AMD — Advanced Micro Devices
Identity
| Field | Value |
|---|---|
| Name | Advanced Micro Devices, Inc. |
| Ticker | AMD |
| Exchange | NASDAQ |
| Sector | Semiconductor — Foundry, IDM & Design |
| HQ | Santa Clara, California |
| CEO | Lisa Su |
| Market Cap | ~$314B (March 2026) |
| Business Model | Fabless semiconductor designer (TSMC manufactures) |
Thesis
Core thesis: AMD is a structural share gainer in two of the most important silicon markets on earth -- server CPUs (EPYC eating Intel alive) and AI GPUs (MI series as the only credible NVIDIA alternative). Lisa Su executed one of the greatest corporate turnarounds in semiconductor history, taking the company from near-bankruptcy at $3/share to $300B+. The chiplet architecture gives AMD a 2-3 year design lead in multi-die packaging.
The agentic CPU angle -- this is the overlooked part. Agentic AI workloads (autonomous systems that plan, decide, and act through continuous perception-reasoning-action loops) are fundamentally CPU-intensive in ways traditional inference is not. The Georgia Tech/Intel research paper (arXiv:2511.00739) showed tool processing on CPUs accounts for 50-90% of total latency in agentic workloads. The GPU literally sits idle waiting for the CPU to finish. This means:
- CPU:GPU ratios in AI clusters need to shift from 1:8 toward 1:1 or higher
- OpenAI's AWS partnership implies a 100:1 CPU:GPU ratio at agentic scale ("tens of millions of CPUs" vs. "hundreds of thousands of GPUs")
- Futurum projects CPU demand growth of 34.9% by 2029 -- outpacing GPUs
AMD is the best-positioned x86 vendor for this shift. Turin (192 cores) now, Venice (256 cores, 2nm, 1.6 TB/s memory bandwidth) in H2 2026. Intel is losing share and supply-constrained. Lisa Su at Morgan Stanley (March 2026): CPU demand "far exceeded my expectations." Supply is nearly fully allocated for FY2026.
Single most important thing that must go right: AI GPU revenue must keep growing -- MI400/MI450 needs meaningful adoption in training, not just inference. On the CPU side, the agentic thesis needs agent adoption to actually scale (Gartner warns 40% of agent projects may be canceled by 2027).
Expected holding period: 2-3 years.
Exit criteria: Sell if AI GPU revenue growth stalls for two consecutive quarters, if NVIDIA's CUDA moat proves insurmountable (AMD share stays below 10%), or at $350+ (80% upside from March 2026 entry).
Business
What They Do
AMD designs high-performance computing chips across four segments:
| Segment | FY2025 Revenue Share | What It Is |
|---|---|---|
| Data Center | 48% (~$16.6B) | EPYC server CPUs + Instinct AI GPUs + Pensando NICs |
| Client | ~25% | Ryzen desktop/laptop CPUs |
| Gaming | ~17% | Radeon GPUs + semi-custom console chips (PS5, Xbox) |
| Embedded | ~10% | Xilinx-derived FPGAs and adaptive SoCs |
Data Center is where the action is. It grew 32% in FY2025 to a record $16.6B and is inflecting hard. Everything else is fine but not the story.
EPYC Server CPUs -- The CPU Renaissance
5th Gen "Turin" (current): 192 cores, 3nm/4nm, 150B transistors, up to 5GHz. Over 1,600 cloud instances (50% YoY increase). Now >50% of total server revenue. AMD has ~34% server CPU revenue share (41% in Q2 2025), approaching Intel parity.
6th Gen "Venice" (H2 2026): 256 cores on the world's first 2nm server CPU. 1.6 TB/s memory bandwidth per socket (2.5x Turin). Doubles CPU-to-GPU bandwidth. PCIe Gen6. This chip is purpose-built for the agentic world -- massive core counts for parallel agent execution, extreme memory bandwidth for context management, lower inter-chiplet latency.
