AI Factories , Trainium3, Bedrock & Nova Garman also moved into one of the biggest shifts in AWS strategy. AI factories are the cloud giant's attempt to bring hyperscale training inside customer walls. Many governments and enterprises have racks of data centre capacity but not the talent to stitch together giant clusters. "Why can't we help more customers—the ones who really need this large-scale infrastructure, see what our expertise, our services, are understanding?" he said. These AI factories allow AWS to drop its full stack inside a customer environment while meeting sovereignty rules, essentially giving the feeling of owned infrastructure without the pain of building it. Then came silicon. AWS has quietly shipped more than one million Trainium chips. Moreover, at re:Invent, it unveiled the Trainium3 UltraServers and previewed Trainium4. According to Garman, Trainium4 delivers "over 6x the FP, 4x performance, 4x more memory family and 2x more high bandwidth memory capacity." He also revealed that it doubles power efficiency. The point was not just raw compute. Garman said inference loads now resemble training. "There's not going to be an experienced application of a system built that doesn't rely on inference." He wants Trainium to become the base for training, low-latency inference and giant agent systems that never sleep. Alongside all this came AWS's biggest CPU upgrade yet. Graviton5 will power the new M9g EC2 instances. AWS said it delivers up to 25% higher performance than the previous generation while improving energy efficiency. It uses 192 cores, a 5x larger L3 cache, faster memory speeds and 3nm technology. Bedrock continued its march toward becoming the world's largest neutral model hub. It now serves more than one lakh customers, and will add 18 new open-weight models across Google, MiniMax, Mistral, NVIDIA and OpenAI gpt-oss. Garman said customers are running many models at once. "This mix and match is going to be normal." AWS also refreshed its Nova line. - Nova Light cuts costs on reasoning tasks.
- Nova Pro handles heavier reasoning across video and documents.
- Nova Sonic adds multilingual speech-to-speech.
The unified Nova multimodal model accepts text, images, video and speech inputs for teams that want a single system that can "output different forms of text and imagery" without maintaining multiple workflows. [AIM Exclusive at AWS re:Invent 2025] AWS positions India as a priority market for regulated, multi-model and sovereign deployments across IT, BFSI, GCCs and mobility. AIM Network was reporting live from Las Vegas with exclusive insights from industry experts. |
Комментариев нет:
Отправить комментарий
Примечание. Отправлять комментарии могут только участники этого блога.