Advanced Micro Devices
Fabless: AMD designs the silicon and outsources fabrication to foundries (primarily TSMC), selling chips to hyperscalers, OEMs, and channel, plus semi-custom royalties (game consoles). Increasingly sells rack-scale AI systems (Helios) after the 2025 ZT Systems acquisition.
The thesis on this name
State of AI Compute
Asymmetric optionality on the opening bare-GPU-inference market via MI400/MI450 + the Meta up-to-6GW win — but ~55% GM vs Nvidia's ~75% proves it captures VOLUME not PRICING; a watch-to-long on MI400 volume proof and a second mega-design-win, not yet a core position.
State of Nvidia
Credible #2 on memory-bound inference (432GB HBM4, 19.6TB/s), ~2.5x slower/GPU vs Rubin, not yet a scale-up-fabric peer. The NVLink challen…
Earnings, margins, COGS & capex
FY2025 revenue grew 34% to a record $34.6B as Data Center (+32% to $16.6B) and Client & Gaming (+51% to $14.6B) more than offset flat Embedded ($3.5B, −3%). Q1 FY26 accelerated to +38% ($10.3B) on a 57% Data Center surge ($5.8B) from EPYC server CPUs and Instinct MI350 GPUs, with non-GAAP gross margin reaching 55%. The structural story: AMD captures VOLUME (share gains, two 6GW AI mega-deals) but not Nvidia's PRICING — its ~55% gross margin sits ~20 points below Nvidia's ~75%, proving it competes on price/TCO, not premium. FCF hit a record $2.6B in Q1 FY26 (tripled YoY).
Income statement — where each revenue dollar goes
% of revenueOf every $1 of revenue, ~50¢ is cost of goods and ~39¢ operating expense, leaving ~11¢ of operating profit (~12¢ net).
Revenue trend
Margins
up
up
up
up
up
COGS structure
COGS is dominated by foundry wafer costs (TSMC, including 2nm/N2 for MI400 compute chiplets), HBM memory (HBM3E today, HBM4 on MI400 — a major and rising cost line bought from SK Hynix/Samsung/Micron), CoWoS/advanced packaging, and substrates. Mix is the swing factor: rich Data Center mix lifts margin, while the low-margin nature of bare-GPU-inference wins (vs Nvidia's full-stack pricing) and export-control inventory charges have weighed on it. Memory/component cost inflation is now pressuring the Client & Gaming side.
Capex
FY2025 capex was a record ~$974M (~2.8% of revenue) — low because AMD is fabless. It funds R&D test/lab equipment, data-center engineering (ZT Systems rack integration), and IT, not fabs. The heavy fixed-cost burden (fabs, capacity) is borne by TSMC and the memory vendors; AMD's true 'capacity capex' is its TSMC wafer and HBM prepayments/commitments embedded in purchase obligations.
Latest earnings
Beat — revenue $10.3B vs ~$9.84B consensus and above own guide ($9.8B ±$300M); non-GAAP EPS $1.37 vs ~$1.29 consensus.
Q2 FY26 revenue ~$11.2B ±$300M (~+46% YoY, +9% QoQ), well above ~$10.5B consensus; non-GAAP gross margin ~56%. Management flagged Client/Gaming softness in 2H26 from rising memory/component costs.
- Data Center revenue
- $5.8B, +57% YoY (beat ~$5.56B consensus)
- Client & Gaming revenue
- $3.6B, +23% YoY
- Embedded revenue
- $868M, +6% YoY
- Non-GAAP gross margin
- 55% (+170bps YoY)
Growth drivers
- Instinct AI GPU ramp — MI350 shipping now; MI400/MI450 + Helios rack-scale on track for 2H 2026 — the inflection for the two 6GW deals
- Two anchor AI mega-deals — Meta up-to-6GW (~$60B/5yr) and OpenAI 6GW, ~12GW total committed, each with up to 160M-share warrants
- EPYC server-CPU share gains — AMD raised its 2030 server-CPU TAM view to $120B (35%+ CAGR) and targets $100B+ data-center revenue by 2030
- Client (Ryzen) share gains in desktop/notebook/commercial PCs amid AI-PC refresh
- Embedded (FPGA/adaptive SoC, ex-Xilinx) recovery as industrial/comms demand normalizes
- Sovereign-AI and second-source demand as buyers seek a Nvidia alternative for inference
Reported financials — SEC EDGAR
Audited GAAP figures pulled from SEC filings · latest filing 2026-02-04. The audited primary-source spine — not financial advice.