Why CPUs matter for agentic AI: In agentic workflows, the CPU handles tool calling, API requests, JSON parsing, orchestration, RAG/context management, and code execution. These tasks account for 50-90% of total latency. The GPU generates tokens; the CPU does everything else. AMD's EPYC line has higher per-core performance than anything Intel ships, more cores than ARM alternatives, and a massive x86 software ecosystem where every agentic framework (LangChain, AutoGen, CrewAI) is already optimized.
AI GPUs -- The NVIDIA Chase
MI300X/MI325X are the current generation, deployed at Meta, Microsoft, Oracle (50,000 GPU cluster). MI350 (mid-2026) brings CDNA 4 architecture. MI400/MI450 (H2 2026) with HBM4, 432GB capacity, 19.6 TB/s bandwidth is the next major catalyst.
AMD is the "second source" for hyperscale AI. They won't dethrone NVIDIA, but they don't need to. Even 15-20% of the AI accelerator market justifies a significantly higher stock price.
Full-Stack Platform
AMD is the only vendor offering CPU + GPU + NIC + rack-scale:
- EPYC CPUs -- orchestration, tool calling, context management
- Instinct GPUs (MI350, MI450/MI455X) -- inference and training
- Pensando NICs (Pollara 400) -- secure, low-latency networking
- Helios rack-scale platform -- 72 MI450 GPUs + 18 Venice CPUs per rack, 2.9 exaFLOPS FP4
Competitive Moat -- Moderate
(+) Chiplet architecture leadership -- 2-3 year lead in multi-die design. (+) Deep TSMC partnership with early access to 2nm. (+) x86 duopoly with Intel -- decades of software compatibility. (+) Hyperscaler "second source" lock -- Microsoft, Meta, Oracle all deploying AMD GPUs. (-) NVIDIA's CUDA ecosystem is the widest moat in AI. Switching costs for training are enormous. (-) Fabless model means no manufacturing differentiation.
Management
Lisa Su is one of the best CEOs in semiconductors, full stop. She took AMD from near-bankruptcy to a $300B+ company. MIT PhD in electrical engineering. Track record of on-time execution across Zen architecture, chiplets, EPYC, and MI300. Named to TIME 100 Most Influential People. Holds ~$765M in AMD stock -- she eats what she cooks.
Key quotes from March 2026 (Morgan Stanley TMT Conference):
- "The CPU portion of the business has actually far exceeded my expectations in terms of demand."
- Top customers told her CPU demand alongside AI was "under-forecasted."
- Expects server CPU market to grow "strong double digits" in 2026.
- Projects data center segment revenue growth of >60% annually over 3-5 years.
- AI business to reach "tens of billions in annual revenue" by 2027.
No significant governance red flags. The management bench is deep with experienced semiconductor operators.
Financials
Key Metrics (FY2025 / Q4 2025)
| Metric | Value | Notes |
|---|---|---|
| Revenue (FY2025) | $34.6B | +34% YoY, record |
| Revenue (3-yr CAGR) | ~18.4% | Strong secular growth |
| Gross margin (non-GAAP) | 52% FY / 57% Q4 | Expanding -- mix shift to Data Center |
| Operating margin (non-GAAP) | ~22% FY / 28% Q4 | Strong operating leverage |
| Free cash flow (Q4) | $2.1B | Record, inflecting sharply |
| Cash | $10.6B | Up 46% QoQ |
| Long-term debt | $2.35B | |
| Net cash | ~$7.4B | Fortress balance sheet |
| Debt/equity | 5.3% | Minimal leverage |
| ROIC | 6-9% | Below WACC (~15%) -- partly Xilinx goodwill artifact |
Red flag worth watching: ROIC of 6-9% below WACC. This is partly an artifact of the Xilinx acquisition inflating invested capital with $50B+ of goodwill. On a cash-on-cash basis, returns are better. But AMD needs to grow into its capital base -- monitor this.
