AWS Summit Paris 2025: Bedrock, SageMaker and Mistral AI at the heart of the technological revolution

With 150 sessions covering technology, strategy, sovereignty, and organisational transformation issues, the AWS Summit Paris brought together more than 10,000 IT professionals on April 9. The event allowed attendees to discover the latest technological advances and use cases in various fields, from CloudWatch metrics to talent management in Generative AI, including vector search and AI workload optimisation on EC2.

The GenAI Experience space hosted around ten interactive demos, allowing visitors to experience the applications of generative artificial intelligence directly. Meanwhile, the AWS for Industries space offered industry-specific construction, media, gaming, healthcare, and financial services solutions.

Devoteam, sponsor of the event, marked its presence with several demos at its booth: a generative AI challenge in the artistic field, an AI-generated podcast to relive the 2024 Olympic Games in partnership with L’Équipe, and a coaching experience in a basketball league. Devoteam also hosted two conferences: “ An assistant coach to reach Pro A ” and “AWS Skill Builder: preparing for a competition in the Cloud.”

In this article, we look at the opening keynote hosted by Julien Groues, AWS VP France and Mai-Lan Bukovec, AWS VP Technology. From computing to processors and storage to the convergence of data analytics and AI, let’s explore how AWS’s offering is evolving with the challenges of its customers, such as Safran, Qonto, and Owkin.

Concrete Commitments in France

Drawing an analogy to Amazon’s 1995 origins, Julien Groues highlighted how the company revolutionized online book sales. He emphasized that customer feedback became a crucial innovation driver for transforming the entire shopping experience.

He then detailed several major AWS commitments in France:

• A 6 billion euro investment to guarantee the presence of data and applications on French territory
• Support for the French startup ecosystem, notably through a presence at Station F since 2017
• A program to support future French AI champions
• A commitment to train 800,000 people in cloud and AI professions by 2030
• Environmental objectives with a zero-carbon ambition for 2040
• The launch of the “Sista AI” program. Sista AI will support 20 French startups founded by women in the field of AI, starting in September 2025

Generative AI in France: considerable but unevenly exploited potential

The figures presented by Julien Groues reveal a strong but mixed dynamic in France’s adoption of AI. A new French company integrates artificial intelligence technologies into its operations every three minutes. The overall adoption rate of generative AI reaches 30% in France, which remains below the European average of 42%.

Analysis by company size highlights significant disparities. French startups have a remarkable adoption rate of 68%, exceeding the European average by 10 points. Large companies follow with 53%, while SMEs lag behind with only 28% adoption.

The benefits for companies that have taken the plunge are tangible. 90% of them see an increase in their revenues and 66% record significant productivity gains.

Owkin: AI at the service of the medical revolution

Jean-Philippe Vert, Owkin’s R&D director, presented an AI healthcare use case during the keynote.

Owkin, a leading biotech company, develops solutions based on AI and the analysis of complex medical data. Jean-Philippe Vert presented the three main areas of their mission:

1. Discovering new treatments using AI
2. Accelerate and reduce risks when moving from a drug candidate to a marketable product
3. Developing AI-based diagnostic tools to identify patients likely to benefit from specific treatments

Jean-Philippe Vert highlighted the major challenges his company faces, including accessing and securely handling medical data from numerous hospitals, as well as transforming this data into actionable knowledge. A particularly crucial issue for the pharmaceutical industry is reducing the clinical trial failure rate, currently at 90%, to offer more accessible treatments that reach more patients.

The partnership between Owkin and AWS is structured around several axes:

• Access to significant computing power for training AI models on large amounts of data
• Deploying models to the scientific community, using solutions like AWS Bedrock to ensure scalability of services
• The implementation of secure solutions for the storage and access of sensitive medical data
• The creation of innovation communities, as illustrated by organising a hackathon on glioblastoma (brain cancer).

What is the future of healthcare technology? Jean-Philippe Vert explained that while AI was initially used to improve the productivity of researchers and doctors, the future lies in increasing human capabilities beyond mere efficiency.

“What we are building today is a new biological superintelligence that combines the best of the latest innovations in AI with access to data and automated laboratories,” he said, mentioning, in particular, a project that will soon be launched in New York in collaboration with AWS.

