The Secret to Finding the Next Broadcom Is Hiding in Plain Sight

Amazon has launched Trainium3, a custom 3nm AI training chip that the company says delivers 4x the compute and 4x the energy efficiency of Trainium2 and has enabled up to 50% reductions in training and inference costs for clients including Anthropic. OpenAI is reportedly considering using Trainium3 in exchange for a potential $10 billion investment, underscoring demand for lower-cost alternatives to Nvidia GPUs. AWS — which accounts for 18% of Amazon's revenue but 66% of operating profits — grew at a 20% annual pace in Q3, and wider adoption of Trainium3 workloads could materially bolster AWS growth and Amazon’s equity performance into 2026.

Analysis

Market structure: Winners are AMZN (AWS/Trainium3), AVGO (custom AI silicon partners) and large cloud customers (Anthropic) that can cut training costs 30–50%; losers are marginal GPU OEMs and Nvidia if share shifts materially. Custom chips trade performance for price — expect downward pressure on GPU ASP growth but sustained TAM growth for datacenter compute; 3nm supply constraints at TSMC could cap rapid Trainium3 rollout near-term. Cross-assets: stronger AWS/AI adoption is risk-on (equities up, IG spreads tighten, 2s10s may steepen); semiconductor volatility and USD funding flows (AI capex in dollars) will spike option vols for NVDA/AVGO/AMZN.

Risk assessment: Tail risks include a failed OpenAI/AMZN deal, Trainium3 under-delivering vs H100 (high-impact), or export/regulatory action limiting model-training chips — each could erase >30% of trade thesis value within 6–12 months. Short-term (days–weeks) moves will be driven by benchmarks and Qs; medium (3–12 months) by customer wins and AWS growth cadence; long-term (>12 months) by software lock-in and models optimized for alternatives. Hidden dependencies: compiler/tooling, customer migration costs and TSMC capacity; catalysts: AWS guidance, Anthropic/OpenAI confirmations, public Trainium3 benchmarks.

Trade implications: Direct trades: establish a tactical 2–3% long AMZN overweight into 2026, sized for +30–40% upside if AWS growth sustains >18% YoY, stop -15% if AWS growth falls <10% for two consecutive quarters. Buy 1–2% AVGO as leverage to custom silicon adoption; target +20–30% in 12 months. Protect existing NVDA exposure with 3–6 month 10–15% OTM puts or replace with a collar; avoid naked short NVDA. Options: buy AMZN Jan 2027 140/240 call spread (~12–18 month) financed by selling 3–6 month calls to capture time for Trainium adoption.

Contrarian angles: Market underestimates customer inertia — most large models are currently optimized for CUDA, so conversion costs could limit Trainium3 share to 10–25% by end-2026, not full GPU substitution. Conversely, consensus may underprice pricing pressure on GPU ASPs, hurting NVDA near-term margins (20–30% downside risk in a prolonged shift). Historical parallel: specialized chips (ASICs) displaced GPUs in niche ML tasks but did not fully unseat dominant general-purpose incumbents; unintended consequence: falling GPU prices accelerate smaller-model experimentation, increasing total cloud spend but compressing per-unit vendor margins.

Key Decisions for Investors

• Establish a tactical 2–3% long position in AMZN (stock) sized to a concentrated sleeve for 12–18 months; target 30–40% upside by end-2026 if AWS revenue growth remains >=18% YoY for next two quarters. Set stop-loss at -15% or exit if AWS growth falls below 10% for two consecutive quarters or if Trainium3 benchmarks lag H100 by >30% on key metrics.
• Initiate a 1–2% long in AVGO as a semiconductor exposure to custom AI silicon adoption; take profits to reduce to 0.5% on a 25% gain, and add on pullbacks of >=15%. Monitor Broadcom’s design wins each quarter — add if new hyperscaler contract announced or revenue beat >5% vs consensus.
• If you hold NVDA >2% position, hedge with 3–6 month 10–15% OTM puts (buy protective puts) or implement a collar (sell 1–2 month calls) to defend against a 20–30% valuation re-rate driven by custom-chip substitution. Do not short NVDA naked; if seeking short exposure, use small size via call overwrites or buy-to-open put spreads.
• Buy an AMZN 12–18 month call spread (example Jan 2027 140/240) financed by selling nearer-term (3–6 month) calls to fund cost — target asymmetric upside to capture multi-quarter AWS/Trainium adoption while collecting premium. Size so max loss = 1–2% portfolio; monitor quarterly AWS growth and public Trainium performance releases as exit/roll triggers.