AMD Deep Research

Business Model & Unit Economics

How does AMD make money? Analyze its product lines (CPU/GPU/EPYC/Instinct), pricing power, revenue mix, and margin structure.

AMD Business Model & Profitability Structure

Product Lines, Pricing Power, Revenue Mix, and Margin Analysis

Executive Summary

AMD (Advanced Micro Devices) is a fabless semiconductor company that monetizes chip design capability through three reportable segments: Data Center (EPYC server CPUs and Instinct AI GPUs), Client & Gaming (Ryzen CPUs for consumer and commercial PCs, semi-custom gaming console SoCs, and Radeon discrete GPUs), and Embedded (Xilinx-heritage FPGAs and adaptive SoCs). FY2025 total revenue was $34.6B (+34% YoY), with Data Center contributing approximately 48% for the first time1. The most significant structural shift documented across FY2021 to Q1 2026 is Data Center's revenue share rising from 22% to 56%, driven by the dual ramp of Instinct AI GPUs and EPYC server CPUs2.

Key Findings

Data Center Share Rose from 22% to 56% (Q1 2026), Driven by EPYC and Instinct

Data Center is AMD's largest and fastest-growing revenue segment. FY2025 full-year Data Center revenue was $16.6B (+32% YoY)3, accelerating to $5.8B in Q1 2026 (+57% YoY, +7% QoQ)4, representing 56% of quarterly revenue. Going back to FY2021, the segment stood at $3.7B (22% of total) — a roughly 4.5x expansion in absolute dollar terms over five years2.

The segment contains two core product families5:

EPYC Server CPUs (Turin / Genoa / Milan generations): Deployed across all 10 of the world's largest hyperscalers and in more than 60% of the Fortune 1006. Q1 2026 server CPU revenue grew more than 50% YoY; Q2 2026 guidance calls for more than 70% YoY growth7. AMD's current server CPU revenue market share is approximately 40%, with a stated target of exceeding 50%8. AMD has updated its server CPU TAM estimate to more than $120B by 2030 (CAGR >35%), with Agentic AI head-node CPU demand cited as the largest new component7.

Instinct AI GPUs (MI300 / MI325 / MI350 / MI450 roadmap): Data Center AI revenue exceeded $5B in FY20249. Q4 2025 Instinct shipments set a record, led by the MI350 series ramp10. AMD issued warrants to both OpenAI and Meta (up to 160M shares each at $0.01/share), vesting based on Instinct GPU purchase milestones — Meta separately committed to deploying up to 6 gigawatts of AMD GPUs11.

Data Center segment operating margin trajectory: FY2024 27.7% → FY2025 21.7% (depressed by MI308 export control charge) → Q4 2025 33% (benefiting from MI308 inventory reserve release) → Q1 2026 28%412.

Ryzen ASP Rising Sharply: FY2025 Full-Year +15%, Q3 2025 Single-Quarter +38%, Driven by Mix Shift

AMD's client CPU business (Ryzen) has shown the most pronounced ASP appreciation in recent periods. The multi-year progression is as follows131415:

The ASP acceleration reflects a product mix shift upward — growth of Ryzen AI 400-series, commercial PC, and premium mobile SKUs, where average prices are materially higher16. Lisa Su characterized this in the FY2025 earnings call as "ASP mixing up the stack" and identified enterprise as the priority growth area for 202616. The FY2023 ASP decline of 16% illustrates the cyclical vulnerability of the client business: PC market weakness combined with supply chain inventory corrections drove both prices and volumes lower simultaneously14.

EPYC Pricing Philosophy: Volume First, ASP Gains from Generational Mix, Not Price Premium

EPYC server CPU growth follows a different logic than Ryzen: it is primarily unit-driven rather than ASP-driven. CFO Jean Hu disclosed explicitly at the BofA Global Technology Conference (June 2026) that approximately 2/3 of server CPU growth — both the >50% Q1 2026 result and the >70% Q2 2026 guidance — came from unit expansion, with only 1/3 from ASP increases17.

"2/3 of that actually are unit growth. So majority of our revenue growth actually come from unit expansion versus ASP increase... the ASP will continue to increase because increasingly, you need a very high core count CPUs." — Jean Hu, BofA Global Technology Conference, June 202617

The ASP uplift mechanism was also clarified by Jean Hu: each new EPYC generation ships with higher core counts, which structurally raises ASP — this is not active market pricing power over Intel18. Turin (Zen 5) crossed 50% of total server CPU revenue in Q1 202618. AMD's stated pricing policy is to pass through inflationary cost increases to customers, while deliberately avoiding using supply tightness to extract price premium19.

On memory costs: AMD does not directly procure DRAM for server or PC deployments — memory is purchased by OEM/ODM partners and hyperscalers themselves and does not flow through AMD's P&L20. This structure largely insulates EPYC gross margin from memory price swings, in contrast to consumer/gaming products.

Competitive positioning: AMD frames EPYC as delivering superior TCO versus Intel Xeon, with a portfolio covering "every price point, every core count, every workload"21. Against ARM-based server CPUs (which carry ASPs of $3,000-$5,000, above current EPYC ASPs), AMD's counter-positioning emphasizes breadth of workload coverage versus ARM's point-product approach22. Next-generation Venice (Zen 6, TSMC 2nm) is positioned as delivering "substantially higher performance per socket and per watt versus competitive x86" and "more than 2x throughput per socket versus leading ARM-based AI solutions"23.

Instinct GPU Gross Margin Below Corporate Average, but Generational Improvement Is an Explicit Management Commitment

Instinct GPU gross margin structure is the primary reason the Data Center segment's blended margin runs below the sum of its parts. Jean Hu has confirmed across multiple earnings calls:

• Instinct GPU gross margin is currently below AMD's corporate average24

• Each successive generation (MI300 to MI350 to MI450) should carry higher gross margin than the prior25

• At the start of each generation's ramp, margins tend to be lower; they improve within the generation as yield improves, test costs decline, and scale builds25

• MI450 (Helios rack-scale system, ramping H2 2026) is expected to begin production below corporate average gross margin, then improve as volume scales26

"the gross margin should progress each generation. So it's kind of a dynamic gross margin, but in the longer term you should expect each generation should have a higher gross margin." — Jean Hu, FY2025 Q4 Earnings Call25

AMD's current strategy is to prioritize Instinct revenue scale over near-term GPU margin maximization24. HBM (High Bandwidth Memory) is the most significant cost pass-through component in Instinct GPU COGS, representing a meaningful share of total GPU cost structure.

Embedded Segment: Highest Margin, Xilinx FPGA Heritage, High Entry Barriers

Embedded is AMD's highest-margin reportable segment, with segment operating margin consistently in the 36-39% range2728 — materially above Data Center (21-28%) and Client & Gaming (16-20%). The segment includes FPGAs, Versal Adaptive SoCs, embedded x86, and System-on-Modules acquired through Xilinx, serving aerospace and defense, industrial, communications, and automotive end markets.

The structural sources of Embedded's premium margin: (1) IP-intensive products that resist commoditization; (2) high customer switching costs from deep hardware customization; (3) long design cycles — once a design win is secured, revenue typically flows for multiple years.

Embedded underwent a significant inventory correction cycle: FY2023 $5.3B → FY2024 $3.6B (-33%) → FY2025 $3.5B (-3%), reflecting channel destocking and mixed end-market demand. Cumulative design win backlog has exceeded $50B since the Xilinx acquisition, with $17B added in FY2025 alone, providing a visible pipeline for future revenue conversion29.

Gross Margin Structure: Mix-Driven, GAAP vs. Non-GAAP Gap Primarily from Xilinx Acquisition Amortization

*Q4 2025 GAAP gross margin of 57% includes a $360M one-time release of previously written-down MI308 inventory reserves; normalized gross margin excluding this release and MI308 China revenue was approximately 55%30. The non-GAAP vs. GAAP gap is driven primarily by acquisition-related intangible amortization in COGS (FY2025: $1.03B) and stock-based compensation31.

FY2025 one-time item: AMD recorded approximately $800M of inventory and related charges on Instinct MI308 GPUs in Q2 2025 due to U.S. export restrictions on China shipments, then reversed approximately $360M of those charges in Q4 2025 after receiving export licenses. Net FY2025 COGS impact: approximately $440M, depressing GAAP gross margin by approximately 1.3 percentage points3233. Excluding this item, FY2025 GAAP gross margin would have been approximately 51%.

Gross margin drivers (CFO Jean Hu framework)34:

• Tailwinds: Rising Data Center revenue share (EPYC server CPU is margin-accretive); Client mix shift toward premium commercial; Embedded recovery (highest-margin segment)

• Headwinds: Instinct GPU ramp-phase margin dilution at each new generation launch; Gaming semi-custom revenue decline (neutral to mildly positive, as semi-custom carries below-average margins)

OpEx Structure and Operating Leverage: Heavy R&D Investment, with OpEx Growth Trailing Revenue as the Long-Term Model Commitment

FY2025 SG&A grew +52% versus R&D +25%, as AMD built out its enterprise server and commercial PC go-to-market infrastructure35. Jean Hu stated in the Q1 2026 earnings call that for full-year 2026, R&D growth is expected to outpace SG&A growth36. The All Other operating loss of approximately $4B (FY2025) — comprising $2.3B of Xilinx acquisition-related intangible amortization and $1.6B of SBC37 — is the principal reconciling item between segment operating income totals and GAAP total operating income.