Valuation (as of March 2026, $192.43)
| Metric | Value |
|---|---|
| Trailing P/E (GAAP) | ~64-76x |
| Forward P/E (FY2026E) | ~30x |
| EV/EBITDA | ~48x |
| P/FCF | ~48x |
Context: EV/EBITDA has ranged from 22x (2022 trough) to 66x (2023 AI hype peak). Current 48x is mid-to-upper range. At 30x forward P/E, the market expects ~25-30% earnings growth for several years. Achievable given AI GPU ramp + server CPU share gains, but requires flawless execution. NVIDIA at 30-35x forward with 80%+ GPU share is the comp -- AMD at 30x with 8% share means you're betting on share expansion.
Growth Trajectory
- Lisa Su's 3-5 year targets: 35% annual revenue growth, data center segment >60% annual growth
- AI business: "Tens of billions in annual revenue" by 2027
- Server CPU: approaching Intel parity in 2026
- Server CPU supply nearly fully allocated for FY2026
Catalysts & Risks
Bull Case
-
Structural CPU demand shift. Agentic AI changes CPU:GPU ratios from 1:8 toward 1:1 or higher. Not cyclical -- architectural. The server CPU market was considered mature/ex-growth before this.
-
Best-positioned x86 vendor. Venice at 256 cores, 2nm, 1.6 TB/s bandwidth. Intel is losing share and supply-constrained. ARM is at 15% datacenter share, well below targets.
-
TAM expansion is enormous. CPU demand growing 34.9% through 2029 (Futurum). AMD share gains on top of a growing market = powerful double tailwind. AI data center market: $1T by 2030 (Lisa Su estimate). Agentic AI TAM: $155B by 2030 (BofA).
-
Pricing power. Supply shortages + demand surge = higher ASPs. Revenue share (41%) already exceeds unit share (27%), showing premium pricing that should expand.
-
CPU de-risks the GPU story. Even if Instinct GPU underperforms vs. NVIDIA, EPYC alone drives "strong double digit" data center revenue growth. Two engines, not one.
-
Full-stack advantage. Only vendor with CPU + GPU + NIC + rack-scale platform (Helios). Integrated optimization heterogeneous setups can't match.
-
Lisa Su confirms it in real-time. "Far exceeded my expectations." Supply fully allocated for 2026.
Bear Case
-
NVIDIA Vera CPU. The biggest competitive risk. Vera is purpose-built for agentic processing (88 cores, 3.4 TB/s bisection bandwidth, standalone deployment). Meta already committed. NVIDIA can bundle with GPU ecosystem. But Vera is unproven and not yet shipping.
-
ARM competition. Hyperscalers building custom ARM CPUs they fully control (Graviton, Axion, Cobalt). If agentic workloads run fine on ARM, hyperscalers route demand to their own silicon. ARM's 15% share could grow.
-
NVIDIA CUDA dominance. The widest moat in AI. ROCm is improving but still breaks on custom kernels and cutting-edge training. If customers can't port training workloads, AMD's GPU TAM is capped at inference.
-
China export risk. $800M charge already taken, ~$1.5B annual revenue at risk from MI308 curbs. Broader worldwide controls reportedly under consideration.
-
Agentic AI may disappoint. Gartner: 40% of agent projects canceled by 2027 due to infrastructure cost overruns. If agent adoption stalls, the CPU demand thesis weakens.
-
TSMC capacity constraints. TSMC prioritizes AI accelerators over CPUs. If not enough 2nm capacity for Venice, AMD is supply-constrained at peak demand.
-
Valuation already reflects it. 48x EV/EBITDA. Any stumble gets punished -- see the -17% post-earnings drop (worst since 2017) on a mild Q1 guide miss.
Near-Term Supply Catalyst
Server CPUs are in shortage as of March 2026:
- AMD: 8-10 week delivery lead times for select EPYC
- Intel: up to 6-month lead times, reallocating PC CPU capacity to servers
- TSMC prioritizing accelerators over CPUs, creating supply tension
- 2026 defined by synchronized shortages across compute and accelerator segments
Hyperscaler CapEx (2026 Context)
| Company | 2026 CapEx Estimate |
|---|---|
| Amazon (AWS) | ~$200B |
| Alphabet (Google) | $175-185B |
| Meta | $115-135B |
| Microsoft | ~$100B+ |
| Total Big 5 | $660-690B |
~75% (~$450B) directly tied to AI infrastructure. Significant CPU buildout within this for agentic workload preparation.
