Forget AMD: Amazon Declares War on Nvidia by Selling Its Own AI Chips

Amazon is exploring selling its custom AI chips to third parties, extending Trainium and Inferentia beyond AWS and potentially turning a roughly $50 billion annualized internal chip business into an external revenue stream. The article frames this as one of the most credible competitive threats to Nvidia, though Nvidia still leads with $193.7 billion in fiscal 2026 data center revenue and a strong CUDA software moat. AWS’s 28% cloud share and Amazon’s $139 billion in 2025 operating cash flow give it the scale to bundle chips with cloud services and pressure AI infrastructure economics.

Analysis

This is less a near-term share-shift story than a margin-allocation story: Amazon is trying to convert AI capex from a third-party cost center into a vertically integrated profit pool. The second-order effect is that every incremental Trainium deployment strengthens AWS’s pricing leverage on compute, storage, and networking simultaneously, which matters more than chip gross margin by itself. If AWS can subsidize silicon with cloud attach, it can pressure Nvidia without needing outright performance parity, especially for inference workloads where latency and total cost matter more than benchmark bragging rights.

The bigger competitive risk for Nvidia is not immediate unit loss but the precedent this sets for other hyperscalers and large enterprises. Once one platform vendor proves it can expose custom silicon externally, procurement teams will push for multi-source architectures and negotiate harder on GPU pricing, especially on multiyear cloud commitments. That creates a slower but real erosion in Nvidia’s mix and bargaining power, while beneficiaries further down the stack likely include packaging, advanced substrate, and memory vendors if custom accelerators proliferate even when Nvidia share doesn't collapse.

Timing matters: the market may overreact in the next few sessions to headline risk, but the operating change is a months-to-years thesis. The main reversal risk is ecosystem inertia — if developer tooling and model portability remain CUDA-anchored, external adoption of AWS silicon could stay contained to cost-sensitive inference rather than meaningful training displacement. A second risk is that Amazon’s external chip push could face channel conflict if it alienates large AWS customers who prefer neutral infrastructure over being sold a quasi-competitive semiconductor product.

Contrarianly, the consensus may be underestimating how defensible Nvidia remains in training and overestimating how fast Amazon can monetize chips outside its own cloud. The better trade is not a blunt anti-Nvidia bet, but positioning for relative dispersion: AWS gains are more plausible than a full Nvidia reset. If Amazon succeeds, the market should reward the platform economics first and only later mark down Nvidia’s terminal growth assumptions.

Key Decisions for Investors

• Long AMZN / short NVDA pair for a 3-6 month horizon: thesis is AWS monetization improves while Nvidia faces gradual pricing pressure; size modestly because CUDA moat limits near-term downside for NVDA.
• Buy AMZN on any post-headline dip over the next 1-2 weeks, targeting a 6-12 month re-rating as the market capitalizes AWS's silicon attach and cross-sell optionality; risk is execution drag on external chip commercialization.
• Avoid aggressive outright short NVDA; instead consider a put spread 4-6 months out to express slower multiple compression if hyperscaler custom silicon adoption broadens, with defined premium at risk.
• Watch AVGO and memory/packaging names as secondary beneficiaries if custom accelerators proliferate; a relative long basket versus NVDA is preferable to single-name directional shorts.
• If AMZN confirms any external chip customer wins, add to the AMZN/NVDA pair and trim on sharp NVDA drawdowns of >5% in a day, since initial overreaction is likely before fundamentals reprice.