Amazon CEO Andy Jassy announced at the end of June:Amazon’s self-developed chip business annual revenue has exceeded 20 billion US dollars, a year-on-year increase of more than 100%. This number covers three product lines: Graviton processors, Trainium AI training chips and Nitro security chips. If calculated based on comparable market pricing, the equivalent independent revenue will be close to US$50 billion. This means that the AI chip market’s “Nvidia alone” pattern is substantially loosening.
Where did the $20 billion come from?
Trainium is a dedicated chip for AI training developed by Amazon. It is different from Nvidia GPU - it is specifically optimized for large language model training from the design stage, rather than general-purpose GPU computing. This results in significant reductions in training costs, and this cost advantage accrues over multi-year contracts. at present,OpenAI, Anthropic, Meta and UberBoth have signed a long-term purchase agreement with Amazon for Trainium.
This $20 billion figure validates an important trend: Enterprise AI labs and cloud service providers are actively reducing their sole reliance on Nvidia and instead investing in custom chips that can match specific workload needs.
The new pattern of AI chips among four parties competing for hegemony
As of June 2026, the AI chip market has formed a clear four-party competition pattern:
- NVIDIA: Still leading in the field of general-purpose GPU training, the CUDA ecosystem is the largest moat
- Google TPU: The most mature custom chip alternative, which has been iterated for many generations.
- OpenAI Jalapeño: The first self-developed chip, specializing in inference cost optimization, teaming up with Broadcom to challenge NVIDIA’s inference hegemony
- AmazonTrainium: The most widely adopted enterprise training alternative with long-term orders from the largest customer base
It is worth noting that Qualcomm has just acquired Modular for US$3.92 billion - this transaction will provide a "chip-independent" software deployment layer, allowing developers to freely switch between different AI chips. This is equivalent to building an "open alternative" outside of CUDA, further accelerating the diversification process of AI chips.
What does it mean for users?
Competition in AI chips has intensified, and the most direct beneficiaries are end users. Multi-vendor competition means: AI inference and trainingCosts continue to fall(More training capacity means cheaper API prices),Supply is more stable(no longer stuck on the waiting list for NVIDIA GPUs), andMore customized AI services(Different chips are suitable for different tasks).
For observers of China’s AI industry, Amazon Trainium’s success has an additional meaning—it provesBuilding custom AI chips from scratch is a viable business path, even in the face of Nvidia’s absolute lead.
Summarize
$20 billion in annualized revenue isn’t an “interesting data point” — it’s a clear sign that the AI infrastructure power dynamics are being reorganized. Nvidia is still strong, but the days of "only one choice" are ending. This means cheaper, faster, and more diverse options for everyone who uses AI services.
Continue to pay attention to the core trends of the AI industry? access AI Dash ——Discover the best AI tools.