Amazon Web Services has given its artificial intelligence strategy With a series of announcements that strengthen both its own chip portfolio and its technological alliance with Nvidia, the company detailed how it will integrate Nvidia technology into its upcoming AI processors and unveiled new servers designed for training and deploying large-scale AI models at its major annual cloud computing conference in Las Vegas.
This move positions AWS even more prominently in the race for artificial intelligence computing, a field where it competes directly with giants like Microsoft, Google, and Meta. Far from simply reselling third-party GPUs, Amazon is betting on a combination of its own chips, specialized interconnects, and agreements with Nvidia to to offer European and global companies more performance and cost options.
Trainium4: Amazon's next chip will speak Nvidia's language
One of the most notable announcements was the confirmation that the company's future AI training chip, known as Trainium4, will incorporate interconnection technology Nvidia NVLink FusionThis solution enables ultra-fast communication between processors, which is critical when thousands of machines have to work in coordination to train large language models.
NVLink is considered one of Nvidia's greatest assets in high-performance data centers, and until now it was primarily associated with configurations based solely on the company's own GPUs. Amazon's decision to integrate this technology into Trainium4 means that its future AI clusters will be able to combine AWS chips and Nvidia GPUs in the same infrastructure, facilitating Hybrid architectures.
Although AWS has not given specific dates for the arrival of Trainium4 to the market, it has made the objective clear: to offer a platform that allows customers to continue taking advantage of the Nvidia software ecosystem, especially CUDA, while migrating part of their workloads to hardware designed by Amazon in search of better availability and price.
This compatibility is especially relevant for organizations in Europe and Spain that have already standardized around Nvidia GPUs and optimized libraries, but are encountering capacity limitations or increasing costs when scaling their generative AI projects.
New servers with Trainium3: more power and less energy
While the development of Trainium4 continues in the background, AWS has already put one into production. new generation of servers Based on the Trainium3 chip. These devices, available from the Las Vegas conference itself, are geared towards training large models and high-volume inference in cloud AI services.
Each server integrates 144 Trainium3 chips And, according to the company, it offers more than four times the computing power of AWS's previous generation of AI hardware. Furthermore, it does so with approximately 40% lower energy consumption, a key factor in a context where electricity costs and network constraints are beginning to impact data center expansion.
Dave Brown, vice president of computing and machine learning services at AWS, emphasized that the goal is not just to boast about raw power, but to demonstrate to customers that it exists. a competitive alternative in terms of price-performance ratio compared to traditional GPUs. The company avoids giving absolute figures, but insists that the leap compared to the previous generation is remarkable.
This efficiency improvement is particularly interesting for European operators, where regulatory and social pressure on the energy consumption of digital infrastructure is constantly increasing. Less electricity for the same volume of training means more sustainable data centers and, potentially, more concise invoices for business customers.
Within Amazon's roadmap, Trainium3 not only seeks to gain ground against Nvidia, but also to reduce dependence on external providers and strengthen a proprietary line of AI chips that can evolve with cycles more controlled by the company.
AI factories: dedicated infrastructure in our own data centers
Another announcement that has generated a lot of attention is the launch of the so-called AI Factories from AWS, a product designed for large companies and public administrations that want to run advanced artificial intelligence systems on their own data centers, without giving up integration with the Amazon cloud.
The model is simple on paper: the client provides the physical space and energy, and AWS takes care of installing, managing, and maintaining the AI ​​system, connecting it with the rest of the platform's services. In this way, European companies with stringent regulatory requirements, or governments concerned about data sovereignty, can maintain total control over sensitive information without exposing it to external infrastructure.
The term AI Factories is not accidental. Nvidia uses the same concept to refer to its own hardware systems optimized for artificial intelligence, and in this case, the AWS solution is being built precisely in collaboration with the GPU manufacturer. Amazon's AI Factories will combine Blackwell chips from Nvidia and the new Trainium3, relying on AWS cloud networking, storage and security.
In addition to hardware, these facilities can be integrated with managed services such as Amazon Bedrock —to orchestrate and deploy foundational AI models— and AWS SageMaker, geared towards the development and training of proprietary models. For companies, this means a high-performance AI environment, but deployed under a hybrid cloud scheme that better adapts to local data regulations.
Meanwhile, other major providers like Microsoft are also moving in the same direction, with local data centers and solutions designed for data sovereignty. Amazon's commitment to its AI Factories reflects the extent to which artificial intelligence is pushing cloud giants towards more hybrid modelsmoving away from the purely centralized model of a decade ago.
Nova and Sonic models and the boost to AWS's AI offering
Alongside the hardware innovations, Amazon used its conference to strengthen its software side with new versions of its artificial intelligence models grouped under the brand NovaThe company presented Nova 2, an evolution that promises greater speed and responsiveness compared to the previous generation.
One of the variants of Nova is capable of interacting with users not only through text, but also through images, voice and videoThis expands the potential use cases in sectors such as customer service, online education, and content creation. This multimodal capability places it on par with other leading proposals in the generative AI market.
In addition, AWS showcased a model called Sonic, geared towards voice interactions. According to Matt Garman, CEO of Amazon Web Services, this system can respond to spoken commands with voice output of "human-like" quality, opening the door to more natural conversational assistants for public and private services.
Although the company admits that its models still face the challenge of gaining market share against competitors such as ChatGPT (OpenAI), Claude (Anthropic) o Gemini (Google)AWS's business figures for the last quarter point to solid growth. Sales for the division increased by around 20%, driven largely by demand for AI computing and infrastructure.
For European businesses, this expanded catalog means more options when choosing an AI platform, both in terms of ready-to-use models and in terms of... infrastructure on which to train proprietary solutions adapted to each sector and local regulations.
Fierce competition in the AI ​​chip and infrastructure race
All these releases occur in a context of intense competition in the market artificial intelligence chipsNvidia maintains a dominant position thanks to its GPUs and the strength of its CUDA ecosystem, but companies like Amazon, Google, Microsoft, and even traditional processor manufacturers are investing millions to avoid falling behind.
In the case of AWS, the investment in Trainium3 and the future Trainium4 is part of a dual strategy. On the one hand, it seeks reduce dependence on external suppliers and better control the costs and availability of its infrastructure. On the other hand, it wants to offer customers hardware alternatives that allow them to optimize spending without sacrificing performance or compatibility with Nvidia tools.
From a market perspective, the combination of its own chips, collaboration agreements with Nvidia, and products like AI Factories positions Amazon as a player capable of competing not only in cloud services, but also in the very foundation of the hardware which drives the new wave of artificial intelligence applications.
For companies and public administrations in Spain and Europe, this scenario translates into a wider range of possible infrastructures—from the public cloud to on-premises or hybrid data centers—with the option to choose between different configurations of price, performance, and data sovereignty. In a sector where technological decisions have direct impact on costsRegulatory compliance and innovation capabilities, the arrival of Trainium3, future integrations with NVLink Fusion and AI Factories add new cards to the deck that will foreseeably accelerate the adoption of large-scale artificial intelligence solutions even further.