Operating leverage has begun to materialize: FY2025 operating income grew +94% on +34% revenue (approximately 2.8x leverage ratio)38; Q1 2026 operating income grew +83% on +38% revenue39. Lisa Su committed in the FY2025 earnings call: "we've always said in our long-term model that OpEx should grow slower than revenue. And we would expect that in 2026 as well"40.

Gaming Segment: Highly Cyclical, Semi-Custom Entering Decline, Discrete Radeon GPU Is the Growth Vector

Gaming (semi-custom game console SoCs for PlayStation and Xbox, plus discrete Radeon GPUs) is the most cyclical of AMD's segments: FY2022 peak $6.8B → FY2024 trough $2.6B (-62%) → FY2025 recovery $3.9B (+51%). The FY2025 rebound was driven by semi-custom volume, but management has guided that semi-custom SoC revenue will "decline by a significant double-digit percentage" in 2026 as the PlayStation/Xbox cycle enters its seventh year41. The next Xbox (AMD semi-custom SoC) is expected in 2027, which would restart the semi-custom ramp. Radeon discrete GPUs (RDNA 4 / RX 9000 series) represent the growth vector within Gaming, though at an absolute scale materially below peak semi-custom revenue.

Multi-Period Segment Comparison Tables

Segment Revenue Time Series (USD billions)

The table below shows the full FY2021-Q1 2026 revenue trajectory to illustrate the structural shift in AMD's revenue mix. Source: AMD 10-K FY2025 (file_id 114051), 10-Q Q1 2026 (file_id 164007), 10-Q Q3 2025 (file_id 218005).

1234

Segment Operating Margin Comparison (excluding All Other)

Note: Segment operating margins reflect each segment's own revenue and directly allocated costs, excluding All Other (which contains Xilinx amortization of $2.3B and SBC of $1.6B). These figures are therefore higher than the company's reported GAAP operating margin42.

272812

Ryzen ASP Decomposition Table (FY2023 to Q3 2025)

The table below decomposes Ryzen client CPU revenue growth into ASP and volume components, illustrating that recent ASP gains reflect mix shift toward premium SKUs rather than market pricing power. Source: AMD 10-K FY2025 (file_id 114051), 10-Q Q3 2025 (file_id 218005).

13141543

Gross Margin and Key P&L Metrics

The table below shows the multi-year progression of gross margin (GAAP and non-GAAP), R&D investment, and operating income — illustrating AMD's improving profitability structure as the revenue mix shifts toward Data Center.

*Q4 2025 includes $360M MI308 inventory reserve release; normalized GAAP gross margin approximately 55%30

313233353839

One-Time Factors and Accounting Basis Notes

The following are short-term or one-time factors that have limited bearing on long-term structural analysis, listed to help readers distinguish signal from noise:

1. MI308 Export Control Charge (FY2025 one-time): AMD recorded approximately $800M of inventory charges in Q2 2025 related to U.S. export restrictions on Instinct MI308 GPUs for China, then reversed approximately $360M in Q4 2025 after receiving export licenses. Net FY2025 COGS impact: approximately $440M3233. Excluding this item, FY2025 GAAP gross margin was approximately 51% and non-GAAP approximately 52%.

2. Q4 2025 Gross Margin: 57% Reported vs. ~55% Normalized: The reported 57% includes the $360M one-time inventory reserve release. Management provided the ~55% normalized figure for comparability30. Both are factually correct; the difference is a matter of which metric better represents underlying run-rate performance.

3. FY2025 Segment Reporting Structure Change: Beginning Q1 2025, Client and Gaming were merged into a single reportable segment. The FY2025 sub-disclosures of Client $10.6B and Gaming $3.9B appear in MD&A only, not as separate GAAP reportable segments1.

4. Gaming Semi-Custom 2026 Decline: Management has guided to a "significant double-digit percentage" decline in semi-custom SoC revenue in 2026, reflecting the console lifecycle — not AMD's underlying competitive position or Data Center trajectory41.

5. Non-GAAP Gross Margin FY2025: 51% vs. 52%: Source materials contain both figures. AMD's formal non-GAAP disclosure in the FY2025 earnings call states 52%; the 51% figure appears to be a rounded GAAP-adjacent estimate. This analysis uses 52%.

Sources, Information Limitations, and Unresolved Conflicts

Information limitations:

• The breakdown of Instinct GPU versus EPYC revenue within the Data Center segment is not separately disclosed in audited financial statements; the ">$5B Instinct FY2024" figure comes from management commentary, not audited GAAP disclosure

• Segment-level gross margins are not separately disclosed (AMD reports combined cost of sales and operating expenses at the segment level, preventing gross margin isolation)

• No absolute dollar ASP data are available for EPYC or Instinct SKUs (AMD does not publicly disclose unit pricing)

• The split between discrete Radeon GPU and semi-custom revenues within the Gaming sub-segment is not fully disclosed

Unresolved conflicts: (1) Non-GAAP FY2025 gross margin stated as both 51% and 52% in different source documents — 52% used here per formal earnings call disclosure. (2) Data Center operating margin of 27.7% appears in both FY2024 annual results and Q1 2026 quarterly results — this is a numerical coincidence, not a data error, and has been noted accordingly in the analysis.

Moat & Competitive Landscape

What are AMD's competitive moats vs Intel and NVIDIA? Analyze IP advantages, fab-less model, switching costs, and competitive positioning.

AMD Competitive Moats vs Intel and NVIDIA

Executive Summary

AMD operates across four distinct moat dimensions: (1) a structural IP and architecture advantage built on chiplet/packaging leadership, Infinity Fabric interconnect, and an x86 cross-license that constrains Intel's M&A options; (2) a capital-light fab-less model that converts Intel's $105 billion PP&E burden into AMD's structural R&D focus advantage, amplified by privileged "lead HPC customer" access to TSMC's N2 node; (3) accumulating switching costs from enterprise server qualification depth (3,000+ OEM solutions, 1,600 EPYC cloud instances) and multi-gigawatt co-engineering partnerships with hyperscalers; and (4) a widening process node lead over Intel in CPU, with a meaningful but acknowledged software gap behind NVIDIA in GPU. The most significant near-term competitive change is AMD's shift in GPU software strategy — using AI (Claude/Codex) to accelerate ROCm development — described by AMD's VP of AI as "the biggest leap since the early days of our strategy."44

Key Finding 1: Chiplet IP and Infinity Fabric — AMD's Deepest Hardware Moat

AMD's chiplet architecture and Infinity Fabric interconnect represent the most difficult-to-replicate element of its IP stack. CTO Mark Papermaster described Infinity Fabric as "the hidden gem" — a proprietary die-to-die interconnect that enabled three distinct MI300 variants (CPU-only, CPU+GPU, GPU-only) from a single silicon base.45 The 4th-generation Infinity Fabric is in production; the 5th generation targets 112 Gb/s scale-up for MI400/Helios.45

On chiplet packaging, AMD has reached its 4th generation of 3D V-Cache stacking with AMD CTO confirming: "still without a competitor in the industry to that capability."46 From first-gen EPYC to Helios, AMD achieved a 30x increase in available silicon for compute and memory.46 MI400 doubles packaging interconnect density versus MI300.47

SerDes IP from the Xilinx acquisition (112 Gb/s → 224 Gb/s in MI400 → 448 Gb/s roadmap) powers both PCIe Gen 6 for Venice and the scale-up network architecture of Helios.48 AMD's formal patent portfolio as of December 2025: approximately 18,900 patent matters worldwide (12,600 issued + 6,300 pending), per the FY2025 10-K.49 AMD's FAD 2025 CTO presentation cited 13,000+ issued patents covering Infinity Fabric, chiplet integration, SoC architecture, and security building blocks.50

The AMD-Intel x86 Cross-License: Structural IP Moat

AMD and Intel maintain a broad Patent Cross-License Agreement (Exhibit 10.10, dated November 11, 2009) covering both companies' entire patent portfolios, including x86 and AMD64 ISAs.49 A change-of-control by either party terminates the agreement, giving AMD potential leverage in any Intel M&A scenario.51 AMD and Intel also co-operate a joint x86 advisory group, explicitly protecting "billions and billions of lines of code that customers have out there with x86."52 AMD was also first to integrate a dedicated NPU on the same SoC as an x86 CPU for AI PCs.53

Key Finding 2: Fab-less Model Economics — Capital Efficiency vs Intel IDM, Process Leadership vs Both

Capital structure divergence: Intel's IDM model produced net PP&E of $105.4 billion as of December 2025, with $34.5 billion in construction-in-progress. Intel's own 10-K acknowledges this creates "higher capital expenditures and R&D spending than many of our fabless competitors."54 Intel absorbed $878 million in inventory reserves from 18A early-ramp cost overruns in FY2025, plus $950 million in non-cash asset impairments.55 NVIDIA's fab-less model yields comparable capital efficiency to AMD: net PP&E of $10.4 billion in FY2026.56

Process node access: AMD's Venice (Zen 6) is the first HPC product in the industry confirmed as taped out and in volume production on TSMC N2 (2nm GAA transistors), announced jointly by Lisa Su and TSMC CEO C.C. Wei.57 TSMC confirmed N2 entered high-volume manufacturing in Q4 2025 with good yield.58 Intel's competitive P-core response (Diamond Rapids) is targeted for mid-2027 on Intel 18A-P — approximately a 12-month process node lead for AMD in server CPU.

