not a pure cyclical bounce. Q1-26 Client was +26% YoY.
• Gaming and Embedded are the low-quality counterweights. Gaming ($3,910M, +51% in FY2025) is semi-custom console SoCs in the seventh year of a console cycle; management already guides FY2026 semi-custom revenue to decline by a “significant double-digit percentage,” and second-half gaming to fall >20% versus first-half on memory/component cost pressure. Embedded ($3,454M, −3%) is the legacy Xilinx franchise, still digesting an industrial inventory correction.

Forward opportunities. The bull catalysts are concrete and dated: MI350 ramping now, MI400 series plus the “Helios” rack-scale system as a 2026–2027 inflection (management’s stated answer to NVIDIA’s NVLink/rack-scale advantage), EPYC “Venice” on the server roadmap, AI PCs lifting Client ASPs further, and sovereign-AI demand. (ASSUMPTION) The >35% revenue CAGR and >$20 non-GAAP EPS targets from the November 2025 Analyst Day depend almost entirely on Instinct scaling against a single dominant incumbent — these are aspirations, not commitments.

Verdict — high-quality growth, but front-loaded into one bet and partly cyclical at the edges. The Client share gain (units + ASP) and the EPYC server franchise are durable, evidence-backed, high-quality growth. The Data Center GPU ramp is real and large but is the opposite of de-risked: it is concentrated, hyperscaler-customer-concentrated, margin-dilutive (below), and contested by a competitor with ~90% share and a deeper software moat. Gaming and a chunk of Embedded are cyclical/one-time and rolling over. Netting it out: the quality of the growth that has already happened is high; the quality of the growth being underwritten by the share price is unproven and binary. The FY2022 Xilinx step-up should never be counted as organic momentum.

6. Financial Quality

Five-year financial summary (GAAP; $M except per-share):

(i) FY2025 non-GAAP operating income is implied: GAAP $3,694M + $2,254M intangible amortization + $1,638M SBC. Reported FY2025 non-GAAP GM was 52%; non-GAAP diluted EPS $4.17.

Gross-margin trajectory and mix. (FACT) GAAP GM bottomed at 44.9% in FY2022 (Xilinx intangible amortization newly burdening COGS, plus the PC downcycle), recovered to 49.5% in FY2025, and reached 52.8% GAAP / 55% non-GAAP in Q1-26. The recovery is genuine but the mix dynamics cut against the bull narrative: Data Center GPU is dilutive to corporate gross margin. Instinct carries lower margins than EPYC server CPUs or the ~65–70%-gross-margin Embedded/FPGA franchise. The evidence sits in the segment math — (FACT) Data Center operating income rose only $121M (to $3,603M) on a $4,056M revenue increase in FY2025, and DC operating margin actually fell to 21.7%. Two culprits: Instinct mix, and ~$440M of net inventory/related charges from the U.S. export ban on the MI308 China part. The MI308 saga also flatters more recent optics — Q4-25 GM of 57% included a $360M MI308 reserve release; ex-release and ex-China the underlying Q4 GM was ~55%. (INTERPRETATION) As Instinct grows from a multi-billion to a “tens of billions” business, it mathematically pulls blended GM toward the low-50s unless EPYC and server-CPU strength (management cites server CPU +70% in Q2) offsets it. The 55–58% long-term GM target is therefore a mix-and-scale bet, not an entitlement.

Operating leverage is real and accelerating. GAAP operating margin went 1.8% (FY2023) → 7.4% → 10.7% (FY2025) → 14.4% (Q1-26), with operating income growing faster than revenue as the business scales past the FY2023 trough. R&D, however, is rising fast — FY2025 R&D was $8,091M (23.4% of revenue), up from $5,872M in FY2023, and Q1-26 opex grew 42% YoY. This is appropriate for a company in a roadmap race, but it means margin expansion is being earned against a heavy and growing fixed-cost base; any revenue stall would de-lever hard, exactly as it did in FY2023.

The GAAP-to-non-GAAP gap — and why half of it is not legitimate. (FACT) FY2025 non-GAAP diluted EPS of $4.17 sits 57% above GAAP $2.65 — a $1.52/share, ~$2.5B pre-tax wedge. It decomposes into two very different items in the −$4,007M “All Other” segment: $2,254M of acquisition-related intangible amortization and $1,638M of stock-based compensation. The amortization is a non-cash, finite-lived purchase-accounting artifact of the Xilinx/Pensando deals; excluding it is defensible. (INTERPRETATION) The SBC add-back is not. SBC is a real, recurring, dilutive economic cost — AMD pays employees ~$1.64B/year in equity it would otherwise pay in cash, and non-GAAP simply pretends it is free. Adding it back to GAAP operating income inflates the figure to ~$7,586M and lifts non-GAAP operating margin to ~22%, roughly double the 10.7% GAAP reality. So the honest read of FY2025 operating economics is between the two poles: better than GAAP (because intangible amortization is genuinely non-economic) but materially worse than non-GAAP (because SBC is genuinely a cost). Non-GAAP is flattering, by approximately the SBC amount.

SBC and dilution — managed, but cash-funded. (FACT) SBC as a percentage of revenue is falling: 6.1% (FY2023) → 5.5% → 4.7% (FY2025), a healthy direction as revenue scales. Diluted shares rose only modestly, 1,625M → 1,636M from FY2023 to FY2025, because AMD repurchased $1,316M of stock (12.4M shares) in FY2025. (INTERPRETATION) Dilution is contained — but note the mechanism: roughly a full year of buyback cash ($1.3B) was consumed almost entirely offsetting one year of SBC ($1.64B), and still didn’t fully neutralize it. The buyback is not returning capital so much as mopping up issuance.

FCF generation and conversion. (FACT) FCF (OCF from continuing operations minus capex) was $5,519M in FY2025, up from $2,405M in FY2024, and Q1-26 set a record at ~$2.6B (25% of revenue). Conversion is improving, but the multi-year pattern shows FCF is volatile and working-capital-sensitive: it fell to $1,121M in FY2023 and FY2025 absorbed a $2,189M inventory build to support Data Center demand (inventory $7,920M at year-end). AMD is asset-light (capex only $974M, ~2.8% of revenue — it fabs at TSMC), which is the structural reason FCF can scale quickly when the top line moves.

