Amazon SageMaker

You could also self-host it in the cloud.

So like ECS service, for instance, or like the SageMaker server, something like that.

and then using your applications, and then you have all possible kind of things.

Like some of the reported usage, what people come up with, obviously, like Vlad mentioned email, so you can have tool calls.

It says AI.

Generative AI inference recommendation for Amazon SageMaker now available in the SageMaker AI Studio.

So Amazon SageMaker AI now offers generative AI inference recommendation in SageMaker AI Studio, giving customers a guided low-code, no-code pathway.

to find the best inference or configuration for the workload.

This builds on the API-based launch on April 2026.

The interesting bit is that this is for SageMaker, so this is when you are building your own, enhancing your own models.

This is...

Yeah, this is managed models, isn't it? This is

We are providing AI compute with custom-built chips in Frontier and Tranium, along with other GPUs that we provide.

Alongside sits the Amazon SageMaker for custom building of models, as well as data foundation for AI.

Attached to this layer, the middle layer, is for AI and agent development software and services.

You all would have heard of Amazon Bedrock.

So that's where the whole open model or open weights self-hosted becomes the key.

Whether it's self-hosted locally or you're on AWS, you have SageMaker or you have your own EC2 instances and you're running your own show.

And you're getting an open-weight model that's running from there, right? So that's kind of what I think about from the raw horsepower.

But that's just the inference piece.

Para lo último en la capa de infraestructura cloud, le damos la palabra a Amber.

AWS y Unsloth publicaron esta mañana cuatro patrones concretos de implementación para ejecutar modelos de lenguaje de gran escala cuantizados en S2, SageMaker, EKS y ECS, reduciendo la huella de memoria en un 75 % y los costos de inferencia hasta en un 80 %.

Es una señal de costos muy significativa para las empresas que aún dudan entre APIs propietarias y modelos autoalojados.

Por otro lado, AWS también activó la gestión nativa de casos dentro de Quick Automate, lo que permite a los desarrolladores conectar flujos de trabajo agénticos con seguimiento completo del ciclo de vida y con puertas de aprobación humana, sin necesidad de andamiaje personalizado.

Last week, Indian company Razorpay announced the launch of Razorpay Vulcan, India's first AI transformer model built specifically for payments, Perth announcement.

The model combines Razorpay's payment data with NVIDIA's accelerated computing and AWS cloud infrastructure, including Amazon SageMaker.

For millions of local shoppers, a single failed transaction, a delayed OTP, or a fraudulent charge is reason enough to revert to cash.

As the country commerce market heads toward a projected $250 billion by 2030, the reliability of digital payments has become an economic priority.

We have to have the right integration to begin with across all of the layers of the AI lifecycle in order to secure things properly, whether it's model development or runtime environments or inference and usage or identity or the data layers, etc.

And so, for example, when we talked to you guys a while ago about this integration factor with Falcon, and it works across whether you're using SageMaker or Bedrock or the IAM Identity Center, it's because you have to put all of those pieces together there.

And to operate quickly means it's not only about being able to look back forensically at what happened, because yes, that's useful, but guess what? You're too late if you're doing forensics.

It's about how close to absolute real time do you have evaluation of what's currently going on, the changes in the threat landscape, the way your systems are operating, and building rules, detections, and taking actions like shutting things down.

Para conocer lo que está moviendo la infraestructura para desarrolladores, le damos la palabra a Amber.

AWS y Unsloth publicaron conjuntamente cuatro patrones de despliegue para modelos de lenguaje de gran escala, cuantizados en S2, SageMaker, EKS y ECS.

Las guías prometen hasta un 75 % de reducción en memoria y un 80 % de recorte en costos de inferencia.

Es un manual concreto para equipos que buscan ejecutar modelos capaces sin los presupuestos de la frontera tecnológica.