Meet the Company That's Quintupled Its Share of AI Chip Shipments in 2 Years

Amazon's custom AI chip business is scaling rapidly, with CEO Andy Jassy citing a >$20 billion annual run rate in Q1 and saying it could generate $50 billion of revenue as a standalone operation. Trainium3 demand is strong, nearly all 2026 capacity is sold, and Trainium adoption should improve AWS margins and free cash flow by reducing reliance on Nvidia-priced chips. The article frames this as a growing competitive threat to Nvidia and a potential multi-tens-of-billions-dollar profit lever for Amazon.

Analysis

The key market implication is not simply that custom silicon is taking share from Nvidia; it is that hyperscalers are converting AI demand from a vendor-margin problem into an internal infrastructure optimization problem. That shifts value from the chip OEM to the cloud operator, because every unit of workload migrated to captive accelerators improves both gross margin and supply assurance while lowering sensitivity to NVIDIA lead times and pricing. The second-order winner is anyone selling adjacent tooling, packaging, networking, and power infrastructure into a more diversified accelerator stack, since custom chips increase total deployment complexity rather than reduce it.

Amazon stands out because the chip business is no longer a science project — it is scaling enough to become a material lever on AWS economics. The important nuance is that this is likely to show up first in operating leverage, not in headline revenue: lower effective compute cost should widen AWS margins before the market fully capitalizes the standalone chip revenue stream. Over the next 2-4 quarters, the bigger catalyst is not third-party chip sales but internal substitution, which can offset capex anxiety and support earnings acceleration even if top-line cloud growth is merely steady.

For Nvidia, the risk is less immediate demand collapse and more a gradual mix erosion in the highest-volume enterprise accounts, where custom silicon can absorb incremental training/inference spend. That dynamic may cap multiple expansion even if absolute GPU demand remains healthy, because investors will start discounting a lower long-run share of wallet in hyperscale. The contrarian take is that consensus may be underestimating how quickly “custom” becomes the default for mature workloads, especially as reasoning and agentic applications push customers to optimize cost per token rather than chase peak performance.

The main reversal risk is execution: custom silicon only matters if software stacks, developer tooling, and supply chain yield keep pace. If next-gen accelerators fail to ramp cleanly, or if frontier-model training continues to demand the flexibility of general-purpose GPUs, Nvidia’s ecosystem moat holds longer than bears expect. Watch for signs of third-party uptake outside the top labs; if adoption stays confined to internal workloads, the market may be overpricing the long-term revenue opportunity.

Key Decisions for Investors

• Long AMZN vs short NVDA on a 3-6 month horizon: the pair captures margin expansion at AWS versus share-of-wallet pressure at Nvidia; best entry is on any post-earnings pullback in AMZN or strength-driven squeeze in NVDA.
• Add to AMZN common or call spreads into the next two quarterly prints: the convexity is in margin surprise, not revenue surprise, with the setup favoring upward AWS EBIT revisions over the next 2-4 quarters.
• Initiate a relative-value long GOOGL / short NVDA basket as a hedge on accelerating custom silicon adoption across hyperscalers; thesis is that internal accelerator economics support cloud margin resilience even if overall AI capex stays elevated.
• Avoid shorting NVDA outright; if expressing the bear case, use put spreads 4-6 months out to isolate multiple compression risk from a gradual share-loss narrative while limiting the risk of near-term AI demand strength.
• For higher risk tolerance, buy AMZN Jan-2027 call spreads: the payoff is tied to the market eventually capitalizing AWS as an AI infrastructure platform with chip economics layered in, while downside is limited to premium paid.