Decision Log
2026-05-12 PM | REVISED Post-Q1 2026 PT (re-anchored to spot $458) | see amd-q1-2026-synthesis
Spot live-pulled $458.79. Forward P/E 36x. Mkt cap $748B. Earlier same-day entry's PT framework was anchored to stale $192; that math is superseded.
Revised PTs.
- 12M PT: $400 (32x FY27 EPS $9.5; spot is $58 above PT — market got ahead of FY27)
- 5yr PT: $600 (= Tarot T4 = Lisa's $20 × 30x; +31% / +5.6% CAGR — market-like return)
- 5yr stretch: $750–900 (analyst-day bottom-up: $25–31 EPS × 28–30x)
RSU mechanics corrected (from Pink's prior claude.ai chat, since I had it wrong this AM):
- Baseline = 60-day avg ending Mar-15-2026 ≈ $203 (not $200; Lisa 10b5-1 sales at $197-203 confirm)
- Vesting = 90-day trailing average must reach hurdle, sustained — not single-day touch
- T1/T2/T3 fixed dollar values rounded to $5: ~$325 / $410 / $455; T4 = $600 absolute
- At spot $458 with 90-day avg only $242: tranches NOT yet vested. Trajectory: T1 ~Jun-Jul, T2 ~Aug-Sep if AMD holds. T4 needs $140+ more spot first
The honest read at $458. From $192 Mar 9 entry to $458 May 12 = +139% in 9 weeks. The post-Q1 rerate has happened. Forward returns are now market-like (5yr base CAGR ~1.7%, bull ~9%, stretch ~15%).
Action.
- Trim 25–33% here at $458 if conviction is "Lisa hits $20 EPS, not the $31 stretch" — locks post-Q1 windfall, keeps remainder as T4 option
- Hold full only if conviction = analyst-day bottom-up ($31 EPS, $900 PT)
- Do NOT add at $458 — wait for pullback to $350–380 or Q2 confirmation early August
- Mental stop tightened to $300 (50% retrace of post-Q1 rally), not $145
2026-05-12 AM | (SUPERSEDED) Initial Post-Q1 Read
Earlier entry anchored PT framework to wiki's $192 entry without pulling live spot. Replaced by PM revision above after claude.ai chat reconciliation surfaced RSU mechanic corrections and live spot at $458.
2026-03-09 | Pre-Buy Analysis | $192.43
Verdict: BUY. Start 2-3% position at $190-195. Add 1-2% on either pullback to $175-180 (200-day MA support) or Q1 earnings confirmation. Target 4-5% total position. Max loss tolerance 25% (mental stop at $145).
Rationale: Best combination of business quality, growth trajectory, and entry timing. Stock 28% below 52-week high ($267) after post-earnings selloff. Trading between 200-day MA ($179 support) and 50-day MA ($221 resistance) at ~30x forward earnings. Lisa Su is an elite operator, Data Center inflecting, MI400/MI450 in H2 2026. Even permanent #2 status in AI GPUs at 15-20% share justifies significantly higher price.
Conviction: Medium-high. Excellent business, reasonable correction entry, but NVIDIA competitive risk is real.
Behavioral check: Narrative seduction risk on "AMD is the next NVIDIA." CUDA moat is real and AMD has only 8% AI GPU share. The correction from $267 reduces FOMO risk -- this is a pullback entry, not a chase.
Downside scenario: AI GPU revenue disappoints, stock de-rates to 35x forward on $5 EPS (vs. $6.50 consensus) = ~$175 fair value. Deeper bear (AI bubble, 20x multiple) implies $100.