Balance sheet — fortress, with a goodwill caveat. (FACT) At year-end FY2025: cash + short-term investments $10,552M ($12.3B by Q1-26), total debt $3,222M, equity $62,999M — net cash positive by ~$7.3B. No liquidity or solvency concern. The caveat is composition: goodwill ($25,126M) plus acquisition intangibles ($16,705M) total $41.8B, or 54% of the $76.9B balance sheet — overwhelmingly the Xilinx purchase price. No impairment has been taken; the 10-K discloses standard impairment-test sensitivity but no charge, and given Embedded’s still-solid 36% segment margins, an impairment is not imminent. (OPEN QUESTION) But a sustained Embedded downturn would put that $25B goodwill at risk, and the FPGA business has been the weakest segment for two years.

ROIC/ROE — depressed, but by accounting, not operations. (FACT) FY2025 ROE was ~7.2% and normalized ROIC ~5.8%. (INTERPRETATION) Both are structurally suppressed by the Xilinx purchase-accounting goodwill/intangibles inflating invested capital and book equity. Strip the $41.8B of goodwill/intangibles and ROIC on the operating asset base is ~23% — i.e., the underlying business earns strong returns; the headline ratio is a denominator artifact of overpaying (in equity) for Xilinx at the 2022 valuation peak. The reported single-digit ROE is real for a shareholder buying the book, but it understates operating quality.

Verdict — economics do improve with scale, and reported profitability is partly real. The operating leverage (1.8% → 14.4% operating margin), asset-light FCF generation, falling SBC intensity, and ~23% ex-PPA ROIC all confirm that scale genuinely improves the economics — this is a real, improving business, not a cash-burner. But two honest qualifications govern the verdict: (1) non-GAAP overstates economics by adding back recurring SBC, so true operating margin sits well below the ~22% non-GAAP figure; and (2) the margin engine of future growth (Instinct GPU) is the least profitable part of the mix, so the 55–58% GM / >35% operating-margin long-term targets require a mix-and-scale outcome that has not yet been demonstrated. Reported profitability is real where it counts (cash, returns on operating assets), inflated where the company most wants you to look (non-GAAP), and depressed where the headline ratios sit (GAAP ROE/ROIC, on Xilinx goodwill).

7. Capital Allocation

Capital allocation is where AMD’s transformation story meets its sharpest scrutiny. Lisa Su has rebuilt AMD from a near-bankrupt also-ran into a credible #2 in accelerated computing — an undeniable operational achievement. But operational excellence and intelligent capital allocation are different things, and on the latter the record is mixed: a genuinely shrewd strategic vision repeatedly funded with equity issued at or near valuation peaks, a buyback that does not even offset its own dilution, and, most recently, an anchor-customer warrant that hands away up to ~10% of the company.

M&A — strategically right, financially expensive. The defining capital event remains the Xilinx acquisition — ~$49 billion, all-stock, closed February 2022 (FACT). Strategically it is defensible and arguably correct: Xilinx brought FPGA/adaptive computing and the embedded franchise that is today among AMD’s most profitable and most stable segments, diversifying a company that was otherwise a pure x86 duopolist. The problem is the price and the currency. AMD paid a full strategic premium in all stock, at the early-2022 valuation peak, using its own then-elevated shares — the textbook moment to issue equity, which is precisely why it transferred so much value to Xilinx holders. The deal sits on the balance sheet as the bulk of $25.1 billion of goodwill (12/27/2025) and generates a permanent ~$2.3 billion/year of acquisition-related intangible amortization (FY25; $2.4B FY24) — a non-cash but real drag on GAAP earnings quality that will persist for years (FACT, FY25 10-K). Goodwill has never been impaired, which is the bull’s defense; but “we overpaid in stock at the top and it hasn’t been written down” is faint praise. Pensando (~$1.9B, 2022) added DPUs/networking and is strategically coherent though small.

The most instructive recent deal is ZT Systems — $4.4B, closed March 31, 2025 (FACT). Here management showed genuine discipline: AMD wanted ZT’s rack-scale design team and IP to compete with NVIDIA’s full-system approach, but did not want to become a low-margin contract manufacturer. So it retained the design business and, by October 2025, divested the ZT manufacturing operation to Sanmina (FACT), settling contingent consideration with former ZT holders. Buying the capability and selling the commodity is exactly the right instinct — CFO Jean Hu framed it at the November 2025 Analyst Day as proof of “an M&A machine” weighing acquisitions against organic investment and buybacks. INTERPRETATION: ZT is the template of good allocation; Xilinx is the cautionary tale of good-asset/peak-price.

R&D — the right place to spend, but the bar is rising. R&D ran $8.09 billion in FY2025, up sharply from $6.46B (FY24) and $5.87B (FY23) — roughly 21–25% of revenue depending on the denominator (FACT). This is the one form of capital deployment that is unambiguously necessary: AMD cannot stay in the accelerator race against NVIDIA’s vastly larger budget without it, and the EPYC/Instinct roadmap is the product of it. But through a Marathon capital-cycle lens, AMD is pouring capital into a capital-attracting boom — high returns drawing in capital and competition — which raises the ROIC hurdle every spend must clear. R&D earns its keep only if the MI400/Helios and EPYC Venice generations convert into durable share and margin; the spend is a bet, not a moat.

Buybacks — not even treading water. The repurchase authorization is large (up to $14B; $9.4B remaining at 12/27/2025), but execution is modest and, more damningly, ineffective at the share-count level. FY25 open-market repurchases were $1,316M ($862M FY24, $985M FY23), plus ~$0.6B/year buying back stock merely to cover tax withholding on employee equity (FACT). Despite all of this, shares outstanding still rose from 1,622M to 1,630M over FY25. INTERPRETATION: the buyback is not reducing the share count — it is failing to fully offset $1.64B of annual SBC dilution. A buyback that runs alongside net share issuance is, in substance, a partial SBC-offset dressed up as capital return, not value-accretive capital allocation. And what is repurchased is bought at elevated prices, consistent with the firm’s habit of transacting in its own shares near highs rather than on weakness.