Intel IDM execution failures: Intel's FY2025 10-K disclosed supply constraints at the Intel 7 node that limited ability to meet market demand in both CCG and DCAI, with constraints expected to persist into Q1 2026.59 Intel Foundry external revenue was only $307 million in 2025 out of Intel's $52+ billion total revenue.60 Intel Q2 2026 foundry operating loss: $2.1 billion.70

Supply risk (AMD): AMD's fab-less model concentrates geopolitical supply risk in Taiwan and Malaysia/China for assembly per FY2025 10-K disclosures.61 TSMC supply has been constrained since 2024 for AI customers.62 Mitigation: TSMC Arizona Fab 21 Phase 1 in high-volume production; ~30% of N2+ capacity eventually targeted for Arizona.63 Lisa Su noted "very deep relationships across the supply chain" as a competitive advantage.64

AMD GPU cadence acceleration: AMD's annual CDNA GPU cadence now targets flop rate doubling every year post-2025 (vs. every two years before), enabled by consistent TSMC node access. CDNA 4 (MI355X) ramping; CDNA 5 (MI450/Helios) next year; CDNA 6 (MI500) in design.65

Key Finding 3: Server CPU Switching Costs — Qualification Depth Is AMD's Strongest Moat vs Intel

Enterprise server qualification creates compounding switching costs:

• OEM solution depth: Leading server providers offer over 3,000 solutions optimized for AMD's 4th and 5th generation EPYC CPUs. Large businesses deploying EPYC on-premises more than doubled in 2025.66

• Cloud instance breadth: EPYC-powered cloud instances grew >50% YoY to nearly 1,600 in 2025, with hyperscalers launching 500+ new AMD-based instances.67

• Customer velocity: Q1 2026 server CPU revenue grew >50% YoY, with both cloud and enterprise each growing >50%. Lisa Su confirmed: "More customers validating and ramping platforms at this stage than with any prior EPYC generation."68

• Market share: AMD holds 46.2% server x86 CPU revenue share in Q1 2026 (vs. 33.2% unit share), indicating AMD wins the premium high-core-count segment.69

• Intel's response: CEO Lip-Bu Tan confirmed "we continue working on improving the competitiveness against our competitor" — relying on process recovery (18A roadmap) rather than ecosystem lock-in.70

Co-engineering partnerships: AMD-Meta agreement covers a custom GPU accelerator co-designed on MI450 architecture for up to 6 gigawatts of Instinct GPUs across multiple product generations.71 The AMD-Anthropic deal (July 22, 2026) includes AMD's $5 billion equity investment, 2 gigawatts of MI455X in Helios racks (first gigawatt H1 2027), and multi-year ROCm co-development using Claude AI.72 These co-engineering agreements entangle customers' future hardware roadmaps with AMD's product development across multiple generations.

Financial co-marketing programs: AMD 10-K discloses MDF programs providing reimbursement for AMD-integrated product marketing — creating financial disincentive to switch.73

Source: Mercury Research via third-party press; Q1 2026 from AMD Q1 2026 earnings call.6968

Key Finding 4: Software Moat Gap vs NVIDIA — The Primary Structural Constraint, Under Active Attack

NVIDIA's CUDA ecosystem is the most durable competitive moat AMD faces in AI accelerators. NVIDIA's FY2026 10-K discloses over 7.5 million developers use CUDA — a 20-year ecosystem spanning 6,000+ supported applications.74 NVIDIA describes its full platform as "hardware, systems, software, algorithms, libraries, AI models and training data sets, and services."75 Jensen Huang explicitly described NVIDIA's strategy as "accelerating all of the world's tools so that it runs on CUDA" — targeting every industrial toolchain from Cadence to Adobe.76 NVIDIA acknowledged that China export foreclosure "helped our competitors build larger developer and customer ecosystems to challenge us worldwide."77

AMD ROCm trajectory: ROCm is described as "fully battle tested" in top supercomputers and by most demanding hyperscalers.78 ROCm is embedded at rack level in Helios systems (communication libraries and workload tuning), making it an architectural requirement.79 The AMD-Anthropic ROCm collaboration (Claude) and AMD-OpenAI (Codex) — with AMD VP of AI Vamsi Boppana describing feature development time reduced from "a team of 20 people, six months" to "dramatically cut down" — represent AMD's most direct investment to date in narrowing the software moat gap.4480 AMD plans to begin shipping AI-native ROCm platform capabilities in August 2026.80

AMD's open-source strategy creates an asymmetry vs NVIDIA's proprietary stack: because AMD publishes instruction sets, compilers, and toolchains openly, AI coding assistants can learn AMD's platform "right off the bat" while NVIDIA's proprietary toolchains require more curated AI training data.81

AMD's AI accelerator market share target: double-digit share with >80% AI revenue CAGR over 3-5 years.82 Current estimated share: 5-7% (third-party estimates only). NVIDIA holds 80-86% of the data center GPU market per third-party range estimates (not a primary NVIDIA disclosure).

Key Finding 5: NVIDIA's CPU Entry and OEM Design Win Pipeline — New Vectors

NVIDIA's Vera CPU (custom Arm cores, co-designed with Rubin GPU) claims "1.5x faster performance per core, 2x performance per watt, and 4x density per rack compared to x86-based alternatives." NVIDIA projects $200 billion new CPU TAM and "nearly $20 billion in total CPU revenue" visibility in 2026.83 This opens a new competitive front that AMD did not face in prior server CPU competition cycles.

AMD's response: Lisa Su characterizes ARM as "point products relative to a portfolio" — AMD's Venice spans throughput-optimized, power-optimized, cost-optimized, and AI-infrastructure-optimized (Verano) variants.84 AMD targets >50% server CPU market share within both x86 and the overall market.

AMD's embedded and semi-custom design win pipeline: over $45 billion in cumulative won semi-custom design win content across aerospace/defense, automotive, data center, and communications. Embedded design wins exceeded $14 billion in 2024, targeting >$16 billion in 2025, totaling over $50 billion at 21% CAGR.85 Gaming console semi-custom (~$20 billion lifetime revenue) demonstrates AMD's track record of multi-generation single-source OEM relationships.86 For Helios rack-scale, HPE and Lenovo publicly committed to Helios rack systems in 2026.87 OEM/ODM manufacturers Sanmina and WeWin are co-building and validating Helios racks.88

Scope Notes and Reconciled Conflicts

1. CONFLICT-9 (Data Center segment boundary): AMD Data Center FY2025 revenue of $16.6B is the combined CPU+GPU segment — not GPU-only. Instinct GPU sub-total is not separately disclosed.

2. CONFLICT-5 (Server CPU TAM): $120B figure (Q1 2026 call, May 2026) has been superseded by $220B (July 2026 AI Roundtable). Both are management estimates, not third-party data.

3. CONFLICT-3 (CUDA developers): 7.5 million (NVIDIA FY2026 10-K, primary source) used throughout; 5.9 million from secondary blog posts is outdated.

4. CONFLICT-11 (NVIDIA AI market share): Range 80-86% reflects different third-party methodologies — not a single hard figure from NVIDIA filings.

5. AMD Q2 2026: Results releasing August 4, 2026 after market close — not yet in evidence base. Guidance was $11.2B revenue.

6. NVIDIA $76.7B cumulative R&D: A calculated cumulative across annual disclosed figures since inception, not a single-line disclosed figure in NVIDIA's 10-K.

Management, Capital Allocation & Culture

Evaluate AMD's management team led by Lisa Su, capital allocation track record, M&A strategy (including Xilinx acquisition), R&D investment, and corporate culture.

AMD Management, Capital Allocation & Culture

Executive Summary

Under Lisa Su (Chair, President & CEO since 2014), AMD grew annual revenue from approximately $5 billion to $34.6 billion by FY2025 and established a consistent product execution cadence across CPU, GPU, and — through the $48.8 billion Xilinx acquisition (closed February 2022) — FPGA and adaptive computing. The most quantitatively significant recent development is the FCF trajectory: operating cash flow from continuing operations reached $6.5 billion in FY2025 and FCF hit a record $2.6 billion in Q1 2026 (25% of revenue)9091, reflecting a business model generating cash at materially higher rates than during the 2021–2023 integration period. Capital allocation priorities have been R&D-first ($8.1 billion in FY2025, 23.4% of revenue)92, with M&A as the second lever and share repurchases as a tertiary return mechanism.

Key Finding 1: Lisa Su's Execution Track Record — Revenue Scale and Market Share

Lisa Su serves as Chair, President and CEO of AMD (confirmed at the AI 2026 Investor Roundtable)93. Her leadership period encompasses AMD's transformation from a near-bankruptcy position (~$5 billion annual revenue) to a $34.6 billion revenue company in FY2025, a +34% YoY increase92.

The most concrete execution metric is server CPU market share. At the BofA Global Technology Conference (June 2026), CFO Jean Hu stated AMD reached 46% server CPU value share in Q1 2026, with CPU revenue growing 50%+ in Q1 and guided to grow 70%+ in Q2 202694. AMD was also the first to publicly forecast AI-driven CPU TAM expansion — from $60 billion to $120+ billion to potentially $200 billion by 2030 — a forecast that was initially met with skepticism and has since been materially validated94.

