Semiconductors / AI Chips

AMD: The High-Quality Second Place

Key Partners

• TSMC — sole leading-edge foundry partner (3nm/5nm); AMD has no own fabs • Microsoft — exclusive Xbox Series X/S SoC supplier; Azure data center EPYC buyer • Meta, Google, Oracle — large-scale Instinct GPU customers • Lenovo, HP, Dell — Ryzen PC and EPYC server OEM partners • PyTorch/TensorFlow community — ROCm ecosystem compatibility partners

Key Activities

• CPU/GPU microarchitecture development and generational iteration (1-2 year cadence) • ROCm software ecosystem development and developer community maintenance • Process migration planning and IP adaptation with TSMC • Custom chip co-design with cloud providers (AI accelerator variants) • Export compliance management (downspecced GPU variants for restricted markets)

Key Resources

• x86 architecture license (shared with Intel; only two companies worldwide hold this) • Zen CPU microarchitecture family (server and consumer) • CDNA GPU microarchitecture (data center AI acceleration) • Xilinx FPGA IP library and embedded market position • TSMC 3nm/5nm leading-edge process exclusivity

Value Propositions

• EPYC: better price-performance than Intel Xeon in servers, improving every generation • Instinct GPU: lower hardware cost than Nvidia H100, ROCm ecosystem improving steadily • Single x86 vendor covering CPU + GPU, reducing supply chain fragmentation in hyperscale data centers • AI NPU integration into consumer Ryzen chips, addressing on-device AI inference demand

Customer Relationships

• Cloud providers: multi-year strategic procurement agreements tied to infrastructure refresh cycles • OEM partners: joint product roadmap development, locked in 12-18 months ahead • Developer ecosystem: ROCm open-source framework, dev tools, and community for technical stickiness • Gaming community: FidelityFX Super Resolution and other technologies to maintain PC gamer loyalty

Channels

• Direct sales to cloud providers and enterprise data centers (OEM direct supply) • PC market through OEM partners (HP, Dell, Lenovo) • Distributors and resellers for SMB and embedded markets • E-commerce retail channels for consumer CPUs and GPUs

Customer Segments

• Hyperscale cloud providers (AWS, Azure, Google Cloud) — primary buyers of EPYC CPUs and Instinct GPUs • Mid-size cloud providers and HPC supercomputing centers — cost-driven GPU buyers • PC OEM manufacturers (HP, Lenovo, Dell) — Ryzen CPU procurement • Gaming console makers — sole-source SoC supplier for Microsoft Xbox Series X/S • Telecom and industrial equipment vendors — Xilinx FPGA customer base

Cost Structure

• Wafer fabrication costs (TSMC): largest component of cost of goods sold • R&D investment: ~$6-7B/year (~18-20% of revenue) • Xilinx acquisition amortization ($31B acquisition; ongoing goodwill and intangible amortization) • Sales and marketing: includes ROCm ecosystem promotion and cloud provider technical support • Export control write-offs: $440M in 2025 (MI308 GPU China market)

Revenue Streams

• Data Center (EPYC CPU + Instinct GPU): $16.6B, 48% of total (FY2025) • Client (Ryzen CPU) + Gaming (console SoC): $14.6B, 42% of total (FY2025) • Embedded (Xilinx FPGA): $3.5B, 10% of total (FY2025) • Non-GAAP gross margin ~54%; free cash flow $5.5B

Editor's Take

AMD posted record full-year revenue of $34.6 billion in 2025, up 34%. The Data Center segment hit $16.6 billion — nearly half the company. That same year, export controls forced AMD to write off $440 million in MI308 GPU inventory earmarked for China. The turnaround Lisa Su has run since 2014 is one of the most dramatic reversals in semiconductor history: from near-bankruptcy to directly challenging Intel's server dominance and positioning as Nvidia's most credible alternative in AI accelerators. The only question that matters now is whether AMD can close the gap in the AI accelerator race, or whether it's permanently locked into the role of profitable second place.

I. Decoding the Business DNA

AMD is fundamentally an architecture company. It doesn't manufacture chips — fabrication is outsourced entirely to TSMC. Its strategic assets are the x86 architecture license (shared exclusively with Intel), its GPU microarchitecture, and the engineering capability to design competitive silicon every 12-18 months. The fabless model gives AMD the freedom to migrate to each new process node without carrying the capital expenditure of a factory.

In 2025, AMD operates three distinct business lines at different competitive maturity levels:

Data Center: EPYC CPUs continue taking server market share from Intel; Instinct GPUs (the MI300X generation) have established themselves as the primary alternative to Nvidia's H100/H200. Data Center revenue reached $16.6B, up 32% — the company's only high-confidence growth engine. [Source: AMD FY2025 Annual Report]

Client + Gaming: Ryzen processors remain strong in consumer markets; the AI PC trend is pulling demand for on-chip NPUs. Xbox Series gaming console SoC contracts are entering a down-cycle as the console generation matures, but PC CPU momentum offset the gaming decline. Combined Client + Gaming revenue: $14.6B, up 51%. [Source: AMD FY2025 Annual Report]

Embedded: The Xilinx FPGA business serving telecom and industrial markets. After collapsing in the 2023 inventory correction, Embedded revenue hit $3.5B in 2025, down 3% — still recovering. [Source: AMD FY2025 Annual Report]

II. The Earnings Logic

AMD's business model has undergone a fundamental transformation over the past decade — from "cheap alternative" to "architecture premium." In 2015, AMD's gross margins sat around 30%, buying share through price discounts. In 2025, non-GAAP gross margins reached approximately 54%, approaching Intel's. This shift isn't about pricing power in isolation; it's about product mix rotation.

