Amazon Web Services Has Strong Partnership With Anthropic, CEO Says

AWS is promoting its Trainium 3 training platform as a cost‑effective, high‑performance alternative to NVIDIA GPUs, highlighting strong performance in real‑world customer use (notably Anthropic) and deep integration across silicon, server design and data centers. Management said more than half of Bedrock inference tokens run on Trainium servers, AWS added 3.8 GW of capacity last year and plans to roughly double capacity to ~8 GW by end‑2027, while acknowledging supply‑chain constraints for chips and data‑center components and that customer demand will determine the mix between Trainium and NVIDIA accelerators.

Analysis

Market structure: AWS’s push of in-house training silicon and large-capacity rollout (3.8GW added last year, target ~8GW by end-2027) benefits AMZN (AWS), data‑centre suppliers (Equinix/Realty Income style REITs), power/commodity suppliers (copper, transformers) and AI customers like ANTHROPIC who need scale. NVDA remains a winner too — multi‑vendor demand expands total TAM even if share shifts; smaller clouds and GPU-only service resellers are the losers as vertically integrated hyperscalers capture more stack value.

Risk assessment: Near term (days–weeks) expect volatility around product/earnings windows and supply‑news; medium term (3–12 months) capacity bottlenecks (chips, power, networking) could cause delivery slippages; long term (1–3 years) regulatory (antitrust, export controls) or slower software migration to agentized workflows could compress IRRs on heavy AWS capex. Tail risks: major export controls or a large datacenter outage that halts model training would materially hit both revenues and sentiment.

Trade implications: Tactical longs — NVDA and AMZN — with exposure also to EQIX/DLR and copper miners (FCX) capture both compute and infrastructure demand. Use options to express asymmetric upside in NVDA (short-dated call spreads into earnings) and buy-leverage on AMZN 6–18 month call spreads; consider AMZN/GOOGL pair to express AWS share gains through 12 months.

Contrarian angles: Consensus underestimates capex drag — doubling to ~8GW implies heavy upfront spend that can depress AWS margins in 2025–26 even if revenues grow; historically (TPU era) custom silicon supplemented but did not replace GPUs, so NVDA upside is likely underpriced. Monitor customer migration rates (PoC→production) and Anthropic TPU/dual‑sourcing signals as 30–90 day catalysts that could flip relative winners.

Key Decisions for Investors

• Establish a 2–3% long position in AMZN (AWS) with a 12–24 month horizon to capture training‑silicon and Bedrock tailwinds; set a -15% stop and a target of +30–50% on successful capacity commercialization and rising Bedrock revenue share.
• Allocate 1.5–2% to NVDA via a 3‑6 month call spread ~5–10% OTM (debit spread) ahead of next earnings to capture continued GPU demand while limiting premium risk; target 20–40% return on spread, max loss = premium paid.
• Initiate a 1% position long EQIX or DLR (data‑centre REIT exposure) and a 1% position in FCX (copper) with 6–18 month horizon to play infrastructure and commodity upside from datacenter buildout; trim if datacenter utilization falls below 70% industry average.
• Execute a pair trade: long AMZN / short GOOGL sized 1:1 at 2% net exposure over 12 months to express AWS share gains in production workloads; unwind if AMZN/GOOGL relative performance deviates by >15% against thesis within 6 months.
• Monitor regulatory and supply catalysts closely: track DOJ/FTC antitrust filings, EU cloud procurement guidance, and Nvidia chip allocation notices over the next 30–60 days; pause new size increases if concrete antitrust actions or export controls are announced.