A notable CEO-level operational advantage surfaced at the BofA Conference: Matt Ramsay (IR) noted that Lisa Su's personal relationships with TSMC executives help AMD secure fabrication capacity at tight nodes95.

"I think we're very well positioned with the relationships Lisa has personally with the executives in that space to get maybe more than our fair share of the incremental." — Matt Ramsay, AMD IR, BofA Tech Conference June 202695

Management team depth as of mid-2026 (source: AI 2026 Investor Roundtable93):

Key Finding 2: M&A Strategy — Xilinx as Structural Portfolio Expansion, ZT Systems as Capital-Efficient Capability Acquisition

The Xilinx acquisition is AMD's defining M&A event: announced October 26, 2020; closed February 14, 202296.

Dual valuation (both figures are correct with different reference dates):

• $35 billion at announcement (all-stock at 1.7234 AMD shares per Xilinx share)96

• $48.8 billion gross ($46.4 billion net of $2.4 billion cash acquired) at close, reflecting AMD's share price appreciation during the 16-month regulatory review96

Strategic rationale: AMD's pre-Xilinx strategy focused on CPUs and GPUs. Xilinx added FPGAs, Adaptive SoCs, and ACAP products — product categories absent from AMD's portfolio9798. The combination established AMD in embedded markets (automotive, industrial, aerospace & defense, communications) with over 7,000 customers99.

"AMD is focused on high performance computing technology, software and product leadership... As part of our strategy to establish AMD as the industry's high performance computing leader, we announced in October 2020 that we entered into a definitive agreement to acquire Xilinx." — AMD FY2022 10-K98

Post-close Embedded segment performance:

*FY2023 Embedded revenue is an upstream estimate; flagged as requiring direct verification against FY2023 10-K segment table.

FY2024–FY2025 Embedded revenue decline reflects a broad industry inventory correction per management commentary99, not competitive share loss. Design win momentum ($14B → $17B annually) is the leading indicator management points to for eventual revenue recovery.

Lisa Su's own characterization of the Xilinx integration (FAD 2025):

"Xilinx was the first large acquisition that we did. People were always like, Lisa, are you sure you can integrate this and add value and not have synergies? I think what we've shown is excellent execution on the large ones." — Lisa Su, FAD 2025100

Design win growth post-acquisition exceeded the pace of the former standalone Xilinx (FPGA-only portfolio), attributed to AMD's broader product portfolio101.

Acquisition accounting — FY2022 P&L impact:

• Goodwill assigned: $22.784 billion (primarily Embedded segment; not tax-deductible)

• Identified intangibles: $27.308 billion (developed technology $12.295B / 16-yr life; customer relationships $12.290B / 14-yr life)

• FY2022 amortization drag: ~$3.5B/year compressed gross margin from 48% (FY2021) to 45% (FY2022) and operating income from $3.6B to $1.3B106

• No goodwill impairment recorded across FY2022–FY2025107

Adjacent acquisitions — consistent playbook:

• Pensando (May 2022): DPUs and networking; $1.655B total consideration, $1.098B goodwill97

• ZT Systems (March 2025): $3.2B in cash + 8.3M shares; ZT Manufacturing sold to Sanmina for $2.4B cash (October 2025); AMD retained ZT Design Business (IP + engineers). Net effective cost for rack-scale AI design capability: ~$800M103

"In March 2025, we completed the acquisition of ZT Systems for $3.2 billion in cash and 8.3 million shares of our common stock. We retained select intellectual property and employees associated with the design operations (ZT Design Business), and in October 2025, we sold the ZT data center infrastructure manufacturing business (ZT Manufacturing Business) to Sanmina Corporation (Sanmina) for $2.4 billion in cash." — AMD FY2025 10-K103

Lisa Su describes AMD as having built an "M&A machine" — "very confident and comfortable with M&A execution"101.

Key Finding 3: R&D Investment — $1.5B (FY2019) to $8.1B (FY2025), Entirely Headcount-Driven

R&D is AMD's stated first capital allocation priority. Six-year absolute trajectory:

The FY2022 jump (+76%) reflects Xilinx headcount consolidation. FY2024's increase was "primarily due to higher employee-related costs due to an increase in headcount in support of our AI strategy"112. FY2025's +25% ($1.6B) increase was described by Lisa Su as a front-loaded AI roadmap investment: "we have very high conviction in the roadmap that we have" and the OpEx lean-in was "for all the right reasons," with R&D OpEx expected to grow slower than revenue in 2026113.

Headcount as a leading R&D indicator:

• Dec 2024: ~28,000 employees114

• Dec 2025: ~31,000 employees (+~11% net)114

• 2024 Restructuring Plan (Q4 2024, ~4% reduction, $186M charges): explicitly framed as reallocation to AI and enterprise, not a cost-reduction signal115

• 25,000+ engineers in hardware and software (management-stated, FAD 2025)

Management committed at Q1 2026 earnings that R&D will grow faster than SG&A in 2026.

Key Finding 4: Capital Allocation — FCF Ramp and Buyback Discipline

FCF trajectory (AMD definition: OCF from continuing operations minus capex; excludes discontinued operations):

FY2023 was a trough driven by inventory build and working capital; capex is very low given AMD's fully fabless model (~2.8% of revenue in FY2025).

Share repurchase history ($14B total authorization, established May 2021):

AMD pays no dividends. SBC runs $1.4–1.6B/year, largely offsetting buybacks — diluted shares flat at ~1,625–1,637M from FY2023 to FY202592. Contingent dilution: OpenAI warrant for up to 160M shares at $0.01/share, vesting on GPU purchase milestones through October 2030 (vesting trigger is purchase milestones only per 10-K; no stock price condition stated)104.

Balance sheet (Q1 2026): Cash + ST investments $12.3B; total debt $3.2B; net cash ~$9.0B108. Goodwill $25.3B; acquisition intangibles $16.2B — $41.5B combined on total assets of $79.6B108.

Long-term targets (management-stated, Q1 2026 earnings): >$20 EPS over strategic time frame; 55–58% gross margin target121.

ROIC note: AMD does not report ROIC in any filing. Researcher-constructed proxy: operating income $3.7B (FY2025) / total assets $76.9B = ~4.8% ROTA. Adjusted for $41.5B acquisition-related assets: ~10% ROTA ex-intangibles. These are researcher-estimated metrics with denominator sensitivity; not AMD-disclosed figures.

5-year total shareholder return (Dec 2020 – Dec 2025, per 10-K stock performance graph):

• AMD: +134% ($100 → $234)

• S&P 500: +101% ($100 → $201)

• S&P 500 Semiconductors Index: +447% ($100 → $547)

AMD outperformed the broad market but materially underperformed the semiconductor peer index over this window, primarily reflecting NVIDIA's dominance in AI GPU revenue. Period-sensitivity caveat: this window includes the FY2021–2022 peak and subsequent correction; alternative period selections yield materially different relative results.

Key Finding 5: Corporate Culture — Multi-Year Consistency and Talent Investment

AMD's corporate culture is documented with consistent language across FY2021–FY2025 10-Ks: innovative, inclusive, open, respectful. Employee surveys ("AMDer Survey") rank AMD "among the very best for global companies in the technology industry" with manager quality in the top 10% of tech industry peers (management-stated; no independent third-party verification available in local data).

Pay-for-performance culture: FY2025 earnings call cited "higher employee performance-based incentives" as an explicit OpEx driver113. Incentive structure includes short-term and long-term plans with PRSUs for SVP-level and above. Majority of new leaders promoted internally.

Headcount grew net from 28,000 (Dec 2024) to 31,000 (Dec 2025), despite the 2024 Restructuring Plan (~4% reduction)114115 — the restructuring was a reallocation to AI/enterprise growth areas, not a sustained headcount reduction.

Lisa Su adopted a Rule 10b5-1 trading plan (November 2023) for 834,226 shares, a transparency mechanism for pre-planned insider transactions.

Short-term / One-time Factors and Scope Notes

• FY2023 OCF trough ($1.7B): Driven by inventory build and working capital normalization — cyclical, per management characterization.

• FY2022 GAAP operating income decline ($3.6B → $1.3B): Entirely driven by Xilinx amortization (~$3.5B/year); not an underlying business deterioration signal106.

• Embedded revenue decline FY2024–2025: Broad industry inventory correction; design win pipeline ($14B → $17B) is the management-cited leading indicator119120.

• OpenAI warrant: Up to 160M shares at $0.01 — contingent on GPU purchase milestones, not yet vested as of FY2025104.

• ZT Systems net cost framing: ~$800M net is the cash arithmetic; Sanmina also paid in stock and up to $450M contingent consideration is outstanding103.

• Market share figures (46% server CPU value share; "7 of top 10 AI companies using Instinct"): Management-stated; not independently verified third-party data.

Information Sources, Limitations, and Pending Conflicts

Information limitations:

• No Proxy Statement (DEF 14A) locally: CEO pay ratio, exact compensation amounts, PSU performance metrics, board committee structure not verifiable

• Pre-FY2021 financials: only narrative references and FY2021 10-K comparative columns available

• FY2023 Embedded revenue (~$5.3B) is an upstream estimate — flagged for direct verification against FY2023 10-K segment table

• ROIC: not disclosed by AMD; researcher-estimated proxy only

Pending conflict: OpenAI warrant vesting — one EB mentioned "stock price targets"; FY2025 10-K text confirms vesting is based solely on GPU purchase milestones. The 10-K language takes precedence104.

