أمازون تُظهر تفاؤلًا قويًا بمعالجات Trainium المطورة داخليًا لتعزيز أداء الذكاء الاصطناعي في سحابة AWS

Amazon Web Services (AWS) is making bold moves in the AI infrastructure race with its homegrown Trainium accelerators, signaling a strategic shift toward in-house AI chip development to boost profitability and control over its cloud platform. The Trainium2, now fully subscribed and generating multi-billion-dollar annual revenue—up 2.5 times sequentially from Q2—has become a cornerstone of AWS’s AI strategy. According to CEO Andy Jassy, it delivers 30%–40% better price/performance than alternatives on AI workloads, driving strong demand from large customers. Trainium2 is powering not only AI training but also inference, particularly within AWS’s Bedrock and SageMaker platforms. Jassy confirmed that the majority of token processing and generation in Bedrock now runs on Trainium chips—often Trainium2 or older Inferentia2 units—suggesting a gradual phase-out of the Inferentia line in favor of the newer, more versatile Trainium platform. The scale of AWS’s commitment is evident in Project Ranier, a massive AI supercluster built in collaboration with Anthropic. Initially projected to house hundreds of thousands of Trainium2 chips, it now comprises 500,000 chips—with plans to expand to 1 million by year-end—delivering five times the performance of Anthropic’s prior GPU-based clusters used for training the Claude 3 models. This underscores AWS’s push to become a dominant AI infrastructure provider. Looking ahead, Trainium3 is set for preview by year-end 2025 and full deployment in early 2026. Built on TSMC’s 3nm process, it promises twice the performance and 40% better energy efficiency than Trainium2. With projected 4X aggregate capacity and 2X per-chip performance over Trainium2 UltraClusters, it will enable even larger AI model training, especially for partners like Anthropic and other major AWS customers. AWS is simultaneously expanding its datacenter footprint at an unprecedented pace. Over the past 12 months, it has activated 3.8 GW of new capacity, with another 1 GW expected in Q4. The company projects its total datacenter capacity will double by 2027, from around 4 GW in 2022 to an estimated 10–20 GW. This massive buildout, driven by $125 billion in capital spending in 2025 (with $28.2 billion dedicated to AI infrastructure), reflects a deep bet on GenAI. While GPU-based systems remain central to AWS’s offering—especially for customers seeking raw power—homegrown accelerators like Trainium are becoming critical to cost control and margin protection. Estimates suggest that by 2026–2027, half of AWS’s AI compute capacity may be Trainium-based, though GPUs still command higher revenue due to premium pricing. Financially, AWS delivered $33.1 billion in revenue in Q3 2025—a 20.2% year-on-year increase—while operating income rose 9.4% to $11.43 billion. Despite rising costs from deploying both Blackwell GPUs and Trainium systems, the core systems business (compute, storage, networking) grew 33.8% to $19.47 billion in revenue, with operating income up 21.8% to $5.73 billion. As AI workloads surge, compute is rapidly catching up to software in revenue contribution, with AWS’s infrastructure services poised to match or exceed traditional software revenues in the coming years. This marks a return to the early AWS era, when EC2 outpaced S3—now reborn in the age of AI. With Trainium at its core, AWS is not just adapting to the GenAI revolution—it’s leading it.