The OpenAI warrant — buying an anchor customer with equity. The signature 2025 event: alongside a product agreement to deploy 6 GW of AMD GPUs, AMD issued OpenAI a warrant for up to 160 million shares at a $0.01 strike (FACT, Oct 2025 8-K / 10-K). That is ~9.8% of the current 1,630M share count — potentially the single largest equity giveaway in AMD’s modern history. The warrant vests in tranches tied to GPU purchase milestones and specified AMD share-price targets. INTERPRETATION: this is financial engineering to manufacture a marquee reference customer — a customer-acquisition cost paid in equity. It is defensible if the 6 GW translates into the tens of billions of high-margin Instinct revenue management implies; it is value-destructive if deployments slip, or if the share-price vesting conditions are satisfied by multiple expansion (i.e., AMD’s stock rising on AI euphoria) rather than by AMD out-executing. Critically, the company is diluting existing holders by up to ~10% to win business it then books as its own growth — a structure that flatters the narrative while quietly transferring ownership. Open question: the warrant’s P&L geography (potential contra-revenue) is not yet fully clear and bears watching.

Balance sheet — conservative, appropriately. AMD carries ~$12.3B cash+STI against ~$3.9B debt and pays no dividend (and says it won’t soon) (FACT). For a company in a capital-intensive technology arms race with binary product cycles, holding net cash and prioritizing R&D/optionality over a dividend is the correct posture — this is the one box management checks cleanly.

Verdict: capital allocation is competent but not exemplary — strong strategic vision undercut by a repeated willingness to issue equity cheaply. Su’s team picks the right strategic targets (Xilinx’s franchise, ZT’s design capability, OpenAI’s reference halo) and runs a conservative balance sheet. But three patterns should trouble an owner: (1) the company’s primary M&A and partnership currency is its own stock, issued at or near peaks (Xilinx) or as warrants struck at a penny (OpenAI); (2) the buyback does not offset SBC, so the share count grows; and (3) goodwill/amortization from peak-price deals permanently depresses GAAP earnings quality. INTERPRETATION: this is the capital allocation of a company that views its richly-valued equity as a free resource. It has worked so far because the stock has compounded — but that is a function of the AI cycle, not of disciplined per-share value creation. A correct read is: excellent operator and strategist, only a fair allocator of capital on a per-share basis.

8. Changes and Headwinds — Last Two Years

The last two years recast AMD from “the company finally beating Intel in CPUs” into “the only credible #2 to NVIDIA in AI accelerators” — a far larger ambition that brings far larger execution and policy risk.

Strategic and structural changes. AMD reorganized reporting around its AI priorities and built out a full systems strategy rather than selling chips alone — the rationale behind retaining the ZT Systems design team (closed March 2025) while divesting ZT manufacturing to Sanmina (October 2025) (FACT). The headline strategic shift is the pivot to rack-scale, full-stack AI infrastructure (the MI450/MI400-based “Helios” rack), explicitly mirroring NVIDIA’s platform approach and an admission that selling discrete GPUs is insufficient to win the data center. Alongside, AMD signed a string of marquee AI commitments — most prominently OpenAI (6 GW, with the 160M-share warrant) plus partner/sovereign-AI relationships (Oracle, Meta, and government deployments) that anchor the demand narrative (FACT/INTERPRETATION).

Product launches. The Instinct cadence accelerated to roughly annual: MI300X → MI325X → MI350/MI355X → MI400/MI450 (Helios) (FACT, 10-K and Analyst Day). On the CPU side, EPYC “Turin” → “Venice” extends the server franchise that remains AMD’s profit engine and market-share-gain story versus Intel. The strategic question (carried by) is whether this hardware cadence closes the gap to NVIDIA’s CUDA/ROCm software and NVLink scale-up moat; the capital question is simply that this cadence is what the rising R&D bill (now $8.1B) is buying.

Regulatory and litigation — the dominant headwind. China export controls are the clearest negative development. The MI308 restriction forced an ~$800 million inventory and related charge in Q2 2025 (FACT) — a direct write-off of product AMD could no longer ship freely. AMD subsequently obtained some licenses, but in August 2025 U.S. officials signaled an expectation of a 15% revenue-share on licensed MI308 China sales (FACT) — an unprecedented, quasi-tariff arrangement that, if codified, would tax a slice of China revenue off the top and invites litigation/uncertainty. INTERPRETATION: China is now a structural overhang, not a one-time event — both a lost-revenue risk and a margin/policy wildcard with no resolution date (OPEN QUESTION: will the 15% become regulation?).

Other headwinds. Gross-margin mix pressure is real: ramping data-center GPUs at scale carries lower initial margins than the mature server-CPU and embedded businesses, so a faster GPU mix can dilute blended gross margin even as revenue grows — management’s long-term model (55-58% gross margin, >25% FCF margin, per the Analyst Day) is an aspiration the current mix does not yet hit. Competition is intensifying on two fronts: NVIDIA’s entrenched platform above, and hyperscaler custom silicon (Trainium, TPU, MTIA) below, which can in-source the very accelerator demand AMD is chasing.

Leadership and board. Changes have been orderly rather than disruptive: a new Chief Accounting Officer appointed (December 2025), routine EIP bonus determinations (February 2026), and at the May 2026 annual meeting director Jon Olson did not stand for re-election (retiring), with the slate including Su, Talwalkar, and Vanderslice (FACT, 8-Ks/DEF 14A). The same May 2026 8-K disclosed a new unsecured financing facility and a 65,000,000-share increase to the 2023 Equity Incentive Plan — the latter a continuation of the equity-as-currency pattern flagged in No CEO/CFO succession concerns; Lisa Su (Chair & CEO since 2022) remains the key person, and key-person risk around her is genuine.