Revenue — annual (GAAP)
Margins & balance sheet — FY’25
Bull & bear
AMD is the asymmetric call option on the opening merchant AI-inference market: two 6GW mega-deals (Meta, OpenAI) plus the MI400/Helios ramp in 2H26 give it ~12GW of committed demand and a path to $100B+ data-center revenue by 2030, while EPYC keeps compounding server-CPU share — all on a fabless, net-cash, FCF-rich model.
- Second-source inflection: as the AI buildout shifts toward inference, buyers want a Nvidia alternative — AMD is the only credible one, and Meta+OpenAI's ~$60B+ commitments prove it
- MI400/MI450 + Helios rack-scale on track for 2H 2026 is the catalyst that turns committed GW into revenue and lifts data-center mix/margin
- EPYC structural share gains: management raised the server-CPU TAM to $120B by 2030 (35%+ CAGR); CPU franchise alone funds the GPU push
- Financial quality: ~$7.3B net cash, record $6.7B FY25 FCF, GM inflecting 50%→56%, ~2.8% capex intensity — high-quality compounder, not a cash-burner
- Margin upside if MI400 mix and software (ROCm) let AMD price closer to value rather than pure TCO discount
AMD is a watch-to-long, not a core position: the ~55% gross margin vs Nvidia's ~75% proves it captures VOLUME not PRICING, the MI400 volume and a second non-anchor mega-design-win are still unproven, valuation already prices triple-digit growth, and the Meta/OpenAI warrants dilute up to ~20% — so it must execute flawlessly into a Nvidia that isn't slowing.
- Pricing power gap is structural: ~55% GM vs ~75% means AMD is a TCO discounter; bare-GPU inference is the lowest-margin slice of the market
- Concentration + dilution: the bull case leans on two customers (Meta, OpenAI) whose warrants total up to ~320M shares (~20% dilution at full vest); a slip at either is a thesis breaker
- Software moat deficit: CUDA lock-in keeps switching costs high and caps AMD's share and price; ROCm parity remains aspirational
- Valuation leaves no slack: ~60–69x forward earnings already embeds the 80%+ AI-revenue CAGR — MI400 must ship at volume on schedule
- Nvidia is not standing still (Rubin), hyperscaler custom ASICs attack the same inference budget, and HBM4/TSMC-N2 supply could gate the ramp
What it is worth
Forward-earnings multiple cross-checked with a reverse-DCF on the 2030 data-center target. AMD trades at ~60–69x forward EPS (June 2026) on a ~$848B cap; consensus PT ~$500 (Strong Buy, 51 analysts). Peer anchor: Nvidia at ~75% GM commands a premium franchise; AMD's ~55% GM justifies a discount on margin but a growth premium on the second-source optionality.
~$320 (Street low
MI400 slips or disappoints on volume/margin, warrant dilution + Nvidia pace compress the multiple back toward the low estimate)
~$500–530 (consensus ~$500–503
MI400 ramps roughly on plan, margin grinds higher, Client softness offsets near-term)
~$650–670 (MI400 ships at volume on schedule, GM pushes past 56–58%, a third anchor win lands; matches the Street's high estimate)
At ~$532 the stock already prices the 80%+ AI-revenue CAGR and a clean MI400 ramp — reverse-DCF implies AMD must hit roughly its $100B+ 2030 data-center target with margin expansion to justify today's multiple. Watch-to-long: re-rates on MI400 volume proof + a third mega-win; de-rates on any ramp slip or margin disappointment.