Financial Snapshot

MetricValue
Full-Year Revenue (2025)$34.6B
Data Center Revenue$16.6B (48% of total)
Non-GAAP Operating Income$7.8B
GAAP Net Income$4.3B
Non-GAAP Diluted EPS$4.17
Free Cash Flow$5.5B
Export Control Write-Off$440M (MI308 GPU, China)
Q1 2026 Revenue Guidance~$9.8B

[Source: AMD Q4 2025 Earnings Release and Conference Call]

Two pricing strategies running simultaneously: EPYC CPUs are priced 20-30% below Intel Xeon, using cost advantage to win share; MI300X GPUs are priced 20-30% below Nvidia H100, using the same playbook. In the GPU market, however, being cheaper is table stakes — the real question is whether the hardware is sufficient for the task at hand.

MI300X unit economics: AMD Instinct GPU cards sell at roughly $10,000-$15,000 in volume (vs. Nvidia H100 at $20,000-$30,000). The price delta is compelling. But GPU value isn't ultimately determined by the sticker price — it's determined by the software stack running on top. Nvidia's CUDA ecosystem carries 15+ years of accumulated code, model libraries, framework optimizations, and developer habits. AMD's ROCm is closing the gap, but the gap remains.

The export control overhang: In 2025, the US government added AMD's MI308 GPU (a downspecced variant designed for China) to its export control list. AMD wrote off $440 million in inventory. This isn't a one-time accounting event — it's AMD's permanent exit from China's data center GPU market, which was a meaningful revenue source. [Source: AMD Q4 2025 Earnings Call, export control disclosures]

III. Growth Flywheel and Moat

AMD's growth story in 2025 runs on a single engine: the AI compute demand explosion, with AMD positioned as the primary beneficiary alternative to Nvidia.

The flywheel works like this: EPYC CPU penetration at hyperscalers (AWS, Azure, Google Cloud) gets AMD hardware into the data center rack. Once a customer's server infrastructure runs on AMD CPUs, the incentive to co-deploy AMD GPUs increases — avoiding heterogeneous vendor compatibility friction. GPU shipment volume drives ROCm ecosystem maturation; more AI frameworks optimize for AMD hardware, creating a positive feedback loop.

This flywheel started turning in 2025, but its angular velocity is still far below Nvidia's. Nvidia's moat is ecosystem depth: the code, models, frameworks, and developer muscle memory accumulated on CUDA cannot be priced away. AMD's ROCm has proven itself in HPC contexts (national supercomputing labs); its penetration in general AI inference workloads is still early-stage.

EPYC's server CPU moat is more durable. AMD took server CPU market share from near-zero in 2017 to approximately 25% in 2025 through four successive Zen architecture generations. That share was won on merit, not subsidy. The x86 design barrier is extremely high — ARM-based alternatives (AWS Graviton, Microsoft Cobalt) represent a longer-term structural threat, but that's not the dominant variable in 2025.

IV. Risks and Concerns

Nvidia's software moat: The AI accelerator competition isn't about peak FLOPs — it's about which GPU makes the AI engineer's job easiest. CUDA has a 15-year head start; PyTorch and TensorFlow are natively optimized for Nvidia hardware. ROCm compatibility improved significantly in 2024-2025, but there's a material difference between "works" and "first choice." Large-scale model training (clusters of 10,000+ GPUs) requires tight integration with interconnect topologies and ecosystem components like NVLink — areas where AMD's competitive position remains substantially weaker.

Geopolitical and export control exposure: The China data center GPU door is closed. If future policy expands controls to additional AMD SKUs, the company's total addressable market in global AI chips will continue shrinking. Management cannot fully hedge this risk.

TSMC concentration: 100% of AMD's leading-edge production (3nm/5nm) runs through a single foundry partner. TSMC capacity constraints, geopolitical tension in the Taiwan Strait, or natural disaster risk are external risks that AMD cannot internalize. This is the fabless model's structural vulnerability.

Gaming console cyclicality: Xbox Series X/S SoC contracts delivered a significant revenue contribution, but console generation cycles are finite. The next-generation console's chip may not flow entirely to AMD.

V. The Endgame

AMD is no longer the challenger buying market share on price. It has earned a position in the server CPU market through genuine architectural superiority, and it has achieved "credible Nvidia alternative" status in the AI accelerator market. $34.6 billion in annual revenue and ~54% non-GAAP gross margins indicate that product pricing power has been rebuilt from scratch.

The actual endgame depends on how the AI compute race evolves. If large model training remains CUDA-centric indefinitely, AMD is permanently second place. But if the inference market explodes — edge deployment, on-premise model hosting, mid-tier cloud provider purchasing — AMD's price-performance proposition and progressively maturing ROCm stack create a real opportunity to take larger share. Inference workloads are less dependent on software ecosystem depth than training workloads. That's AMD's entry point.

When Lisa Su became CEO in 2014, the company held $150 million in cash and its core products were being crushed by Intel on every front. She spent ten years rebuilding AMD into the second-largest x86 CPU vendor and the second-largest GPU accelerator maker in the world. That transformation already happened — and the numbers prove it.

AMD in 2025 is a high-quality number two: real moat, genuine growth momentum from the structural expansion of AI compute demand, and competitive pressure that will not disappear. The market is large enough that even without taking significant share from Nvidia, AMD has space to grow and sustain healthy profitability.

Craft your own business model