Insider transactions — a clear, if unsurprising, read. The Form 4 corpus over the trailing months contains zero open-market purchases (code P) by any officer or director (FACT). Activity is uniformly option exercises (M), open-market sales (S), tax withholding (F), and grants (A) — distribution, not accumulation. Lisa Su’s transactions are all sales and grants, with her sales executed under Rule 10b5-1 plans (e.g., an 85,000-share sale), which mutes the signal but does not reverse it. Other officers (Norrod, Papermaster, Grasby, Denzel, Guido) sold into strength at prices ranging roughly $200 to $522 (FACT). Insiders collectively own only ~0.41% of shares, so alignment runs through equity compensation (Su’s FY25 Summary-Comp-Table total was $55.16 million, ~$50.6M of it stock/options) rather than through founder-style ownership. INTERPRETATION: none of this is alarming for a serial-acquirer growth company with heavy equity comp — but it is the opposite of conviction-buying. No insider is using their own cash to buy AMD shares at these levels.

Verdict: the last two years strengthen the strategic position but weaken the risk profile and earnings quality — net, they raise both the upside and the stakes. The AI pivot, systems strategy, and customer wins materially expand the addressable opportunity and are real positives. But they are accompanied by genuine new negatives: a China overhang with an $800M scar and an unresolved 15%-revenue-share threat, GPU-mix margin pressure, intensifying competition from both NVIDIA and hyperscaler ASICs, and a capital-structure pattern of issuing equity (warrant + 65M-share plan increase) to fund the ambition. Insider behavior — uniform selling, ~0.41% ownership — offers no offsetting vote of conviction. These changes make AMD a bigger, more exciting, and more fragile thesis than two years ago, not a safer one.

9. Risk Analysis (Risk Matrix)

The risks below are ordered by their bearing on the thesis. The dominant feature of AMD’s risk profile is that nearly every material downside is correlated to a single variable — the durability of the AI data-center capital cycle. AMD has diversified its product portfolio (EPYC, Instinct, Ryzen, Radeon, Xilinx/embedded, semi-custom) but its valuation and forward earnings are now concentrated on one bet: that hyperscaler/sovereign/AI-native AI infrastructure spend keeps compounding and that AMD captures “double-digit share” of it. That concentration is the through-line of the matrix.

Top risk — AI-capex digestion meets a 99.7th-percentile multiple (Risks 1 + 2, the joint tail). The single most important fact about AMD’s risk profile is that its valuation and its largest growth driver are exposed to the same shock. AMD trades at ~21x EV/sales — above NVIDIA’s ~20x P/S and near Broadcom’s ~24x — on roughly a quarter of NVIDIA’s operating margin. That multiple is sustainable only while the market extrapolates the current AI build-out forward at the Analyst-Day slope. The Marathon capital-cycle lens is unambiguous here: when every credible participant (NVIDIA, AMD, Broadcom, and all three memory makers) is simultaneously expanding capacity into “insatiable” demand, the supply side is being seeded for a digestion phase, and high returns attracting capital is the precondition for mean reversion. AMD does not need a recession to de-rate; it needs only one quarter where a flagship hyperscaler pauses, or where 2027 DC-AI guidance fails to step up, for a multiple resting on extrapolation to compress violently — the more so because there is no valuation floor (P/S at the richest level in the company’s history).

Second — the Instinct-economics question is the load-bearing assumption (Risk 3). The bull case requires not just that AMD sells a lot of MI450/MI500, but that AI silicon eventually carries corporate-average-or-better margins. Management’s own Q1-26 commentary undercuts this on the way up: MI450 is explicitly “below corporate average” gross margin, with EPYC, client mix, and embedded cited as the offsets that hold the blended 55-58% target. In other words, the high-quality, high-margin part of AMD is the CPU/embedded franchise, and the AI-GPU ramp is currently dilutive to margin even as it drives the multiple. The structural reason matters: AMD is the merchant #2 selling against a CUDA-locked incumbent that earns ~60% operating margins precisely because it controls the software ecosystem and prices accordingly. A fast-following open-ecosystem challenger competing partly on TCO and openness has a different, lower margin ceiling. If Instinct settles into a 35-45% gross-margin business rather than a 60%+ one, the >35% operating-margin / >$20-EPS model is reachable only via volume that strains the 80%+ AI CAGR assumption — a higher bar than the price implies.

Third — competitive encirclement from both flanks (Risks 4 + 5). AMD is squeezed between NVIDIA above (CUDA moat, full-stack rack systems, pricing power, ~90% share) and hyperscaler ASICs below (Broadcom/Marvell co-design ~95% of custom silicon; Meta MTIA, Google TPU, Amazon Trainium, Microsoft Maia all in production). The merchant-GPU “second source” thesis — that hyperscalers want an alternative to NVIDIA — is real and is AMD’s single best structural argument, but it is bounded on both sides: the share NVIDIA cedes is contested by AMD and by the customers’ own internal silicon. Management itself sizes ASICs at 20-25% of the accelerator TAM. AMD’s addressable second-source opportunity is therefore the merchant-GPU residual after NVIDIA’s retained share and after in-housing — a large but contested pool, not the open field the headline TAM ($1T silicon by 2030) suggests.

10. Valuation Discussion (Embedded Expectations)

Comparables — the AI-silicon complex. AMD must be valued against the cluster of companies levered to the same AI capital cycle, because the market is pricing them off a shared narrative. The table below pulls each peer’s metrics from comparable-company data (June 2026).

(NVDA/AVGO/TSM/MU figures: peer industry analysis, June 2026; INTC directional. AMD: yfinance/AZI 2026-06-08.)

The table delivers one verdict immediately: on every profit-anchored measure, AMD is the most expensively valued name in the complex relative to the profit it currently earns. Its ~21x EV/sales exceeds NVIDIA’s ~20x P/S and sits just below Broadcom’s ~24x — yet AMD converts those sales at ~11-14% GAAP operating margin against NVIDIA’s ~60%, Broadcom’s ~40%, and TSMC’s ~51%. NVIDIA, the most profitable business in the public market, trades at a lower forward P/E (~24x) than AMD (~37x on next-FY consensus). The only way AMD’s multiple is defensible is if one underwrites a dramatic margin and earnings inflection — i.e., the Analyst-Day model — and assigns it high probability and near-term timing.