Financial Quality & Cash Flow

Analyze AMD's financial quality: revenue growth trajectory, gross margin expansion, free cash flow conversion, balance sheet strength, and earnings quality.

AMD Financial Quality: Revenue Growth, Gross Margin, Free Cash Flow, Balance Sheet, and Earnings Quality

Executive Summary

AMD's five-year revenue trajectory from FY2019 to FY2025 reflects a compound annual growth rate of approximately 31%, accelerating to +34% YoY in FY2025 to reach a record $34.6 billion, with Data Center displacing legacy segments as the primary revenue engine.123124 Free cash flow has expanded from a FY2023 trough of approximately $1.1 billion to $5.5 billion in FY2025 and an annualized run-rate above $10 billion based on Q1 2026 results, supported by the company's fabless model (CapEx ~2.8% of revenue) and structural amortization add-backs from the Xilinx acquisition.125126 The key earnings quality consideration is the persistent $1.52/share gap between GAAP ($2.65) and non-GAAP ($4.17) EPS in FY2025, entirely attributable to acquisition intangible amortization ($1.38/share) and SBC ($1.00/share) — plus a non-recurring $853 million IRS tax benefit that produced an anomalous (2.5%) effective tax rate.127128

Key Finding 1: Revenue Inflected to +34% Growth in FY2025, Driven by a Structural Shift in Segment Mix

AMD's total revenue trajectory from FY2019 through FY2025 shows two distinct phases: organic share-gain growth (FY2019–FY2021, ~56% CAGR) and post-Xilinx expansion with a cyclical dip (FY2022–FY2025).129130 FY2022 revenue reached $23.6 billion (+44%), inflated by the February 2022 Xilinx acquisition consolidating Embedded from $246 million to $4.6 billion. FY2023 reversed -4% to $22.7 billion on PC market weakness (Client -25%), semi-custom gaming decline, and Embedded inventory destocking.131132 FY2024 recovered +14% to $25.8 billion, then accelerated to +34% to $34.6 billion in FY2025.123

The most significant structural shift is Data Center's rise from 22% of revenue in FY2021 ($3.7 billion) to 48% in FY2025 ($16.6 billion, +32% YoY) and 56% in Q1 2026 ($5.8 billion, +57% YoY).133134135 Client and Gaming, combined in FY2025 at $14.6 billion (+51% YoY), was driven by Ryzen processor demand and AI PC adoption. Gaming as a standalone sub-segment grew +51% YoY in FY2025 (semi-custom consoles), but management flagged a "significant double-digit" decline in FY2026 as the PlayStation/Xbox console cycle enters year 7.136 Embedded declined 3% to $3.5 billion in FY2025, but $17 billion in new design wins in FY2025 (+20% YoY, $50B+ cumulative since Xilinx) signals a multi-year pipeline.137

Structural drivers include: AI hyperscaler capex buildout (EPYC 5th Gen CPU + Instinct MI350 GPU); on-prem enterprise EPYC deployments doubled in FY2025; EPYC cloud instances grew >50% YoY to approximately 1,600; 8 of top 10 AI companies using Instinct in production.138

Management's FAD 2025 long-term guidance: total company revenue >35% CAGR from the $34.6 billion FY2025 base; Data Center >60% CAGR; Core businesses >10% CAGR; EPS target >$20 in the strategic timeframe.139

Revenue Trajectory — AMD Total and Segment Revenue (FY2021–Q1 2026)

Note: FY2022 Embedded surge reflects Xilinx consolidation; Client+Gaming combined per FY2025 segment restructure.140 Sources: 129130131132133134135

Key Finding 2: GAAP Gross Margin Expanded from 45% to 50–54%, Driven Entirely by Revenue Mix — Not Cost Reduction

GAAP gross margin trended from 45% (FY2020–FY2022) through 46% (FY2023), 49% (FY2024), 50% (FY2025), 54% (Q4 2025), and 53% (Q1 2026).141142143 The single largest driver at every step was revenue mix — higher-margin Data Center and Client revenue displacing lower-margin Gaming/semi-custom. The FY2025 10-K explicitly states: "increase in gross margin was due to a favorable shift in revenue mix of higher Data Center and Client revenues."144

FY2022's dip from 48% to 45% was directly caused by Xilinx acquisition amortization entering COGS ($942 million in FY2022, rising to $1,031 million in FY2025).145146 Core cost of sales (excluding amortization) grew 35.8% in FY2025, approximately in line with revenue growth of 34%, confirming that manufacturing unit economics held steady; margin expansion is entirely mix-driven.147

The GAAP-to-non-GAAP gross margin bridge is consistently ~2–3 percentage points, driven by $261 million/quarter of Xilinx intangible amortization in COGS, plus approximately $8 million/quarter of SBC.148149 This gap will mechanically narrow as the Xilinx intangible base depletes.

FY2025 included a one-time drag: approximately $440 million net inventory and related charges from U.S. export controls on MI308 GPUs to China (Q2 2025 COGS). Approximately $360 million was reversed in Q4 2025 when licenses were granted.150151 Absent this charge, FY2025 GAAP gross margin would have been approximately 51.3%.

CFO Jean Hu at the Q1 2026 earnings call outlined FY2026 gross margin puts and takes: tailwinds include server CPU growing >70% in Q2 (accretive to GM), Client moving up the value stack, and Embedded recovery (margin accretive); key headwind is MI450 Helios rack-scale system ramp in H2 2026 (below corporate average at initial ramp). Long-term FAD 2025 target: 55–58% non-GAAP gross margin.152153

Gross Margin Trend — GAAP vs. Non-GAAP (FY2020–Q1 2026)

Note: FY2025 GAAP GM includes net ~$440M MI308 charge; Q4 2025 includes ~$360M reversal.150 Sources: 141142143144145146148149152

Key Finding 3: Free Cash Flow Expanded ~5x in Two Years; Fabless Model Keeps CapEx Structurally Low

AMD's FCF trajectory spans four distinct periods: strong pre-Xilinx ($3.1–3.2 billion in FY2021–FY2022), trough post-integration ($1.1 billion in FY2023), recovery ($2.4 billion in FY2024), and step-function acceleration ($5.5 billion in FY2025, $2.6 billion in Q1 2026 alone at 25% FCF margin).125126154

The CFO bridge from GAAP net income to operating cash flow is dominated by non-cash add-backs in FY2025: D&A $750 million + acquisition intangible amortization $2,254 million + SBC $1,638 million = approximately $4.6 billion total non-cash add-backs on a GAAP net income base of $4,335 million. Working capital was a $2.5 billion drag (inventory +$2,189 million to support the GPU ramp + AR +$121 million, partially offset by AP +$410 million).155 OCF/GAAP net income ratio was approximately 1.50x in FY2025, reflecting large non-cash charges from acquisition accounting.

CapEx as a percentage of revenue has remained structurally low (~2.8% in FY2025, or $974 million on $34.6 billion revenue), consistent with the fabless model outsourcing wafer production to TSMC.125 PP&E net was only $2,312 million at end of FY2025.

Q1 2026 was an exceptional quarter: CFO of $2,955 million (+215% YoY vs. Q1 2025's $939 million) and FCF of $2,566 million on $10.3 billion revenue = 25% FCF margin. Working capital was a net positive in Q1 2026 — AR collections drove a $280 million inflow and accrued liabilities grew $713 million from customer advance payments.156

FCF Conversion Table (FY2021–Q1 2026)

Sources: 125126154155156 *Q1 2026 CapEx single-quarter rate; annualized implies ~$1.6B or ~3.9% of revenue run-rate.

Key Finding 4: Balance Sheet Net Cash +$9.1B; Off-Balance-Sheet Commitments Surged to $25.7B

As of Q1 2026 (March 28, 2026), AMD held $12.35 billion in cash and short-term investments against $3.3 billion in aggregate principal debt, for a net cash position of approximately $9.0 billion.157 This compares to $10.6 billion liquidity and net cash of approximately $7.3 billion at FY2025 year-end.158 The $3.0 billion undrawn revolving credit facility (expires April 2027) provides additional contingency liquidity.158

Debt is entirely fixed-rate, staggered: $875 million at 4.212% due 2026, $625 million at 4.319% due 2028, $750 million at 2.375% due 2030, $500 million at 3.924% due 2032, $500 million at 4.393% due 2052. The March 2025 $1.5 billion issuance pre-funded the 2026 maturity.159 Interest expense grew from $92 million (FY2024) to $131 million (FY2025); GAAP interest coverage approximately 28x.160

Current ratio was 2.85x at FY2025 year-end and 2.73x at Q1 2026.157161 Stockholders' equity grew from $57.6 billion (FY2024) to $64.5 billion (Q1 2026), with retained earnings now $6.7 billion (FY2025) vs. $0.7 billion in FY2023.162

Two structural balance sheet features require context: (1) Goodwill + acquisition intangibles = $41.8 billion (54% of total assets $76.9 billion), entirely from the $35 billion all-stock Xilinx acquisition. No goodwill impairment has been taken to date.163 (2) Off-balance-sheet purchase commitments surged from $12.2 billion (FY2025 year-end) to $25.7 billion (Q1 2026), with $18.3 billion due in the remainder of FY2026 alone — primarily wafer/substrate supply agreements for the AI ramp and multi-year cloud arrangements.164165 Lease guarantees of $4.1 billion were also added in Q1 2026.165

Balance Sheet Strength Summary (FY2022–Q1 2026)

Sources: 157158159160161162163164165

Key Finding 5: GAAP/Non-GAAP Gap Dominated by Acquisition Amortization; FY2025 Tax Rate Anomalous

FY2025 GAAP EPS was $2.65 vs. non-GAAP EPS $4.17 — a $1.52/share gap (GAAP net income $4,335 million vs. non-GAAP $6,831 million).127 Bridge components: SBC $1,638 million ($1.00/share); acquisition intangibles amortization $2,254 million ($1.38/share); IRS tax benefit ($853 million, adds to GAAP but excluded from non-GAAP = +$0.52/share in GAAP vs. non-GAAP); other minor items.