Key Data Points
| Metric | Value | Source |
|---|---|---|
| CPU share of agentic latency | 50-90% | Georgia Tech/Intel paper |
| AMD server CPU revenue share | 41% (Q2 2025) | Mercury Research |
| AMD server CPU unit share | 27.3% (Q2 2025) | Mercury Research |
| Venice memory bandwidth | 1.6 TB/s per socket | AMD |
| Venice core count | Up to 256 | AMD |
| CPU demand growth projection | 34.9% by 2029 | Futurum |
| Hyperscaler CapEx 2026 | ~$660-690B (Big 5) | Multiple sources |
| Agentic AI TAM 2030 | $155B | BofA Global Research |
| AI data center market 2030 | $1 trillion | Lisa Su |
| AMD 2025 revenue | $34.6B (+34% YoY) | AMD earnings |
| AMD data center growth target | >60% annually, 3-5 years | Lisa Su |
| OpenAI CPU scale target | "Tens of millions" | AWS-OpenAI partnership |
| Agent scale projection | 50-100B agents (2026) | Industry estimates |
Sources
- A CPU-Centric Perspective on Agentic AI (Georgia Tech/Intel) -- the foundational research paper
- CPUs Are Back: The Datacenter CPU Landscape in 2026 -- SemiAnalysis
- The Forgotten Chip: CPUs the New Bottleneck -- UncoverAlpha
- Can the CPU Market Meet Agentic AI Demand? -- Futurum Group
- AMD Q4 2025 Earnings Call Transcript
- AMD at Morgan Stanley TMT Conference March 2026
- NVIDIA Vera vs AMD EPYC -- Enertuition
- AMD Delivering Open Rack Scale AI for Agentic AI
- AMD EPYC Venice -- Tom's Hardware
- AI Capex 2026: The $690B Infrastructure Sprint -- Futurum
- Lisa Su 35% Annual Growth -- CNBC
- Data Center CPU Dominance Shifting -- Semi Engineering
Topics
- ai-infrastructure
- supply-chain-security
Source updates (auto-maintained)
Beyond The Hype (May 11, 26) - Beyond The Hype - Looking Past Management & Wall Street Hype
The article flags AMD as a company whose stock "does not adequately reflect the major design Helios design wins" due to circular/vendor financing concerns, noting AMD issued stock to finance AI buildouts and faces the same market skepticism as NVIDIA and Broadcom.
Relevant to your thesis: Reinforces the bear case on valuation — market discounting AMD's Helios wins due to vendor-financing optics, which may explain why the stock trades below fundamental value even as the CPU and GPU design win pipeline strengthens.
Source: https://enertuition.substack.com/p/competitive-risk-starting-to-materialize
Beyond The Hype (May 11, 26) - Beyond The Hype - Looking Past Management & Wall Street Hype
The article focuses on DRAM/memory market dynamics (Micron, Samsung, Hynix) and notes AMD's MI4xx/MI5xx as one beneficiary of Rubin delays, but contains no substantive analysis of AMD's CPU or GPU business itself.
Relevant to your thesis: Rubin execution delays cited as a near-term demand tailwind for AMD MI4xx/MI5xx — partially supports the bear case that NVIDIA competitive pressure is timing-dependent, not just structural.
Source: https://enertuition.substack.com/p/an-update-on-the-memory-market
Drop/z Misc (Mar 17, 26) - 2026 Grants Digest Volume 5 - by ZG - Tarot Capital
Lisa Su received a special $75M CEO Value Creation Equity Award in performance-based RSUs on March 15, 2026, with four price hurdles over five years — the highest tranche requiring a $600 stock price (~3x from ~$200), and the lowest implying ~60% upside; vesting requires continuous tenure through year five with nothing before year three.
Relevant to your thesis: Reinforces the management conviction signal — Su took the grant 28% below highs, with the $600 tranche roughly consistent with Lisa Su's own 35% revenue CAGR targets.
Source: dropfile://z Misc/2026 Grants Digest Volume 5 - by ZG - Tarot Capital.pdf