Reconciling the forward-P/E spread (FACT). The quoted forward P/E ranges from ~37x to ~65x. This is not a data error; it is fiscal-year dispersion on a steep ramp. Current-FY (FY2026) consensus non-GAAP EPS sits near $7.37 → ~$65 forward P/E on FY2026. Next-FY (FY2027) consensus is near $12.96 (the MI450/Helios ramp year) → ~$37 forward P/E on FY2027. The ~$5.6 of EPS the Street expects to be added in a single year is the spread. The reader should treat ~37x as “the multiple if you believe the 2027 inflection lands,” and ~65x as “the multiple on the last fully-de-risked year.” Both are demanding; the difference between them is the entire investment debate compressed into two numbers.

Embedded-expectations math. At ~$800B EV and ~21x trailing EV/sales, what must the market be underwriting? Work it from the Analyst-Day model backward. Management’s 3-5 year framework is: >35% revenue CAGR off the $34.6B FY2025 base; 55-58% gross margin; >35% non-GAAP operating margin; >25% FCF margin; and >$20 non-GAAP EPS. A >35% revenue CAGR compounds $34.6B to roughly $115-155B by 2029-2030 (the band reflects the 3-vs-5-year window and whether DC-AI runs at the 80%+ AI CAGR). Hold that against the EPS target: $20+ non-GAAP EPS on ~1.63-1.79B shares (the upper bound includes the OpenAI warrant) implies ~$33-36B of non-GAAP net income — consistent with ~$130B+ of revenue at >35% operating margin and a 13-15% tax rate. The model is internally coherent; the question is solely probability and timing.

Now price it. Apply a normalized exit multiple of 30-35x to $20 of EPS — a multiple appropriate for a high-growth but maturing semiconductor franchise, well below today’s ~37-65x but above the market — and you get a ~2028-2029 value of roughly $600-700 per share. Discount that back to mid-2026 at a 10-12% cost of equity over ~3 years: PV ≈ $430-560. That straddles the current ~$475-490. The conclusion is stark and clean:

At ~$475-490, the market is paying today, in present value, for AMD to substantially hit its own self-set Analyst-Day target near the early-to-middle of the 3-5 year window. There is little embedded margin of safety for a shortfall or a slip, and material upside only if AMD beats the model — DC-AI scaling past $100B faster than planned, gross margin pressing toward the 58% top of the range, or the multiple refusing to normalize.

Scenario analysis (explicit assumptions). Modeled to FY2029 (~3-year horizon), non-GAAP basis, exit multiple applied to that year’s EPS; share count ~1.7B (partial warrant). These are illustrative ranges, not point forecasts.

The asymmetry is the point. The base case roughly matches today’s price three years out (a low-single-digit annualized return before discounting); the bull case requires AMD to beat an already-aggressive management plan; and the bear case — which only requires the AI cycle to behave like every prior capital cycle — implies a halving or worse. The distribution is left-skewed at this price.

What the market is pricing correctly vs. aggressively. Correctly: the EPYC server franchise is a genuine, demonstrable, high-quality compounder — ~40% server share on a path management credibly argues toward >50%, growth that is unit-driven not just price-driven (Q1-26 server growth “much more unit-driven”), accretive to gross margin, and now riding an expanding CPU TAM (raised from ~$60B to >$120B by 2030 in six months on agentic-AI CPU pull). The market is right to pay up for that. It is also right that AMD has real AI revenue and a credible #2 position. Aggressively: the market is extrapolating that Instinct scales to NVIDIA-like blended profitability. It structurally may not. Management itself flags MI450 as below corporate-average gross margin and leans on the CPU and embedded mix to defend the 55-58% target — which means the AI engine driving the multiple is, today, a margin drag, not a margin source. Pricing AMD’s AI franchise on NVIDIA’s economics is the central aggressive assumption embedded at ~$475-490, and it is the assumption least supported by the filings.

Is the AI ramp already priced? Largely, yes — through the early-to-middle of the Analyst-Day window. The reverse-DCF shows the current price ≈ PV of management delivering its own plan. That does not make AMD overvalued in any provable sense (a true believer in the >35% CAGR and margin convergence can construct a fair-to-cheap case on FY2027 numbers); it makes AMD a stock with no cushion — priced such that meeting expectations earns a mediocre return and missing them is punished severely. As an embedded-expectations matter, the burden of proof has shifted entirely onto execution. No price target; the analysis frames ranges and the expectations the price embeds.

11. Variant Perception

Consensus belief. The sell-side is firmly constructive: an aggregate rating near 4.16/5 with ~32 strong-buy ratings, and a narrative that has hardened into orthodoxy — “AMD is the #2 AI compute platform, the inevitable merchant second-source to NVIDIA, an EPYC share-gainer with an expanding CPU TAM, and a multi-year compounder on its way to $20+ EPS.” Importantly, this is not a crowded short (short interest ~2.75% of float). The variant-perception axis here is therefore not bull-vs-bear on direction — almost everyone is constructive — but a disagreement about degree and price: how much of the Analyst-Day model is real, how fast it arrives, and whether the franchise’s AI economics resemble NVIDIA’s or something structurally lower.

The strongest bull case. AMD is at a genuine inflection, and the inflection is bigger than the model. The MI450/Helios generation is the first AMD rack-scale system that competes with NVIDIA at the system level rather than the chip level, arriving exactly as hyperscalers, sovereigns, and AI-natives are desperate for a credible second source and as power/supply constraints make any additional capacity valuable. Lead-customer forecasts already exceed AMD’s initial 2027 plan; OpenAI (6GW) and Meta (6GW) anchor the ramp; and management raised the path to “exceed” the 80%+ DC-AI CAGR. Simultaneously, EPYC is compounding on a server TAM that just doubled (to >$120B) because agentic AI consumes CPU cycles AMD didn’t previously count — server CPU guided >70% YoY. Put the data center >60% CAGR (toward >$100B), client toward 40% share, and embedded recovering, and >$20 EPS isn’t a stretch target — it’s a base case that the steep next-FY consensus EPS ($12.96) already half-validates. On FY2027 numbers (~37x forward) a compounder growing 35%+ is not expensive. Falsified if: DC-AI revenue fails to reach the “tens of billions” 2027 run-rate, or a flagship anchor (OpenAI/Meta) cancels/delays, or MI450 gross margin stays stuck below corporate average into 2028 — any of which breaks the EPS bridge.

