Amazon May Start Selling Its Custom AI Chips to Outside Companies. Should Nvidia Investors Be Worried?

Amazon is reportedly exploring external sales of Trainium AI chips, potentially creating a lower-cost competitor to Nvidia’s GPUs, though the move is still in early discussions with no shipping timeline. Amazon’s custom-silicon business is already running at a $20 billion-plus annual revenue run rate and growing triple digits, while Nvidia’s data center revenue still rose 92% year over year to a record $75.2 billion in the latest quarter. The news is directionally negative for Nvidia’s pricing power over the long term, but the article argues the immediate competitive threat is limited and the market reaction was muted.

Analysis

The market is underestimating how Amazon’s chip move changes bargaining power more than absolute unit demand. If Trainium is sold externally, the real lever is not immediate share loss at Nvidia but a slow reset in procurement psychology: hyperscalers will increasingly treat accelerator supply as a multi-vendor sourcing problem, which tends to compress ASPs before it meaningfully shifts volume. That matters most in 12-24 months, when buyers roll new clusters and negotiate refresh cycles, not in the next few weeks.

The second-order winner is the broader AI supply chain, especially vendors that sit across both custom silicon and merchant silicon. Marvell is strategically positioned because any external Trainium push increases the number of custom datapath, networking, and storage-adjacent components required per rack; even if Nvidia loses a few points of share, the rack count still expands. UBER is a quieter beneficiary only indirectly through lower AI inference costs if cloud pricing eases, but the cleaner read-through is to AI application software more broadly: cheaper training and inference should extend runway for usage-based workloads, not collapse demand.

The biggest contrarian point is that this is less a competitive threat to NVDA’s top line than a margin story for later. Nvidia can absorb share pressure while still growing because the market remains supply-constrained, but once supply catches up, the existence of a credible second source becomes a pricing overhang. That overhang is likely to show up first in gross margin compression and longer contract terms, while reported revenue remains resilient for several quarters. The key risk to the thesis is if Amazon’s external Trainium rollout slips beyond the next 12 months, in which case the market has priced in optionality without any near-term monetization.

Key Decisions for Investors

• Initiate a relative-value long AMZN / short NVDA pair over the next 3-6 months: AMZN gets optionality from external chip monetization, while NVDA carries the higher multiple risk if buyers start anchoring on second-source pricing.
• Buy MRVL on weakness into any post-news consolidation, 6-12 month horizon: the stock has leverage to custom-silicon expansion even if Trainium external sales take time to mature; use a 10-15% trailing stop because timing risk is high.
• For NVDA holders, hedge 25-35% of delta with 6-9 month out-of-the-money puts or a call spread overwrite: the thesis is not revenue collapse, but margin compression and multiple compression if competitive sourcing widens.
• If AMZN breaks out on confirmation of external Trainium customers, add via Jan-2027 calls rather than spot: this is a 12-24 month catalyst, and optionality is the better risk/reward than paying up for immediate execution risk.
• Avoid chasing UBER on this headline alone; treat it as a secondary beneficiary only if lower AI compute costs translate into observable margin expansion in later quarters.