Amazon Moves to Sell Custom AI Chips Externally, Challenging Nvidia

Amazon is making a direct play for the external AI chip market, moving beyond its internal AWS deployments to sell its custom silicon to third-party buyers, according to TechCrunch - a step that places the company in direct competition with Nvidia, which currently controls roughly 80 percent of the AI accelerator market.

What Happened

TechCrunch reported on June 18, 2026, that Amazon intends to offer its Trainium and Inferentia chip lines to customers outside its own cloud infrastructure, signalling a significant shift in strategy. Until now, these chips have been deployed almost exclusively to power Amazon's own AWS services, giving the company cost advantages on its own platform but leaving Nvidia unchallenged in the open market. By selling externally, Amazon would compete with Nvidia's H100 and B200 GPU families, which remain the dominant choice for AI training and inference workloads at major enterprises and research institutions. The move also positions Amazon alongside Google, which already sells access to its Tensor Processing Units (TPUs) via Google Cloud, and Microsoft, which offers access to its Maia chips through Azure. Amazon did not disclose specific pricing or a formal launch timeline in the report.

In a separate development in chip fabrication, Quantum Computing Report covered QTREX Quantum's announcement that it has successfully fabricated a cryogenic chip carrier using a single-build additive manufacturing process - a significant step in reducing the complexity of quantum hardware assembly. The same story was also covered by The Quantum Insider, which noted the carrier is designed specifically for quantum computing applications that require components to operate at temperatures near absolute zero. Traditional cryogenic carriers require multi-step fabrication, making QTREX's single-build approach a potential reduction in both cost and lead time for quantum hardware developers. QTREX did not disclose commercial availability dates or pricing in the announcement.

Why It Matters

Amazon entering the external AI chip market matters primarily because of what it signals about the semiconductor supply chain. Nvidia's near-monopoly on AI accelerators has created well-documented bottlenecks: long lead times, elevated prices, and concentration risk for any organisation building AI infrastructure. US export controls on advanced chips have already reshaped global procurement strategies, pushing buyers in restricted markets toward alternatives and accelerating investment in domestic chip programs in China, Europe, and India. Amazon selling Trainium and Inferentia externally adds a credible alternative at scale, which could reduce dependency on a single supplier and gradually diversify the AI chip supply chain.

The broader pattern here, where hyperscalers develop custom silicon and then commercialise it, mirrors the trajectory of cloud computing itself. Those interested in historical parallels for technology trade conflicts will note that dominant infrastructure providers have repeatedly faced competitive disruption from vertically integrated newcomers willing to undercut on price. For the semiconductor industry, Amazon's move also raises questions about the long-term role of TSMC as the primary fabrication partner for these custom chips - a dependency covered in depth in the TSMC's Monopoly encyclopedia entry.

QTREX's cryogenic carrier development is relevant to this supply chain story in a different way: quantum computing hardware currently depends on a very small number of specialised component suppliers. Single-build additive fabrication, if it proves reliable at scale, could open the supplier base for cryogenic components, reducing another potential chokepoint in next-generation computing infrastructure.

Background

The AI chip market has been shaped substantially by Nvidia's early investment in CUDA, its parallel computing platform, which created deep software lock-in alongside its hardware dominance. As described in the GPU Revolution encyclopedia entry, Nvidia's pivot from graphics rendering to general-purpose parallel computation established the template that every subsequent AI accelerator has had to compete against. Amazon's Trainium chips are designed specifically for neural network training workloads, while Inferentia targets inference - the deployment phase of AI models. Both were developed to reduce AWS's own Nvidia spend, which is substantial given the scale of Amazon's cloud operations.

For cryogenic chip packaging, the underlying challenge is that quantum processors must operate at millikelvin temperatures, placing extreme demands on every component in the signal chain. Fabrication techniques that work at room temperature often fail or introduce noise at cryogenic temperatures, making QTREX's single-build approach technically notable if it maintains signal integrity across the full thermal range.

What to Watch

The key indicator to monitor is whether Amazon secures external enterprise customers for Trainium and Inferentia within the next 12 months. Without third-party adoption, the external sales push remains largely symbolic. Nvidia's response will also be telling: the company has historically maintained pricing power through software ecosystem advantages, and any Amazon chip push will need a compelling software story to accompany the hardware. On the quantum side, watch for QTREX to publish performance data on the cryogenic carrier at operating temperatures - fabrication success is only the first hurdle. Regulatory dynamics around AI chip exports will also continue to influence how much room non-Nvidia alternatives have to grow in markets currently underserved by Nvidia allocations.