The strongest bear case. AMD is priced for perfection at the 99.7th percentile of its own valuation history, on the most cyclical demand driver in technology, while earning a fraction of the margins its peers do. The bear holds three planks. (1) Margin structure: Instinct is dilutive today and may never earn NVIDIA-like economics — AMD is the open-ecosystem #2 competing partly on TCO against a CUDA-locked incumbent, so its AI gross-margin ceiling is structurally lower; the >35% operating-margin target leans on EPYC/embedded mix, not the AI engine. (2) Capital cycle: the entire AI complex — NVDA, AVGO, TSM, and all three memory makers — is building into “insatiable” demand simultaneously; that is the Marathon precondition for a digestion phase, and AMD has no valuation floor to cushion one. (3) Encirclement and dilution: NVIDIA above (~90% share, CUDA, pricing power), hyperscaler ASICs below (~95% Broadcom/Marvell co-design; management concedes ASICs at 20-25% of TAM), and a ~9.8% OpenAI warrant overhang that caps per-share upside as the very deal it underwrites delivers. Falsified if: AMD prints a quarter showing Instinct gross margin converging toward corporate average at scale, or DC-AI revenue durably exceeds the 80% CAGR with the multiple holding — proof that the franchise earns its multiple rather than merely promising to.

The 3-5 assumptions that matter most.

1. Instinct’s terminal margin. Does AI silicon reach corporate-average (≥55%) gross margin, or settle structurally lower (35-45%)? This single variable governs whether $20 EPS is reachable at the planned revenue or requires implausible volume. Most important, least proven.
2. AI-capex durability vs. digestion. Does hyperscaler/sovereign AI spend compound through 2028, or digest? Determines both AMD’s revenue and its multiple — the two move together.
3. Second-source share ceiling. How much of the NVIDIA-ceded, post-ASIC merchant-GPU residual can AMD actually win? Bounds the DC-AI revenue line.
4. EPYC share trajectory. Does server share progress 40% → 50%+ on the doubled TAM? This is the highest-conviction plank — the part bull and bear most agree on — and the genuine high-quality core of the franchise.
5. Exit multiple normalization. Does the market still pay 30-35x for $20 EPS in 2028-29, or does a maturing growth rate compress it toward 20-25x? Governs the return even if every operating assumption hits.

Synthesis. The honest variant-perception read is that consensus is probably right about the business (real inflection, real franchise, real share gains) and aggressive about the price (pricing the AI engine on NVIDIA’s economics and the model on early-window timing, with no margin of safety). The disagreement that matters is narrow but decisive: bull and bear largely agree on today’s facts — Q1-26’s 38% growth, the EPYC strength, the MI450 pipeline — and diverge almost entirely on the durability and margin structure of the AI piece. That is the variant perception: not whether AMD grows, but whether its AI franchise earns the profitability the price already assumes.

12. Fact vs. Interpretation Table

13. Open Questions

1. EPYC unit vs. revenue share. Management cites “>40%” server revenue share. What is the unit share (Mercury Research)? If EPYC ASPs run well above Xeon, unit penetration is lower and the >50% path is harder. (Material to the moat verdict.)
2. Instinct terminal gross margin. The single most important unknown. Does AI silicon converge to ≥55% GM at scale, or settle at 35-45%? Management says “below corporate average” today and gives no quantified terminal figure. (Governs whether $20 EPS is reachable at planned revenue.)
3. OpenAI-warrant accounting geography. Is the warrant’s fair value treated as contra-revenue, an operating expense, or equity issuance cost — and how does it hit reported margins as it vests? Not yet fully clear from disclosures.
4. China AI vs. non-AI split. Of the $7.75B China revenue, how much is MI308/AI (subject to the 15% revenue-share threat and licensing) vs. ordinary client/embedded/console? (Bounds the policy-risk magnitude.)
5. Will the 15% MI308 revenue-share become codified regulation, and does it set a precedent for other parts? Unresolved, no timeline.
6. DC-AI customer concentration. What % of FY2026-27 Data Center GPU revenue is the top 3 customers (OpenAI/Meta/Oracle)? Not disclosed; central to the concentration risk.
7. Goodwill impairment risk. If Embedded’s downcycle persists, at what point is the $25.1B Xilinx goodwill tested down? Segment has been weakest for two years.
8. Helios/MI400 ramp timing. First-gen rack-scale integration — does it ship at volume in 2H26 as guided, or slip into 2027? (Governs the EPS-bridge year.)

14. What Must Be True

For the BULL case to be right (and for the current price to prove conservative):

• Instinct scales AND its margins converge upward. DC-AI revenue compounds at the 80%+ CAGR to “tens of billions” by 2027 and Instinct gross margin climbs toward corporate average as MI400/MI450/Helios volume and software (ROCm) mature.
  • Falsification test: By FY2027, if Instinct gross margin is still flagged “below corporate average” or DC-AI revenue run-rate is below ~$30B, the >$20-EPS bridge breaks and the multiple is unsupported. A single hyperscaler-anchor (OpenAI/Meta) cancellation or a Helios slip past 2027 also falsifies.

For the BEAR case to be right (and for the current price to prove far too high):

• The AI capital cycle digests while AMD’s AI economics stay structurally sub-NVIDIA. Hyperscaler capex pauses or grows slower than the extrapolated slope; AMD — the lower-share, later-ramp #2 — sees allocation cut first; and Instinct never earns more than ~40-45% gross margin because it competes on price/TCO against CUDA lock-in.
  • Falsification test: If AMD prints two consecutive quarters of DC-AI revenue acceleration with blended gross margin expanding (Instinct accretive), and the multiple holds through a capex-wobble headline, the bear’s “no-floor, sub-NVIDIA-economics” thesis is refuted.

The pivot both cases share: the durability and margin structure of the Instinct franchise. Everything else (EPYC strength, the fortress balance sheet, the cyclicality of Gaming/Embedded) is largely agreed. Watch Instinct gross margin and DC-AI revenue run-rate, quarter by quarter — that is the falsifiable core of the entire thesis.

