AWS Bullish on Trainium AI Chips as Demand Soars for Inference and Training

Amazon Web Services is showing strong confidence in its homegrown AI accelerators, with CEO Andy Jassy revealing that the Trainium2 chip is in high demand and driving significant revenue growth. Trainium2, now fully subscribed across AWS data centers, is delivering a 30% to 40% better price-to-performance ratio on AI workloads compared to other available options, according to Jassy. The chip is being used not only for AI training but also for inference, suggesting AWS may be phasing out its older Inferentia series in favor of the more versatile Trainium line. Trainium2 has become a cornerstone of AWS’s AI infrastructure, with the majority of token processing in Amazon Bedrock now running on Trainium chips—either Trainium2 or earlier versions like Trainium1 and Inferentia2. This shift underscores the growing importance of in-house hardware in powering AWS’s AI services. A key highlight is Project Ranier, a massive supercluster built in partnership with Anthropic to train the latest Claude 4.X models. The system, which launched in late 2024, now includes 500,000 Trainium2 chips and is on track to scale to 1 million chips by the end of 2025. This cluster offers five times the performance of previous GPU-based systems used for training Claude 3, marking a major leap in AI training capacity. Looking ahead, Trainium3 is in development with Taiwan Semiconductor Manufacturing Co. using a 3nm process. It is expected to deliver twice the performance of Trainium2 and 40% better energy efficiency, likely meaning higher floating-point operations per watt. The chip will enter preview by the end of 2025 and reach full production volumes in early 2026. AWS already has strong interest from large and mid-sized customers, with plans to scale Trainium3 UltraClusters to offer four times the aggregate capacity and double the per-chip performance of Trainium2. Jassy emphasized the need to deliver the chip at scale and continue improving the software stack, noting that real-world success stories like Anthropic’s use of Project Ranier are building credibility for Trainium. On the infrastructure front, AWS added 3.8 gigawatts of datacenter capacity in the past 12 months, with another gigawatt coming online in Q4. The company expects its total datacenter capacity to double by the end of 2027, up from around 4 GW at the end of 2022. This rapid expansion reflects the massive demand for AI compute, with estimates suggesting that building out 10 GW of new capacity—half GPU, half Trainium—could require $435 billion in spending over 2026 and 2027. In Q3 2025, Amazon reported total revenue of $180.17 billion, up 13.4%, with net income rising 38.2% to $21.19 billion. AWS revenue grew 20.2% to just over $33 billion, though operating income increased only 9.4% to $11.43 billion, pressured by the cost of deploying both Blackwell GPUs and Trainium2 systems. Capital expenditures totaled $36 billion, with about $26.4 billion allocated to IT infrastructure. Of that, $28.2 billion went toward AI clusters, and $2.6 billion to non-AI workloads. Based on spending patterns, Trainium accelerators likely accounted for around 35% of AI infrastructure spending, while GPUs still dominate in terms of cost and revenue, despite Trainium’s growing share of compute capacity. The core systems business—compute, storage, and networking—generated $19.47 billion in revenue, up 33.8%, with operating income of $5.73 billion, up 21.8%. This growth signals that compute is rapidly catching up to software in importance, with AI-driven demand pushing the entire infrastructure stack to new levels.