Advanced Micro Devices, Inc.

▲ Bull Case

• ◆AI GPU ecosystem breakthrough: ROCm 7+ closes the training gap for major frameworks; Microsoft Azure and Google GCP both deploy AMD Instinct at >10% of AI training workload. AMD's AI GPU share rises to 18–20% by FY2028 vs. the ~12% base assumption. Data Center revenue reaches $40B+ by FY2028. FY2027E EPS $13.00 × 40x = $520 (+23%).
• ◆EPYC duopoly pricing power: Intel's 18A manufacturing transition fails to restore competitiveness through FY2027; EPYC pricing power improves as Intel becomes unable to match on performance-per-watt. AMD's x86 server CPU share reaches 52–53%, enabling gross margin expansion to 57–58%. Operating leverage amplifies EPS well above the base case.
• ◆Meta 6GW ahead of schedule + multi-cloud replication: Meta Phase 1 (1–2 GW) completes by mid-2026; Phase 2 begins. Microsoft and Oracle both announce strategic AMD Instinct deployments. AMD's AI GPU backlog extends beyond FY2027, reducing cycle risk.

▼ Bear Case

• ◆BIS MI450 restriction + AI capex normalization: BIS restricts MI450 (and next-gen Instinct) exports — removing $1.5–2.5B China AI GPU workaround revenue — simultaneously with hyperscaler AI capex pausing in H1 2027. Data Center revenue falls from $35B FY2027E peak to $24B by FY2028 (–31%). FY2027E EPS $7.50 × 25x = $188 (–56%).
• ◆CUDA moat holds and multiple re-rates: ROCm fails to become the default second choice for AI training — hyperscalers return to NVIDIA for new builds. AMD is re-rated from "AI GPU company" (50x multiple) to "CPU company with GPU optionality" (22–25x). At $10.69 EPS × 22x = $235.
• ◆Custom ASIC acceleration: Google TPU v5+, Amazon Trainium 3, Microsoft Maia 2, and Meta MTIA 2 together displace both NVIDIA and AMD in hyperscaler deployments. The non-NVIDIA AI GPU market compresses as hyperscalers build vertically. AMD's AI GPU revenue peak is FY2026–2027 with a lower long-term ceiling than consensus projects.

“The central debate: 'Is AMD's AI GPU moat durable enough to justify a 40–50x multiple, or is CUDA the real moat and AMD is a temporary beneficiary?' The bull narrative: AMD has proven at Meta scale that ROCm works; the CUDA-ROCm gap is closing; hardware performance (MI300X 288GB HBM3E) is differentiated for inference. As AI spending scales to $500B+/yr, structural demand exists for a second GPU vendor. The bear narrative: CUDA's 4M+ developers and 20+ years of ecosystem lock-in is not a software bug that can be patched by ROCm. Every major framework (PyTorch, JAX, TensorFlow) has CUDA as first-class and ROCm as afterthought. AMD's GPU wins are in inference and custom configurations — not broadly replicable training workloads. The resolution matters: if AMD is a 'structural AI GPU participant,' a 35–40x multiple is defensible; if a 'cycle-peak GPU winner,' reverting to CPU multiples (22–25x) is probable, with downside to $230–290.”

• ◆Q2 FY2026 earnings vs. $11.2B guide (July 2026) — HIGH magnitude; binary beat/miss significantly impacts bull case probability
• ◆MI450 first shipment confirmation (H2 2026) — HIGH magnitude; validates next-gen GPU roadmap; required for FY2027 Data Center ramp
• ◆Meta 6GW Phase 2 deployment confirmation (H2 2026) — MEDIUM-HIGH magnitude; contract execution de-risks FY2027 Data Center revenue
• ◆BIS export control policy update (Ongoing) — HIGH magnitude; MI450/next-gen restriction would immediately compress valuation
• ◆EPYC quarterly server share reports (Quarterly) — MEDIUM magnitude; each 1pp share gain = ~$200M incremental annual revenue
• ◆Intel 18A process node update (H2 2026) — MEDIUM magnitude; Intel failure extends EPYC moat; Intel success adds competition
• ◆AMD investor day and FY2028 guidance update (Late 2026) — MEDIUM magnitude; first multi-year guidance with MI450 architecture confirmed

• ◆BIS MI450 + advanced GPU restriction (MEDIUM-HIGH probability): Removes $1.5–3.0B revenue, –$0.35–0.55 EPS impact; most important risk combined with AI capex normalization
• ◆CUDA ecosystem durability failure (MEDIUM probability): ROCm fails to close training gap; multiple re-rates from 40x to 22–25x, producing $235–267/share even at $10.69 EPS
• ◆AI capex normalization FY2027–2028 (MEDIUM probability): Data Center revenue declines 14% base to 25% bear scenario; EPS $9.12–7.50 range
• ◆Custom ASIC displacement (LOW-MEDIUM, 5–10yr): Compresses non-NVIDIA GPU TAM; AMD ceiling lower than modeled
• ◆TSMC Taiwan geopolitical disruption (LOW-MEDIUM probability): Existential risk; AMD has no alternative to TSMC N3/N2; entire business model at risk
• ◆HBM supply constraints (MEDIUM probability): HBM4 limited by SK Hynix/Micron capacity; AMD AI GPU ramp delayed
• ◆Intel CPU recovery (LOW near-term probability): 18A success would slow EPYC share gains and pressure pricing at top-of-rack deployments