Last price
$278.39
Picks-and-shovels AI chip supplier; demand inelastic to outcome shift.
Live quote sourced from Yahoo Finance. Prices cited in narrative below reflect the original memo date and may be stale.
AMD's EPYC and MI accelerators power the infrastructure layer for both copilot and autopilot deployments. Sequoia's thesis is agnostic to whether customers build tools or outcomes—both require compute. AMD's exposure is pure enabler: more AI inference, more processing demand, regardless of business model.
AMD is the structural beneficiary of the exact force that is a headwind to NVDA—hyperscaler custom silicon deployment. As Meta, Google, and others launch proprietary inference accelerators, they still need a high-performance CPU partner for orchestration, training, and mixed workloads. AMD's MI and EPYC are becoming the co-processors in custom-silicon stacks.
Custom-silicon buyers (Meta, Google, Amazon) do not optimize solely for inference latency; they optimize for TCO across training, fine-tuning, and inference. EPYC CPUs and MI accelerators work together in these stacks. Unlike NVDA, which loses share as inference moves to custom silicon, AMD gains share because custom-silicon vendors need CPU + accelerator combinations. MI300X/MI350 ramp directly captures the volume that Meta MTIA and Google TPU are shifting away from NVDA.
| Use Case | AMD Position | Trend | Thesis Fit |
|---|---|---|---|
| Foundation-model training (on NVIDIA) | Weak (HPL-like), secondary to GPU | Losing share | Tail |
| Custom-ASIC training clusters | Moderate (EPYC + MI orchestration) | Gaining share | Core |
| Inference at hyperscaler edge | Strong (MI300X/350 ramp) | Growing | Core |
| Consumer/gaming AI | Moderate (Ryzen + iGPU AI) | Emerging | Secondary |
| Embedded AI / automotive | Weak (Ryzen Embedded, minority share) | Stable | Tertiary |
Meta's MTIA co-development through 2030 commits billions of inference dollars. AMD gets a share of this wave via MI chip attach. Even at 30% of the pie, 3-5x growth in custom-silicon inference demand is a multi-year tailwind.
Performance-per-watt metrics are competitive. Customer adoption is accelerating outside NVIDA base (Meta, others). Margin profile (40%+) is structurally superior to EPYC.
Customers who adopt MI for inference benefit from EPYC ecosystem (RDNA, ROCm library maturity, cloud provider support). This lock-in is not as deep as CUDA, but it is real.
AMD's own data-center operations are improving yield and power efficiency via ML-driven optimization. This is a margin boost on top of volume.
Frontier models (Claude, GPT, Gemini) still prefer H200/B200 for training cost and stability. MI300X is viable for fine-tuning and smaller training runs but not the default for 10T+ parameter models.
Meta building MTIA also has incentive to design a custom CPU (avoiding Xeon/EPYC royalties). Google TPU stacks use custom GCE CPUs. Custom silicon TAM could erode both NVDA and AMD simultaneously.
ROCm library adoption is growing but not comparable to CUDA. ML researchers default to CUDA. This limits MI adoption in research-forward workloads, even if hardware is capable.
At P/E ~30x on 25% growth guidance, any MI ramp miss or EPYC pressure triggers multiple compression. Meta or Google custom-CPU announcements would be an instant negative.
AMD is the inference-attached beneficiary of the exact Sequoia thesis that is a headwind to NVDA. Every custom-silicon deployment (Meta MTIA, Google TPU, Amazon Trainium) reduces NVDA's inference share, but increases AMD's MI and EPYC attachment. The thesis question: does inference TAM growth (5-10x) outpace the shrinkage of NVIDIA-centric TAM? If so, AMD wins even with lower unit share. Verdict: Sequoia's thesis is structurally bullish for AMD, with higher execution risk than pure enablers like ASML.
Strong thesis fit as structural compute supplier; business agnostic to outcome-vs-tool arbitrage.