The effective tax rate was (2.5%) in FY2025 — a tax benefit year — due to the $853 million IRS resolution of dual consolidated loss uncertain tax positions approved in April 2025. This is non-recurring. FY2024 ETR was 19%. Unrecognized tax benefits declined from $1.5 billion to $806 million; $571 million in remaining unrecognized benefits represents a contingent future impact.128

Additional FY2025 one-time items: (1) $440 million net MI308 inventory charge in COGS (net of $360 million Q4 reversal).150151 (2) $66 million income from discontinued operations (ZT Systems manufacturing sale).127 (3) Zero restructuring charges vs. $186 million in FY2024.163

Revenue recognized over time declined to 9% of total revenue (from 25% in FY2023), with remaining performance obligations of only $315 million — no evidence of revenue pull-forward.166 Accrued liabilities composition shows no anomalous patterns: customer-related liabilities actually declined from $1,349 million (FY2024) to $1,194 million (FY2025).168

SBC trend: $379 million pre-Xilinx (FY2021) → $1,638 million (FY2025), with annualized Q1 2026 run-rate approximately $1.95 billion. SBC as % of revenue declining: ~6.1% (FY2023) → ~4.7% (FY2025). Net share count essentially flat (1,637M → 1,636M diluted) as $1.3 billion in FY2025 buybacks largely offset SBC dilution.155156167

GAAP vs. Non-GAAP Reconciliation (FY2024, FY2025, Q1 2026)

Note: FY2025 other one-time includes ($853M) IRS tax benefit, $360M MI308 reversal, ($66M) discontinued ops, and minor acquisition/investment items. Sources: 127128141148149169

Short-Term / One-Time Factors and Accounting Notes

The following are flagged as one-time or period-specific items not representative of underlying run-rate performance:

1. MI308 export control charge: ~$800M gross charge in Q2 2025 COGS; ~$360M reversed in Q4 2025 when licenses were granted; net FY2025 GAAP COGS impact ~$440M, depressing GAAP gross margin approximately 1.3pp.150151

2. IRS dual consolidated losses settlement: $853M non-recurring tax benefit in FY2025 producing a (2.5%) effective tax rate. FY2024 ETR was 19%.128

3. ZT Systems discontinued operations: $66M income from sale of ZT Manufacturing in FY2025, not in continuing operations.127

4. Q4 2025 gross margin base: Q4 2025 non-GAAP GM 57% included the MI308 reserve release and initial China shipments. Management described Q1 2026 guide of 55% as a "clean" baseline.170

5. Segment reporting change: Client and Gaming combined into one segment effective Q1 2025; all prior periods retrospectively adjusted.140

Information Sources, Limitations, and Pending Items

Information Limitations:

• Segment-level gross margins not disclosed by AMD (consolidated only)

• No third-party peer benchmarking data available (semiconductor industry folder was empty)

• Q1 and Q2 2024 10-Q filings not locally available; FY2024 interim quarterly FCF not derivable

• FY2020 and earlier 10-K filings not locally available

• Forward FCF guidance for FY2026 not provided by management

• No goodwill impairment sensitivity analysis available in local data

Pending confirmations: AMD wiki has 0 published cards. No SSOT conflicts. The upstream reconciliation task's noted "conflict" on FY2025 non-GAAP EPS has been resolved: $4.17 is confirmed correct from the 8-K non-GAAP reconciliation table. The "~$8.00+" figure referenced in one preliminary task appears to be a misread of the long-term FAD EPS target (>$20), not an actual reported figure.

Risks & Bear Case

What are the key risks and bear case for AMD? Analyze Intel recovery, NVIDIA dominance in AI, China export controls, TSMC dependency, and other structural threats.

AMD Bear Case: Key Risks and Structural Threats

Executive Summary

AMD's bear case centers on four compounding structural exposures that each independently constrain the company's long-term revenue ceiling: (1) a roughly 12:1 revenue gap versus NVIDIA in AI accelerators, reinforced by a 20-year software moat that AMD has self-disclosed it cannot overcome quickly; (2) China export controls that have already generated an $800 million inventory charge and reduced China AI GPU revenue to "not material" in Q1 2026, against a baseline of 22.4% of total FY2025 revenue from China; (3) 100% dependency on TSMC for all leading-edge chips, with AMD's entire 2026–2028 CPU and GPU roadmap gated on TSMC N2 capacity that remains geographically concentrated in Taiwan; and (4) an Intel recovery that — while currently ASP-led rather than unit-share-led — is backed by a multi-year manufacturing roadmap and a September 2025 NVIDIA partnership that AMD formally flagged as a material competitive risk in its own 10-K.

Key Finding 1: NVIDIA's ~12:1 Revenue Lead in AI Accelerators Is Self-Disclosed by AMD as a Structural Threat

NVIDIA's FY2026 Compute & Networking (data center AI) revenue reached $193.5 billion, up 67% year-over-year, driven by Blackwell platform demand and a 142% increase in data center networking revenue171. AMD's Data Center segment full-year FY2025 revenue (10-K segment table, corrected) was $16.6 billion, up 32% year-over-year187.

Data reconciliation note: EB 101367 (from AMD FY2025 earnings PDF) states "$19.8 billion, +69% YoY." The upstream reconciliation task found this figure is inconsistent with AMD's 10-K segment table, which shows Data Center = $16,635 million ($16.6 billion), FY2024 = $12,579 million ($12.6 billion), implying +32% growth, not +69%. The $19.8 billion figure has no primary-source support. This analysis uses the 10-K filing figure as authoritative.

This produces a revenue ratio of approximately 12:1 in NVIDIA's favor. AMD does not separately disclose Instinct GPU-only revenue. Third-party estimates place AMD Instinct revenue at approximately $5–8 billion in FY2025 — not from AMD filings and should be treated as approximations only.

AMD's own annual report acknowledges the competitive dynamic directly172:

"Similarly, Nvidia Corporation (Nvidia) leverages its market position in data center GPU, financial resources, and proprietary software ecosystem to promote its systems and influences customers who do business with us. Our competitors' business practices, including allocation strategies, pricing actions, product mix and introduction schedules, licensing terms, marketing arrangements, product bundling strategies, lack of software inoperability and business acquisitions can limit customers' ability to choose alternative products, including ours."

NVIDIA's CUDA moat is quantified in its own filings: CUDA was introduced in 2006 (20 years ago), with NVIDIA investing over $76.7 billion in cumulative R&D since inception, supporting "hundreds of domain-specific software libraries, frameworks, SDKs, and APIs"173. Jensen Huang states explicitly: "NVIDIA is the only platform that runs every frontier AI model"174, with NVIDIA "growing share in inference... very, very quickly" and serving the "second category" of AI data centers (AI-native clouds, sovereign AI, enterprise, industrial) "almost uniquely"175.

NVIDIA Q1 FY2027 reported sovereign AI revenue up 80%+ YoY across nearly 40 countries, and swept all MLPerf inference benchmarks with Blackwell Ultra, delivering a 2.7x throughput increase and 60% cost-per-token reduction on GB300 in six months176.

ROCm vs. CUDA — AMD's own language implies a catch-up posture: AMD's CTO describes ROCm as "fully battle tested" in HPC and hyperscaler environments and frames AMD's open-source strategy as distinct from NVIDIA's "more proprietary walled garden solutions"179. AMD's Q1 2026 transcript describes "significantly accelerating" ROCm development cadence180 — language consistent with catching up. AMD's AI 2026 Investor Roundtable shows investor questions still centered on ROCm competitive gap181, and ROCm feature development timelines were described as "dramatically" reduced via AI tooling — implying prior timelines were long182.

Hyperscaler custom silicon creates a TAM ceiling: AMD formally acknowledges in its 10-K that "some of our customers are internally developing their own data center microprocessor products and accelerator products which could impact the available market for our products"177. Google (TPU v7 Ironwood), AWS (Trainium 3), Microsoft (Maia 200), and Meta (MTIA) are all building competing silicon — the same customers that underpin AMD's largest GPU deployment partnerships (Meta 6 GW, OpenAI, Anthropic)183.

Multi-period data center revenue comparison:

Note: AMD Data Center includes CPUs + GPUs + DPUs. NVIDIA Compute & Networking includes AI GPUs + networking. Ratios are approximate and compare non-identical product sets.

