Amazon announces new AI chips, closer Nvidia ties — but it's cloud capacity that matters most

At AWS Re:Invent 2025 Amazon unveiled Trainium3 — its in‑house AI training/inference chip with roughly four times the compute performance, energy efficiency and memory bandwidth of prior generations — and said early customer tests cut AI training/inference costs by up to 50%. AWS also launched "AWS Factories," combining Trainium accelerators with Nvidia GPUs for on‑prem deployments, as Amazon accelerates capacity after supply constraints: CEO Andy Jassy cited 20.2% YoY AWS growth and >3.8 GW added in the past 12 months. Analysts (Wells Fargo, Oppenheimer) expect Amazon to add >12 GW by year‑end 2027 (each GW ~ $3bn annual revenue), potentially supporting up to $150bn incremental annual AWS revenue and delivering meaningful upside to 2026–27 AWS revenue amid intense competition from Microsoft Azure and Google Cloud.

Analysis

Market structure: AWS' dual strategy (Trainium3 + Nvidia-on-prem via AWS Factories) strengthens AMZN's bargaining power across workloads — expect AMZN to convert capacity into revenue with each incremental GW ~ $3B/yr (Wells Fargo) and management guiding >12GW added by end-2027 (~$36B potential incremental revenue). Direct winners: AMZN (cloud + services), NVDA (continued GPU demand & software stack), AVGO/GOOGL (TPU ecosystem); losers: small cloud providers, GPU-only rent-seekers and legacy data-center OEMs facing pricing pressure. This shifts pricing power toward hyperscalers who control chip stacks and power allocation, compressing intermediate margins for GPU resellers.

Key Decisions for Investors

• Establish a 2–3% long position in AMZN (equity or 9–12 month LEAP calls) sized to risk tolerances; add on pullbacks toward $230–240. Rationale: capacity build + Trainium3 efficiency supports >20% AWS growth; target 12–18 month upside of 20–35% if capacity milestones (add >3GW/yr) are met.
• Allocate 1–2% to NVDA via a 3-month call spread (buy 1–2% notional at-the-money, sell 5–10% out-of-the-money) to capture near-term GPU demand with limited capital. Close or roll if NVDA implied vol drops >30% or if quarterly orders miss by >10%.
• Relative-value pair: go long AMZN and short MSFT 0.75:1 (size smaller vs long) for 3–6 months to express AWS re-acceleration vs Azure complacency; trim short if MSFT outperforms by >8% in 30 days or AMZN misses AWS growth <15% YoY.
• Reduce small-cap cloud and GPU-reseller exposure by 50% and increase allocations to datacenter infrastructure names (AVGO, GOOGL) if AWS announces >3GW incremental capacity in a single quarter; stop-loss: cut AMZN exposure if AWS growth falls below 12% YoY or if public policy restricts GPU exports to China within 90 days.