Amazon Plans to Sell Custom AI Chips Externally, Challenging Nvidia's Compute Market
Amazon is negotiating to sell its custom AI chips externally for use in other companies' data centers, directly targeting Nvidia's market share in AI compute chips.
This move focuses on the Trainium training chips and Inferentia inference chips, aiming to attract external customers with lower-cost solutions. Amazon CEO Andy Jassy has previously confirmed that the annualized revenue run rate for the chip business has reached hundreds of billions of dollars.
The cloud giant is accelerating the externalization of its self-developed chips, with rising demand from buyers (AI companies seeking diversified supply), while Nvidia's dominance faces pressure from market diversification. Funding may shift towards more cost-effective solutions, driving a restructuring of pricing and supply chains in the compute market.
Source: Public Information
ABAB AI Insight
Amazon's Annapurna Labs has long developed the Trainium and Inferentia series chips, initially serving internal AWS large-scale AI training and inference needs, and has previously supported clients like Anthropic. This external sales move continues its strategy of reducing infrastructure costs, aligning with Google's TPU externalization path.
On the capital front, Amazon is leveraging its own data center scale to validate chip performance before shifting to external sales, focusing resource mobilization on TSMC manufacturing and software optimization. The motivation is to capture the explosive demand for AI compute while diversifying away from reliance on Nvidia, aiming for independent growth of the chip business to hundreds of billions of dollars.
Similar to Google Cloud TPU sales and Microsoft's Maia development, hyperscalers are transitioning from internal use to external supply. This move by Amazon strengthens its competitive position in cloud AI infrastructure.
Essentially, this is a restructuring of the industry chain, where cloud giants lower costs and build ecosystems through vertical integration of self-developed chips, prompting a shift in AI compute pricing power from a single GPU supplier to diversified cloud platforms, accelerating the decentralization of the industry supply chain.
ABAB News · Cognitive Law
Self-developed chips first for internal use, then external; scale validation is the market passport.
A 40% cost advantage outweighs a 20% performance lead, as AI compute enters a cost-performance era.
Supplier dependence is a risk exposure; diversification is the long-term moat.