Key Finding 2: China Export Controls Have Imposed a Realized Revenue Disruption and Remain an Escalating Structural Overhang

Revenue exposure is material: AMD's China (including Hong Kong) revenue in FY2025 was $7.751 billion, representing 22.4% of total FY2025 revenue of $34.639 billion. In FY2024, China represented 24.2% ($6.231 billion of $25.785 billion)191.

Data reconciliation note: Multiple upstream sources describe China exposure as "~20%." AMD's 10-K FY2025 geographic table confirms 22.4%. The 10-K figure is used here as authoritative.

The sequence of restrictions and charges:

• October 2023: BIS issued ECCN 3A090/4A090 restrictions preventing shipment of certain AMD Instinct ICs and Versal FPGAs to China without a license195.

• April 2025: BIS tightened controls with a new license requirement for AMD Instinct MI308. AMD recorded approximately $800 million in inventory and related charges in Q2 2025178192.

• Q4 2025: AMD received some export licenses and shipped approximately $390 million of MI308 to China, reversing ~$360 million of prior inventory reserves. This inflated Q4 2025 gross margin to ~57%; underlying normalized gross margin was approximately 55%187.

• Q1 2026: China AI GPU revenue described as "not material." Data Center AI revenue was down modestly sequentially due to this transition190.

• February 2026: U.S. granted export licenses for AMD Instinct MI325 to certain China-based customers. Critical condition: each unit must undergo U.S. inspection before China export, making it subject to a 25% tariff on U.S. reimportation. China may not allow MI325 imports at all186.

• Ongoing: BIS announced intent to rescind the January 2025 AI Diffusion Rule in May 2025 and issue replacement rules — ongoing regulatory uncertainty for AMD's product design and compliance193. A September 2025 BIS rule expanding Entity List coverage to 50%-owned subsidiaries had enforcement suspended for one year in October 2025 but "can be reimposed at any time"194.

China export control charge timeline:

Second-order structural impact: AMD explicitly acknowledges that export controls "put us in a competitive disadvantage compared to domestic Chinese competitors" and "make it easier for our China-based competitors to develop and sell their own solutions"195. Chinese vendors (Huawei Ascend, etc.) face no equivalent restrictions within China.

AMD's assembly/test chain also carries China exposure: Nearly all AMD product assembly and final testing is performed in China (via Tongfu ATMP JV), Malaysia, and Taiwan188. AMD holds only a 15% equity interest in the ATMP JV with Tongfu Microelectronics (a Chinese company), purchasing $1.7 billion per year from these JVs in each of FY2023 and FY2024189. International sales were 67% of FY2025 net revenue196.

Key Finding 3: TSMC Dependency Deepens as AMD's Entire Leading-Edge Roadmap Moves to N2

AMD relies on TSMC exclusively for all wafers at 7nm and below. GF only covers 12nm/14nm legacy nodes. There is no alternative foundry capable of producing AMD's leading-edge CPU and GPU products197.

AMD's roadmap is now fully gated on TSMC N2:

• EPYC "Venice" (Zen 6) = the first tapeout in TSMC 2nm (confirmed by AMD CTO Mark Papermaster)198. Production ramp announced May 2026.

• MI400 series (CDNA 5) = AMD's next AI GPU generation, targeted for launch in 2026, built on TSMC advanced nodes198.

• TSMC CEO C.C. Wei confirms N2 entered HVM in Q4 2025 with "good yield," ramping at Hsinchu and Kaohsiung in Taiwan199. TSMC's policy is to "prioritize land in Taiwan to support the fast ramp of our newest node."

Capacity allocation risk: AMD has no long-term commitment contracts for many TSMC manufacturing services197. TSMC's top two unnamed customers command 19% and 17% of TSMC FY2025 revenue respectively200 — collectively 36% — leaving AMD competing for residual allocation. Note: the customer identity assignments (Apple/NVIDIA) are inferences from public data, not disclosed facts.

CoWoS packaging is chronically constrained: TSMC doubled CoWoS capacity in 2025 but CEO Wei stated demand "clearly feels like even that is not enough" — supply remains "very tight" into 2026201. AMD's MI300/MI350/MI400 systems are entirely CoWoS-dependent. MI400 doubles packaging interconnect density vs. MI300; Helios represents a 30x silicon density increase per AMD's FAD 2025198.

Geographic concentration at TSMC is extreme: TSMC's 20-F confirms 80% of non-current assets (NT$3.1 trillion of NT$3.9 trillion) are located in Taiwan200. Arizona holds approximately 14% and will not carry meaningful N2 capacity until 2027–2028. TSMC confirmed two earthquake events in 18 months: April 2024 (NT$3 billion net losses) and January 2025 (NT$5.3 billion net losses) — combined approximately $255 million in realized losses200.

TSMC dependency layer summary:

Key Finding 4: Intel Recovery Is ASP-Led — Unit Share Defense Currently Intact, but Structural Threat Is Building

Intel DCAI revenue increased 59% YoY in Q2 2026 to $6.3 billion. Decomposition: server ASPs +48% YoY, server volume +9% YoY184. Intel's recovery is driven by premium product mix, not yet by unit share recapture from AMD. On the PC side, Intel client revenue rose 27% YoY in Q2 2026, but client CPU volumes declined 8% YoY due to supply constraints185 — suggesting AMD may be maintaining or gaining PC unit share even as Intel's revenue recovers.

Structural medium-term risks from Intel's roadmap: Intel is accelerating a "four nodes in three years" strategy with 18A process output running 25% above internal targets and 50% QoQ in Q2 2026. Intel CEO Tan is explicitly re-introducing SMT (simultaneous multithreading) to Xeon CPUs to close AMD Zen's per-thread performance advantage. Intel signed a multi-year LTA with Google covering Xeon + IPU, and Intel Xeon 6 was selected as the host CPU in NVIDIA DGX Rubin NVL8 systems177.

AMD's own 10-K discloses Intel pricing pressure202:

"Intel Corporation (Intel) uses its microprocessor market position to price its products aggressively and target our customers and channel partners with special incentives. These aggressive activities have reduced and may reduce our unit sales and average selling prices for many of our products, adversely affecting our business."

Intel's current non-GAAP gross margin is approximately 41–42% vs. AMD's approximately 55–56%185 — Intel has structural room to price aggressively while remaining profitable.

The NVIDIA-Intel partnership is flagged in AMD's 10-K as a new material risk203:

"in September 2025, Nvidia announced a partnership and investment in Intel to partner on new data center and client platform products. This partnership may result in increased competition and pricing pressure for our products or could prevent us from participating in other opportunities, which could materially adversely impact our business, financial condition and margins."

Intel server recovery decomposition:

Key Finding 5: NVIDIA's Vera CPU Threatens AMD's Last Uncontested Segment

NVIDIA disclosed its Vera CPU as opening a $200 billion new TAM in agentic AI CPUs, with $20 billion in standalone Vera CPU FY2027 revenue visibility175. Every major hyperscaler and system maker is named as a Vera partner. This represents NVIDIA's direct entry into server CPUs — AMD's highest-share business segment — from a company with $215.9 billion in annual revenue171 and $76.7 billion in cumulative R&D invested173. AMD's 10-K identifies competition in Data Center from Intel (CPUs), NVIDIA (GPUs/DPUs), and Arm-based CPU vendors204, with Vera adding a GPU-vendor entry directly into the CPU segment.

Short-Term / One-Time Factors and Calibration Notes

The following are short-term or accounting-related and should be disaggregated from structural analysis:

• Q4 2025 gross margin ~57%: Approximately $360 million was MI308 inventory reserve reversal. Underlying normalized gross margin approximately 55%.

• Q1 2026 Data Center AI sequential decline: Driven by absence of China MI308 revenue vs. Q4 2025 license surge. Not indicative of underlying demand softening190.

• MI450/Helios ramp margin dilution: AMD CFO confirmed MI450 is dilutive to corporate average gross margins during ramp (beginning Q3 2026). Expected to persist through 2026–2027183.

• $800M vs. net ~$440M: The $800M Q2 2025 charge was partially reversed by ~$360M in Q4 2025. Net FY2025 impact approximately $440 million. The "$1.5–$1.8 billion total revenue loss" cited in some web sources is a third-party estimate, not a filing-disclosed figure.

AMD Bear Case Risk Matrix

Information Limitations and Pending Conflicts

Information limitations:

• AMD Q2 2026 actual results (released August 4, 2026 after market close) are not yet indexed — Q2 actual Data Center revenue, gross margin, and Q3 guidance remain unconfirmed from filings.

• AMD Instinct GPU-only revenue is not separately disclosed; third-party estimates ($5–8 billion FY2025) are approximations.

• TSMC capacity allocations by customer are not publicly disclosed.

• BIS replacement rules (successor to AI Diffusion Rule) have not been published.

• Helios/MI450 revenue contribution guidance for Q3/Q4 2026 has not been quantified in a specific dollar range.

Pending conflict — EB 101367: AMD FY2025 earnings PDF states Data Center revenue of "$19.8 billion, +69% YoY." AMD 10-K segment table states $16.6 billion, +32% YoY. The earnings PDF figure is unsupported by the primary filing. The 10-K figure is used as authoritative in this analysis.

Valuation & Margin of Safety

Value AMD using DCF, comparable company analysis, and scenario analysis. What is the margin of safety at current prices and what would be a fair entry price?