15. Source Appendix

See the dedicated Source Appendix (AMD_source_appendix.md) for the full citation list. Primary sources relied upon:

• AMD FY2025 Form 10-K (filed 2026-02-04; fiscal year ended 2025-12-27), Items 1, 1A, 7, 8; Notes 4 (Segments), 5 (Acquisitions/Divestitures), goodwill/intangibles. SEC EDGAR CIK 0000002488.
• AMD Q1-FY2026 Form 10-Q (filed 2026-05-06; quarter ended 2026-03-28).
• AMD FY2021–FY2024 Form 10-Ks for multi-year history.
• AMD DEF 14A (2026-03-27) — executive compensation, incentive metrics, board.
• AMD 8-K corpus (trailing 24 months) — ZT Systems / Sanmina, OpenAI warrant, buyback authorization, exec/board changes, financing facility.
• Form 3/4/5 corpus — insider-transaction read.
• Transcripts: Analyst/Investor Day (2025-11-11), Q4-FY2025 call (2026-02-03), Q1-FY2026 call (2026-05-05), CES 2026 (2026-01-05), BofA Tech Conference (2026-06-02) — AZI transcripts feed, mirrored locally.
• SEC EDGAR XBRL (companyconcept) — revenue, net income, gross profit, R&D anchors.
• yfinance (via scripts/fetch.py) and AZI fundamentals/news feeds — price, market cap, EV, multiples, valuation percentiles, news (accessed 2026-06-09/10). Third-party aggregated data reconciled to filings.
• Public comparable-company data (June 2026): NVIDIA, Broadcom, TSMC, Micron, SK Hynix, Infineon — public filings and market data used for industry, comps, and capital-cycle cross-read.
• Analytical frameworks: Greenwald & Kahn, Competition Demystified; Marathon/Chancellor, Capital Returns (via the installed investment-research-frameworks skill).

All third-party AI scores (AZI sentiment, analyst ratings/targets) are treated as signals, not evidence, and never as a recommendation or price target.

APPENDIX A — Standard Diligence Questionnaire

Report date: 2026-06-09 | An independent equity-research note Supplemental to the memo (not counted toward its length target). Fact/Interpretation/Assumption labeled where it matters.

General

What thoughtful questions have other investors asked about this company? The serious questions have converged on a small set: (1) Will Instinct’s gross margin ever converge toward corporate average, or is AMD structurally a lower-margin “second source” against CUDA lock-in? (2) How much of the AI-accelerator TAM does AMD actually address once hyperscaler ASICs (TPU/Trainium/MTIA/Maia) are netted out? (3) Is the EPYC server-share gain durable now that AMD must defend ~40%+ rather than breach Intel? (4) How dilutive is the OpenAI 160M-share warrant, and does it convert into real high-margin revenue? (5) Is hyperscaler AI capex at a cyclical peak? Notably, the debate is about degree and price, not direction — consensus is broadly constructive (sell-side ~4.16/5; short interest only ~2.75% of float).

Cyclicality & Earnings Nature

Are earnings at a cyclical high or low? Mixed by segment but, in aggregate, closer to an up-cycle than a trough — and the valuation is at an all-time high (P/S 99.7th percentile of AMD’s own history). Data Center is in an AI investment super-cycle (likely mid-to-late boom); Gaming/semi-custom is late in the console cycle and guided down >20% H2; Embedded is recovering off an inventory-correction trough; Client is mid-cycle on share gains. (Interpretation) Consolidated earnings are depressed by the Xilinx amortization tail but flattered by an AI demand peak — net, not a cyclical-low entry point.

Driven by the external environment or internal actions? Both. Internal: genuine, durable EPYC/Ryzen share gains (execution + TSMC node lead). External: the AI-capex wave driving Instinct, and the PC/console/industrial cycles. The high-quality part is internal; the part the multiple depends on is external.

How stable are revenues? Moderately unstable. No recurring/subscription revenue; revenue fell 3.9% in FY2023 (downcycle) and grew 34% in FY2025. Instinct is lumpy and hyperscaler-deal-driven. EPYC and Ryzen are the most stable franchises.

Outlook for products/services? Strong near-term: MI350 ramping, MI400/MI450 “Helios” rack-scale in 2026-27, EPYC Venice, AI PCs. The roadmap cadence is annual and credible; the open question is margin and share capture, not product capability.

How big will this market be? Management frames a >$1T silicon TAM by 2030 (data-center accelerators + CPUs + networking). Growing, global, but (Interpretation) promotional — it counts the ASIC slice AMD does not capture and rests on extrapolated hyperscaler capex.

Business Quality & Competitive Moat

Is the industry getting more or less competitive? More. NVIDIA entrenched above (CUDA, NVLink, ~90% share), hyperscaler ASICs encroaching below, ARM pressuring x86, and capital flooding the AI-accelerator pool (Marathon capital-cycle warning).

How profitable is the business (ROIC, ROE)? Headline ROE ~8% / ROIC ~6% — mediocre, depressed by $41.8B Xilinx goodwill+intangibles (54% of the balance sheet). Strip purchase accounting and operating ROIC ~23% (Interpretation). The truth is in between: a good operating business carried on an over-paid-for balance sheet.

How profitable is the industry / barriers to entry? Bifurcated. Leading-edge manufacturing (TSMC, ASML) and NVIDIA earn extraordinary returns behind deep barriers; merchant challengers like AMD earn less. Barriers to AMD’s own franchises: x86 ISA duopoly (high), FPGA design-in switching costs (high), AI-GPU software/ecosystem (AMD is the challenger — low barrier in its favor).

Can the business be easily understood? Reasonably — four product lines, a fabless model, clear competitors. The hard part is forecasting Instinct revenue and margin.

Can it be undermined by foreign low-cost labor? Not labor — but by subsidized/By foreign state-backed competition (Chinese domestic accelerators, encouraged by US export controls) and by customers in-sourcing (ASICs). The threat is capital and policy, not wages.

Do brands matter? Modestly. “EPYC”/“Ryzen”/“Radeon”/“Instinct” carry weight with OEMs and enthusiasts, but purchasing is driven by performance/TCO/software, not brand loyalty. CUDA is the closest thing to a brand-as-moat in this space — and it belongs to NVIDIA.