SWOT
Strengths
- Only credible full-stack x86+GPU alternative to Nvidia — EPYC is gaining server-CPU share fast and Instinct is the #2 merchant AI GPU
- Two 6GW anchor AI customers (Meta, OpenAI) plus hyperscaler breadth validate the MI400 roadmap commercially
- Fortress balance sheet: ~$7.3B net cash, $6.7B FY25 FCF, ~2.8% capex intensity — fabless asset-light economics
- Leading-edge access: secured TSMC N2 (2nm) capacity for MI400 compute chiplets and dual-source HBM
- Margin and FCF inflecting up (GM 50%→55% non-GAAP, FCF tripled YoY in Q1 FY26)
Weaknesses
- ~55% gross margin vs Nvidia's ~75% — AMD wins volume/TCO, not pricing power; bare-GPU-inference is structurally lower-margin
- Software moat gap — ROCm still trails CUDA's 15+ year ecosystem, raising switching friction and limiting premium pricing
- Single-foundry dependence on TSMC for leading-edge; no in-house fabs
- AI GPU share still only ~5–7% vs Nvidia ~80%+ — the gap in absolute scale is widening, not narrowing
- Client & Gaming exposed to PC cyclicality and rising memory/component costs (guided weaker 2H26); Embedded flat
Opportunities
- Bare-GPU / merchant-silicon inference market opening as buyers demand a second source — AMD's core asymmetric optionality
- $100B+ data-center revenue target by 2030 — $120B server-CPU TAM (35%+ CAGR) — large runway if MI400 volume proves out
- Rack-scale systems (Helios, via ZT Systems) move AMD up the value chain toward higher-content, stickier sales
- Sovereign-AI buildouts and enterprise inference diversifying demand beyond a few hyperscalers
- ROCm/open-software momentum could erode CUDA lock-in over time
Threats
- Nvidia's pace (annual cadence: Blackwell→Rubin), CUDA lock-in, and full-stack pricing keep AMD a price-taker
- Hyperscaler custom silicon (Google TPU, Amazon Trainium/Inferentia, Meta MTIA, Microsoft Maia) competes for the same inference budget AMD targets
- Customer concentration + warrant dilution — Meta and OpenAI each hold warrants for up to 160M shares (~10% each at full vest)
- Export controls / China restrictions (already caused FY25 inventory charges) and geopolitical TSMC/Taiwan risk
- HBM4 supply/cost constraints could cap MI400 ramp or compress margin
Moats, dependencies & bottlenecks
Moats
only AMD and Intel hold x86 licenses Structural duopoly; AMD is the share-gainer with a roadmap/process lead. The most durable part of the franchise — funds the GPU push.
Credible #2 merchant AI GPU, but ~5–7% share and chasing, not leading. Edge is execution + second-source demand, not lock-in.
ZT Systems) Moves AMD up the stack to systems; raises content per deal and switching cost, but Nvidia's NVLink/rack play is ahead.
High switching costs, long design-win lifecycles, sticky industrial/comms/aero customers — but slow-growth and currently flat.
The key gap vs Nvidia's CUDA. Improving and open, but not yet a moat — it's the thing the bear case hinges on.
Dependencies
N3/N2 for CPUs and MI400 compute chiplets) No in-house fabs; all advanced silicon is TSMC. Taiwan geopolitical + capacity-allocation risk is the single largest external dependency.
HBM is a major COGS line and the gating component for MI400; AMD dual/triple-sources to de-risk, but HBM4 supply/cost could cap the ramp.
Bull case leans on two names; each holds up to 160M-share warrants. A pushout or cancellation at either is a thesis-level event.
Data-center revenue rides AI buildout spend; a capex pause hits the growth engine directly.
Already drove FY25 inventory charges; further tightening removes a market and creates write-down risk.