AMD Valuation: DCF, Comparable Company Analysis, and Scenario Analysis

Executive Summary

This analysis synthesizes AMD's FY2025 10-K, Q2 FY2026 10-Q, Q2 2026 earnings transcript, Financial Analyst Day (FAD) disclosures, and peer filings (NVDA, AVGO, INTC, QCOM) into a three-method valuation. At the current price of $482.05 (August 5, 2026), the probability-weighted intrinsic value under base assumptions converges near current levels (~$481), implying essentially zero classical margin of safety. A fair entry price for a 20-25% margin of safety falls in the $370-$420 range under base-to-bull assumptions. The stock is priced for near-perfect execution of the 2027 Data Center doubling thesis plus sustained WACC compression from the current ~17.5-18% to ~12-14% as AMD's risk profile matures.205207208

Key Finding 1: Data Center Is Now the Dominant Value Driver — and Growth Is Accelerating Beyond Prior Guidance

AMD's Data Center segment has crossed a structural inflection point: from $6.50 billion in FY2023 to $12.58 billion in FY2024 to $16.635 billion in FY2025 (+32% YoY), and then to $6.718 billion in Q2 2026 alone (+107% YoY) — a single quarter equaling nearly 40% of the full prior year.205206 Data Center revenue now represents 58% of AMD's total revenue (up from 42% a year prior).207

Lisa Su stated in the Q2 2026 earnings call:

"We now expect data center segment revenue to more than double year-over-year in 2027... we expect revenue to grow substantially above our prior target of greater than 35%, and we expect to significantly exceed our $20 annual EPS target within our strategic timeframe."208

She added that analyst estimates for 2027 Instinct AI revenue (~$30 billion) are "probably too low" and the appropriate framing is "well over 100%" Data Center growth.209210 Two structural elements underpin this: (1) contracted multi-year hyperscaler volumes — Anthropic (2 GW MI450 deployment, first 1 GW beginning H1 2027) and OpenAI (multi-generation GW-scale partnership), plus existing Microsoft and Meta deployments;211 and (2) EPYC server CPU momentum with H2 2026 server CPU growth guided at >80% YoY, accelerating to >70% YoY full-year 2027.208

Data Center Revenue Trajectory:

Sources: AMD 10-K FY2025205206; AMD Q2 2026 transcript207208209.

Key Finding 2: Margin Expansion Is Structural and Ongoing

AMD's non-GAAP gross margin rose from ~52% in FY2025 to 56% in Q2 FY2026 (guided again at 56% for Q3 2026).212213 This 400 bps expansion in approximately two quarters reflects the mix shift toward Data Center and the clearing of the $800 million MI308 export-control inventory charge that depressed Q2 2025 comparables.214 GAAP gross margin follows at 53.8% in Q2 2026 (vs. 49.5% FY2025).205

Operating leverage is compounding the gross margin benefit: non-GAAP operating income reached $3.09 billion in Q2 2026 on $11.536 billion revenue — a 27% non-GAAP operating margin, with Q3 implied at ~28% (revenue ~$13 billion, GM 56%, OpEx ~$3.65 billion).213 GAAP operating margin expanded from 1.8% in FY2023 to 10.7% in FY2025 to 15.9% in H1 2026.205212 CFO Jean Hu committed to growing OpEx slower than revenue, preserving this operating leverage.213

Margin Progression:

Note: The GAAP/non-GAAP gap (~$3.9 billion annually in FY2025, comprising $2.254 billion acquisition amortization and $1.638 billion SBC)206 must be treated consistently within any valuation framework. FY2025 GAAP FCF was ~$5.5 billion ($6.49B OCF minus $974M CapEx)205; non-GAAP FCF (adding back SBC) is approximately $7.1 billion.

Note on FY2024 non-GAAP GM: One upstream table showed "~46%." Per primary source cross-check, the correct FY2024 non-GAAP GM is ~53%; the 46% figure appears to be a dirty data artifact (possible GAAP/period confusion).

Key Finding 3: DCF — The $482 Price Embeds a WACC of ~12%, Not the Current Market ~17.5-18%

WACC Construction (August 2026):

⚠️ The "14.8% normalized WACC" used in sensitivity tables assumes AMD's beta de-rates from 2.49 toward ~2.0 as the business matures — this is a forward assumption, not the current market WACC. The two must not be conflated.215

CapEx acceleration is material: H1 2026 CapEx was $1.197 billion vs. $494 million in H1 2025 (+142% YoY), with construction in progress rising from $508 million to $1.148 billion — reflecting Helios production investment.217 This reduces near-term FCF but is consistent with the contracted demand ramp.

FCFF Projections (Base Case):

DCF Sensitivity — Equity Value Per Share (Base Revenue Scenario):

The current price of $482 is reached in the base revenue scenario only at WACC ~12% and terminal growth ~5%, or in the bull revenue scenario at WACC ~14-15%. At the current market WACC of ~17.5-18%, the base case DCF yields approximately $118-$150 per share.

Key Finding 4: Comparable Company Analysis — AMD Trades at a Substantial Premium on FY2026E Multiples, Near Parity on 2027 Bull Estimates

Peer Valuation Table (August 5, 2026):

Sources: NVDA Q1 FY2027 10-Q and transcript218219; AVGO Q2 FY2026 transcript and 10-Q220221222; INTC Q2 2026 earnings223; QCOM Q3 FY2026 transcript224225; web research (StockAnalysis)215.

AMD's 43.6x forward P/E is 2.0x NVDA's multiple and 1.6x AVGO's, yet AMD's gross margin (56%) is 19 percentage points below NVDA (74.9% GAAP in Q1 FY2027)218219 and 14 points below AVGO (69.5% in Q2 FY2026)222. NVDA generated $75 billion in Data Center revenue in a single quarter with $49 billion FCF219; AVGO guided FY2027 AI semiconductor revenue exceeding $100 billion with Q2 bookings alone over $30 billion.221 These represent the scale and margin benchmarks AMD is targeting.

Comps-Implied Value Range:

The current price is supported only by: (a) AMD maintaining its current elevated multiple (43-44x) on FY2027 base-case earnings, or (b) a bull-case EPS scenario ($14-15) with a moderate premium multiple (~35x).215

Key Finding 5: Scenario Analysis and Probability-Weighted Value

Three Scenario Operating Cases:

Bear case triggers: Export control recurrence (AMD took an $800 million MI308 China charge in Q2 2025, demonstrating precedent);214 NVDA competitive entrenchment limiting AMD Data Center AI share gains; gaming revenue -31% YoY continues;214 PC market softness in H2 2026.214

Bull case triggers: Anthropic 1 GW MI450 deploying H1 2027 as contracted;211 OpenAI multi-year GW-scale ramp continues;211 Lisa Su's statement that 2027 DC AI estimates are "probably too low" proves correct;209 AMD beta de-rates toward 1.5-1.8 as revenue mix stabilizes.

Probability-Weighted Intrinsic Value:

At these weights, AMD is essentially fairly valued at $482. Street consensus of 51 analysts (Strong Buy, average PT $595.64) implies the Street assigns higher probability to the bull case and/or a higher bull price target.215

Key Finding 6: Margin of Safety and Fair Entry Price

Margin of Safety at $482:

Fair Entry Price Framework:

Short-Term / One-Time Factors and Basis Notes

The following factors are short-term or one-time in nature and should be separated from long-term signal:

1. Q2 2025 MI308 charge ($800 million): Inflated Q2 2026 YoY comparables (non-GAAP EPS +246% YoY). Comparables normalize from Q3 2026 onward.214

2. Gaming revenue -31% YoY in Q2 2026: Driven by lower semi-custom sales. Annual FY2025 gaming was +51% YoY off a depressed FY2024 base ($2.595B), but still -37% vs. the FY2023 peak of $6.212 billion — annual and quarterly figures must not be mixed.206214

3. CapEx step-up: H1 2026 CapEx $1.197 billion (+142% YoY) reflects Helios production investment, compressing near-term FCF.217

4. PC market softness: Management guided for a softer PC market in H2 2026, affecting the Client segment.214

5. Net cash figure: Three figures were in circulation across tasks — $7.25B (FY2025 year-end), $8.84B (stale web snapshot), and $9.9B (Q2 2026 10-Q, canonical).216 All equity bridge calculations use $9.9B.

6. FY2024 non-GAAP gross margin: One upstream table showed "~46%"; the correct FY2024 non-GAAP GM is ~53% per primary source cross-check. The 46% figure is a dirty data artifact.

Information Sources, Limitations, and Pending Conflicts

Information gaps:

• FY2027-2028 sell-side consensus EPS/FCF estimates not available from local files; estimated from management guidance

• Current 10-year Treasury yield used at estimate (~4.3-4.4%); not pulled from live market data

• AMD beta 2.49 from web sources — periodicity (2-year weekly vs. 5-year monthly) not verified

• ARM Holdings excluded from peer comps (not in local data sources)

• No sell-side DCF model available locally for terminal assumption cross-check

Pending confirmed conflicts:

• CONFLICT-5 (WACC): Normalized "14.8% base WACC" vs. current market WACC ~17.5-18% — both used with explicit labeling throughout

• CONFLICT-3 (FY2024 non-GAAP GM): "~46%" corrected to "~53%"; flagged in Basis Notes above