Uber is the latest to be won over by Amazon’s AI chips

Uber expanded its AWS contract to run more ride-sharing features on Amazon-designed chips, increasing use of Graviton and beginning a trial of Trainium3. Amazon says Trainium is already a multibillion-dollar business, and the deal signals a competitive win for AWS against prior cloud partners Google and Oracle. The move follows SoftBank's acquisition of Ampere (Oracle sold its stake for a $2.7B pre-tax gain) and may shift cloud workload and AI chip demand toward AWS, with potential sector-level implications for cloud providers and chip suppliers.

Analysis

AWS’s in‑house silicon is a lever for durable gross‑margin arbitrage: by lowering server TCO vs incumbent x86 instances, AWS can offer customers a 15–35% effective compute cost advantage on ARM‑friendly workloads, compressing competitors’ pricing power and forcing multi‑cloud customers to rebalance within 6–18 months as proofs of ROI accumulate. That shift is not binary — it will show up first in high‑volume, latency‑sensitive inference and streaming workloads where per‑inference energy and instance cost matter most, before broader migration of stateful services.

The direct competitive effect on Nvidia is asymmetric: Trainium‑class silicon primarily attacks inference and model hosting economics, which could shave a low‑double‑digit percentage off incremental GPU demand growth for those use cases over 12 months, but Nvidia retains a structural edge in high‑throughput training and ecosystem lock‑in. Second‑order supply‑chain winners include packaging, motherboard and power‑delivery vendors aligned to ARM server designs while traditional x86 supply (and Intel exposure) faces adjustment risk in cloud procurement cycles lasting 12–36 months.

Key catalysts and reversal risks are technical parity (benchmarks vs H100-class GPUs), rollout cadence of customer migrations, and commercial pricing tactics from hyperscalers (temporary subsidies or spot pricing). Expect near‑term newsflow (3–6 months) around trial results and more multi‑year commitments; a single large public benchmark or a major enterprise migration win could accelerate re‑rating, while clear superiority in training workloads from Nvidia would blunt the thesis and reverse positioning quickly.

Key Decisions for Investors

• Overweight AMZN (6–12 month horizon): buy a modest directional call spread (e.g., 6–9 month debit call spread sized to 2–4% of tech book). Rationale: capture AWS re‑rating from differentiated silicon and sticky multi‑year deals; target 20–35% upside vs max loss = premium paid; trim on outsized benchmark misses or antitrust headlines.
• Relative pair — long AMZN / short GOOGL or ORCL (12 month horizon): equal notional sizing, 50–100 bps portfolio tilt. Rationale: express cloud share migration and pricing pressure; target ~15–25% relative outperformance with stop if pair moves 10% against position within 3 months.
• Tactical long UBER (3–9 month horizon): buy UBER shares or buy a 3–6 month call spread sized small (1–2% portfolio). Rationale: infrastructure cost takeout should deliver 50–150bps operating margin tailwind over the next two quarters; expected upside 15%–25% vs defined premium loss if migration stalls.
• Defensive tail hedge on NVDA (3–6 month horizon): purchase small NVDA put spread (cost‑limited, 0.5–1% of portfolio) to protect AI‑heavy exposure. Rationale: guards against accelerated substitution in inference workloads that could compress GPU ASPs; payoff asymmetric if Trainium adoption materially dents GPU demand.
• Monitor catalysts and be ready to add exposure to ARM ecosystem suppliers (server board, memory SOPs) on validated performance wins — initiate after two independent public benchmarks show >=20% TCO improvement versus x86, with 12–18 month holding period and position sizing of 1–3% each.