Nature of competition / customers’ switching costs? Competition is performance-per-watt, TCO, software ecosystem, and supply allocation. Switching costs are high in FPGA (Embedded) and meaningful in x86 (recompilation/validation), but low in merchant GPU (ROCm lowers the cost of trying AMD without making leaving painful) and low in Client/Gaming (commodity).

Financial Condition & Balance Sheet

Assets not fully recognized on the balance sheet? The EPYC/Ryzen franchise value and the ROCm software stack are internally generated and largely unrecognized. Conversely, $25.1B of Xilinx goodwill over-states asset value relative to the depressed Embedded earnings.

Off-balance-sheet liabilities? No unusual ones flagged. Standard purchase commitments to TSMC and operating leases. The OpenAI warrant (up to 160M shares) is a contingent dilution overhang rather than a debt liability.

How conservative is the accounting? GAAP is conservative (full SBC and amortization expensed). Non-GAAP is aggressive — it adds back ~$1.64B/yr of recurring, real, dilutive SBC, inflating operating margin to ~22% vs 10.7% GAAP. Use GAAP plus the amortization add-back; do not credit the SBC add-back.

How CapEx-hungry is the business? Light. Capex ~$974M, ~2.8% of revenue (fabless — TSMC carries the fab capex). This is the structural reason FCF scales fast with revenue.

Capital Allocation & Management

How much FCF, and how is it used? FY2025 FCF $5.5B (Q1-26 record ~$2.6B). Uses: R&D ($8.1B), modest buybacks ($1.3B), M&A (ZT $4.4B), no dividend. (Interpretation) The buyback fails to offset SBC — shares rose — so it is SBC-mopping, not value-accretive return.

Significant acquisitions recently? Xilinx ($49B all-stock, 2022 — good asset, peak price/currency); Pensando ($1.9B, 2022); ZT Systems ($4.4B, 2025 — design retained, manufacturing sold to Sanmina; the template of good allocation).

Buying back shares? Yes but ineffectively — $9.4B remaining on a $14B authorization; FY25 repurchases didn’t reduce the share count.

Issuing large amounts of new shares to insiders? SBC ~$1.64B/yr (4.7% of revenue, falling); a 65M-share EIP increase approved May 2026; and the OpenAI penny-warrant for up to 160M shares (~9.8%). Equity is treated as a freely-spent currency.

Compensation policy? Lisa Su FY25 Summary-Comp-Table total $55.16M (~$50.6M equity). Incentives tie to revenue, non-GAAP operating income, and relative TSR (per proxy). Alignment is via comp, not ownership — insiders hold only ~0.41%.

Motivations of management? Operationally elite, execution-driven; Su’s reputation is staked on the AI roadmap. (Interpretation) The incentive structure rewards growth and non-GAAP metrics (which flatter SBC), and the penny-warrant/equity-issuance pattern suggests management optimizes for strategic wins and narrative over per-share discipline.

Valuation & Market Data

ADR, MLP, or K-1 issuer? No — ordinary US common stock (NASDAQ: AMD), 1099 reporting. No K-1.

Dividend policy? None, and none expected near-term. Appropriate for a company in a capital-intensive product race; capital goes to R&D and optionality.

How profitable is the business? GAAP operating margin 10.7% (FY25) → 14.4% (Q1-26); non-GAAP ~22% (overstated). Gross margin ~49-53%. Real but well below NVIDIA (~60% op) / Broadcom (~40%) / TSMC (~51%).

Is net income diverging from cash from operations? Largely consistent — OCF exceeds GAAP net income (non-cash amortization + SBC added back in OCF), and FCF is healthy. The divergence to watch is the working-capital/inventory build ($7.9B inventory, +$2.2B) supporting the Data Center ramp.

Risks & Downside

What factors would cause the stock to decline? A hyperscaler AI-capex digestion quarter; a flagship anchor (OpenAI/Meta) cancellation/delay; Instinct gross margin failing to improve; multiple compression off the 99.7th-percentile P/S; an EPYC share stall as Intel 18A closes the gap; China/MI308 escalation; a Helios/MI400 ramp slip.

Risk of a catastrophic loss? Low-probability/high-impact tails: a Taiwan/TSMC disruption (binary, uninsurable — all leading-edge silicon is TSMC-fabbed) and loss of Lisa Su. Neither is base-case but both are severe.

Chance of a total loss? Negligible. Net-cash balance sheet (~$8B), real cash-generative franchises (EPYC/Embedded), and no solvency risk. The realistic downside is a 40-60% de-rating (bear scenario ~$170-240), not a zero.

Recent News & Events

Has the business environment changed recently? Incrementally. The tape is quiet (per the news sweep) — a £2B/5yr UK AI commitment (Imperial/Cambridge) and a Broadcom-led semi sector selloff. The bigger 2025-26 changes: the OpenAI 6GW deal + warrant, MI350 launch, Analyst-Day target raise, the segment reorg to three segments, and the China/MI308 controls. None is a fresh thesis-breaker as of the report date.

Significant acquisitions? ZT Systems (2025) + Sanmina divestiture — see Capital Allocation.

Change in accounting policies? Segment reporting changed to three segments (Q1 2025), with prior periods recast. No other material accounting-policy change flagged.

Recent changes — new markets, facilities, management? Pivot to rack-scale “AI factory” systems (Helios); new CAO (Dec 2025); director Jon Olson retiring (May 2026); new unsecured financing facility (May 2026). Orderly, not disruptive.

APPENDIX B — Source Appendix

Report date: 2026-06-09 | An independent equity-research note Primary sources before secondary; every non-obvious fact in the memo traces here. Third-party aggregated/AI data is labeled and was reconciled to filings.

A. Primary — SEC filings (EDGAR, CIK 0000002488)

B. Primary — Management calls & events (AZI transcripts feed, mirrored to output/AMD/transcripts/)

C. Secondary / third-party (signals, reconciled to filings)

E. Frameworks

• Greenwald & Kahn, Competition Demystified — moat taxonomy, share-stability & ROIC tests, EPV.
• Chancellor / Marathon, Capital Returns — supply-side capital-cycle analysis, asset-growth anomaly.
• (Installed skill: investment-research-frameworks.)