Advantages
- Only credible full-stack alternative to Nvidia across both CPU and GPU — unique second-source position
- EPYC server-CPU share momentum with a roadmap/process lead over Intel
- Fabless, net-cash (~$7.3B), FCF-rich (~$6.7B FY25) balance sheet funds the AI push without dilution-by-necessity
- Two 6GW anchor deals + hyperscaler breadth de-risk the demand side of the MI400 ramp
- Chiplet architecture and ZT Systems rack integration move it up the value chain to higher-content systems
Weaknesses
- ~55% gross margin vs Nvidia ~75% — captures volume/TCO, not pricing power
- CUDA-vs-ROCm software gap caps share and premium pricing
- AI-GPU share still only ~5–7%; absolute scale gap vs Nvidia is widening
- Heavy reliance on two anchor customers (Meta, OpenAI) with large warrant dilution (~20% combined at full vest)
- Single leading-edge foundry (TSMC) and HBM-supply dependence
- Client & Gaming cyclicality + rising memory/component costs (guided weaker 2H26); Embedded flat
Bottlenecks
- HBM4 supply and cost — the gating input for the MI400/MI450 volume ramp in 2H26
- TSMC N2 (2nm) and CoWoS/advanced-packaging capacity allocation, shared with Nvidia and Apple
- ROCm software maturity / developer mindshare — the friction that limits GPU share and pricing
- Rack-scale system integration throughput (Helios via ZT Systems) — execution risk on first large deliveries
- Skilled GPU/software engineering talent in a market where Nvidia sets the comp bar
Top signals & trends
Top signals
The single most important proof point — converts ~12GW committed into recognized revenue and validates the bull thesis. Watch for slip.
Our board's explicit watch trigger — a second non-anchor mega-win would move AMD from watch-to-long toward core.
Tests whether AMD can price closer to value vs pure TCO discount; the gap to Nvidia's ~75% is the whole VOLUME-not-PRICING debate.
Supply constraints could gate the ramp or compress margin; a binding bottleneck caps upside.
Management already guided softness — a drag that offsets data-center strength near-term.
Vesting confirms demand but dilutes; track the ~320M-share overhang against deployment.
Trends
Opens the merchant/bare-GPU market where a second source matters most — AMD's core optionality.
Hyperscalers and sovereigns actively diversifying; the structural tailwind behind the Meta/OpenAI wins.
Competes for the same inference budget AMD targets — the same customers building in-house.
Validates demand but inflates COGS and creates allocation risk; pressures gross margin.
EPYC franchise compounding; AMD raised its outlook to 35%+ CAGR.
Drives the guided 2H26 Client & Gaming softness.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
Sole leading-edge foundry — N3 for current parts, N2 (2nm) for MI400 compute chiplets. The critical dependency.
Lead HBM3E/HBM4 supplier (Korea-listed); HBM is a major COGS line and the gating MI400 input.
HBM4 partner (named primary on MI455X) and advanced-packaging/foundry option (Korea-listed).
US-based dual-source HBM supplier — proven on MI350, extending to MI400.
Upstream EUV lithography monopoly enabling the TSMC nodes AMD depends on.
Acquired 2025 — rack-scale system integration for Helios; manufacturing arm being divested while AMD keeps design.
Up-to-6GW Instinct deal (~$60B/5yr), custom MI450; holds warrant for up to 160M AMD shares. Anchor AI customer.
6GW Instinct deal starting 2H26 with MI450; warrant for up to 160M shares (~10%). Private; the second anchor.
Azure deploys EPYC and Instinct MI-series; major hyperscaler customer (and custom-silicon competitor via Maia).
OCI is an early large Instinct GPU adopter for AI cloud capacity.
Server OEMs shipping EPYC + Instinct systems to enterprise and cloud (HPE, SMCI, Lenovo also).
Game-console SoCs (PlayStation, Xbox) — the semi-custom royalty stream in Client & Gaming.
The dominant force: ~80%+ AI-GPU share, ~75% gross margin, CUDA lock-in, annual cadence (Rubin). Sets the price AMD discounts against.
x86 CPU duopoly partner/rival in servers (Xeon) and PCs; Gaudi AI accelerators trail. AMD is taking its server share.
Leads hyperscaler custom-ASIC design (co-designs Google TPU, Meta MTIA); attacks the same inference budget AMD targets.
Custom-silicon and AI-networking competitor in the hyperscaler accelerator/interconnect stack.
Trainium/Inferentia, TPU, Maia — the same hyperscalers are both AMD customers and self-supply competitors for inference.
Emerging data-center AI-inference and PC (Snapdragon X) competitor; peripheral today but expanding into AMD's client/edge.