AMD vs. Nvidia: The AI Supercycle Is Big Enough for Both. Here's the Better Buy.

Nvidia's AI dominance is substantial: revenue rose from under $17B in fiscal 2021 to $216B in fiscal 2026, it commands ~90% GPU market share and a >$4T market cap. AMD has secured large GPU deals with OpenAI and Meta (bringing “hundreds of millions” in revenue) and is well positioned as the data-center CPU leader for rising inference and agentic AI demand. The note concludes AMD offers more upside for investors today given its smaller size and under-penetrated CPU and inference opportunities relative to Nvidia.

Analysis

The second-order winners are not just GPU vendors but the software and fabrication nodes that make alternate stacks viable. If ROCm adoption by hyperscalers accelerates, it lowers switching costs for customers and forces incumbents to defend margin through pricing or vertical integration — that dynamic amplifies upside for foundries (TSMC) and EDA/packaging suppliers while compressing GPU OEM gross margins over a 12–24 month window.

The key structural risk is substitution and efficiency gains rather than demand destruction: model quantization, sparsity techniques and purpose-built accelerators (inference LPUs / TPUs / FPGAs) can reduce aggregate GPU training and inference spend per model by 20–50% over a 2–3 year cycle. Regulatory and contract concentration risks (single-hyperscaler dependency or export controls) are asymmetric: they can rapidly reprice a leader while creating durable niches for well-positioned CPU and inference vendors.

Consensus is pricing a winner-take-most outcome for training GPUs; that underweights the coming bifurcation between large-scale training and pervasive, low-latency inference/agent orchestration. Practically, the next 6–18 months will be dominated by (a) hyperscaler procurement choices, (b) early commercial evidence of agentic workloads running on CPU+inference stacks, and (c) margin signaling from component suppliers — each is a binary catalyst that will re-rate incumbents and their supply chains differently.

Key Decisions for Investors

• Pair trade (12–18 months): Long AMD / Short NVDA equal notional sizing (start 1–2% portfolio). Thesis: capture AMD rerating from CPU + inference optionality while hedging macro and AI-capex cyclicality; target asymmetric payoff ~+40–60% on AMD leg vs limited short-side drawdown ~15–25% if NVDA mean-reverts. Risk management: tighten pair if spread moves >25% adverse; stop-loss at 12% portfolio move.
• Buy AMD directional convexity (9–15 months): Debit call spread on AMD (buy longer-dated OTM calls, sell further OTM calls) to express upside while capping premium. Position size small (0.5–1% portfolio); objective 2.5–4x return if CPU/inference adoption accelerates in next 12 months, limited downside to premium paid.
• Selectively long supply-chain exposure (18–36 months): Buy TSMC (TSM) or ASML-sized positions (1–2% each) to play sustained capex growth if multi-vendor GPU/accelerator demand expands. Risk/reward: 20–40% upside if multi-year node/ASML cadence holds; downside 15–25% on a hyperscaler capex pause — use trailing stop or options collar.
• Vol/arbitrage (0–6 months): Sell NVDA near-term call spreads above current notional (small size, <0.5% portfolio) into rallies to harvest premium against expensive implied volatility, but keep strict tail-hedge (buy deep OTM puts) to protect versus event risk (earnings, supply shock). Expect 20–40% annualized carry with capped catastrophic risk.