I've Changed My Mind on AMD Stock. The AI Supercycle Has Room for More Than Just Nvidia.

AMD struck two large GPU capacity deals — 6 gigawatts each with OpenAI and Meta — that include warrants worth up to 10% of AMD tied to deliveries and stock price, forcing ROCm integration into those customers' ecosystems. Improved ROCm support and the market shift toward inference (where GPUs are less CUDA-dependent) plus materially lower GPU pricing could allow AMD to gain share vs. Nvidia. Separately, agentic AI is expanding demand for CPUs; AMD, as the current data-center CPU leader, is designing new CPU architecture for agentic AI and has acquired ZT Systems to sell pre-configured racks, supporting higher pricing and growth in data-center revenue.

Analysis

The market rotation from training-dominated spend to inference and agentic workloads materially changes the server BOM: inference favors cost-per-inference and throughput over raw FP training horsepower, while agentic agents increase synchronous, latency-sensitive control-plane work that raises CPU utilization per GPU. If buyers trade off peak training performance for improved TCO and higher CPU:GPU ratios, expect a re-weighting of hyperscaler procurement math that benefits suppliers with competitive price/perf and scalable software ecosystems.

Structures where large customers receive equity-like economics tied to delivery milestones create three non-obvious effects: (1) they lengthen strategic vendor relationships because customers have financial upside to see the supplier succeed; (2) they create episodic dilution risk that is event-timed to deployment cycles rather than constant cost of capital; (3) they embed a feedback loop between hyperscaler deployment pace and public equity volatility—positive deployments compress financing risk, negative delays amplify share-price pressure.

Second-order supply-chain winners include rack/system integrators and networking vendors benefiting from higher rack-density CPU-first designs; conversely, incumbents whose moat is software compatibility face latent channel friction if cross-stack portability stalls. Key near-term catalysts are independent inference benchmarks, hyperscaler delivery milestones, and CPU architecture telemetry from early agentic production stacks; adverse catalysts are rapid price moves by entrenched incumbents or any demonstrated software portability gap that reestablishes a one-vendor lock.

Time horizon: inference share shifts can move in 3–12 months as procurement cycles refresh; agentic-CPU demand and material margin expansion are 12–36 months and hinge on sustained production deployments and OEM conversion. Execution and dilution risk make a hedged, event-driven exposure preferable to a naked long.

Key Decisions for Investors

• Initiate a hedged directional position: Long AMD 4% NAV funded by Short NVDA 2.5% NAV (market-value neutral ~60:40). Horizon 6–18 months. Rationale: isolates share-take and CPU-driven margin upside while trimming exposure to broader AI market beta; target asymmetric return of +30–60% vs limited draw if NVDA re-accelerates. Stop: re-evaluate if AMD misses two consecutive data-center revenue guides or if NVDA cuts inference pricing by >20%.
• Options collar to express convexity with defined risk: Buy 12-month AMD ATM call and sell a higher-strike call (1:1) sized to 2% NAV, or buy a 9–12 month call spread (buy ATM, sell +30% strike). Max loss = premium; payoff captures re-rating on deployment beats. Monitor implied vol ahead of earnings—enter after post-earnings vol pullback to avoid IV crush.
• Event-driven short: Buy 9–15 month puts on INTC sized 1–2% NAV or short common if catalyst confirms market share erosion in cloud CPU procurement. Rationale: structural pressure on incumbents who cannot convert latency/agentic workloads into pricing power; reward asymmetric if OEM wins skew to non-incumbent. Cover on any credible product roadmap execution or large OEM contract announced.
• Tactical monitoring alerts (no immediate trade): set alerts for (a) major hyperscaler delivery milestone announcements, (b) third-party inference benchmarks showing cost-per-inference parity delta <10%, and (c) warrant/exercise notices—any two positive events within a quarter should trigger adding to the long exposure; two negatives should trigger de-risking.