The challenge of the complexity of living things requires moving “ from productivity to the new frontiers of science.”

Security and a global network as the foundations of innovation

The conference continued with Mai-Lan Bukovec. With 15 years at AWS, she centered her talk on: “What if you could create anything you imagine?“

AWS’s fundamental approach is based on creating efficient services for specific functions while making it easy to combine them. Two fundamental pillars underpin all of their activities:

1. Security: From the silicon level to the entire technology stack, AWS’s security architecture ensures that no one, not even AWS personnel, can access customer data without explicit authorisation.
2. Scale: Powered by AWS’s global network, which includes over 6 million kilometres of fibre optic cables connecting regions, availability zones, and edge locations, AWS has increased its network capacity by 80% in the past year alone.

Calculation: the engine of innovation

The first key element is compute. AWS offers 850+ EC2 instance types to support diverse computing workloads:

• Critical business applications
• Large in-memory databases
• Complex analyses
• Scientific modeling
• Video transcoding
• Machine learning inference

With the rise of generative AI, accelerated computing instances have seen explosive growth. Bukovec highlighted the 14-year partnership between AWS and NVIDIA, highlighting the upcoming P6 instance family, powered by NVIDIA’s Blackwell GPUs and GP200 Superchips, which is scheduled for release later this year. She also added that NVIDIA has chosen AWS for its generative AI cluster project.

She highlighted EC2 Capacity Blocks for organisations with short-term needs. These Blocks provide immediate and predictable access to high-demand accelerated instances with NVIDIA GPUs, ideal for training and fine-tuning workloads.

Customer case: Canva

The design platform now has over 220 million monthly users and aims to reach one billion. Cameron Adams, Canva’s co-founder, explained how its mission, “to empower users to transform their ideas into seamless design,” has evolved to incorporate AI features for creating whiteboards and presentations.

Scaling generative AI products to Canva’s global user base requires a global presence in a datacenter with appropriate capacity and GPU instances.

“Using AWS ensures we have all the basic functionality we need, secure, reliable, and available when needed .”

Processors: Graviton, Inferentia and Trainium

Mai-Lan Bukovec then discussed AWS’s processor initiatives, starting with Graviton. Now in its fourth generation, Graviton 4 is up to 45% faster than its predecessor for large Java applications while consuming 60% less power at the same price. This performance explains why AWS has deployed more Graviton processors in its datacenters than any other type of processor over the past two years.

Graviton’s success led AWS to develop dedicated AI chips: Inferentia and Trainium. Bukovec highlighted the new Trainium 2 instances, which offer the best price-performance ratio for generative AI and deep learning on AWS, with up to four times faster execution speeds and up to three times lower power consumption than Trainium 1 instances.

Poolside: AI for Software Development

Founded in 2023, Poolside distinguishes itself by specialising in high-impact software environments in industries such as financial services and defence. Its approach prioritises security, deploying models entirely within customer security boundaries and never training base models on proprietary data.

Eiso Kant, CTO of Pooloside, detailed his approach to reinforcement learning based on code execution feedback, where models learn like human developers through trial and error. This method has allowed them to complete hundreds of millions of different tasks, learning from both successes and failures.

The partnership with AWS has enabled Poolside to offer secure fine-tuning and custom model management within customer environments, leveraging EC2 and, more recently, AWS Bedrock.

Revolutionising data storage and management

Mai-Lan Bukovec continued her presentation by focusing on the next essential element of the AWS ecosystem: storage. She recalled that in the early days of cloud computing, development teams designed storage solutions from scratch for each new project, resulting in developers spending approximately 70% of their time building infrastructure rather than innovating.

This challenge gave rise to Amazon S3, which now has over 400 trillion objects. S3 has become a central repository for data AWS services use, including Bedrock knowledge bases, Amazon QuickSight business dashboards, and Redshift high-performance data analytics.

She highlighted the evolution of storage solutions, including the rapid adoption of open table formats like Apache Iceberg. She explained that Apache Iceberg is a high-performance, open-source table format designed for large analytical datasets consisting of hundreds or even thousands of smaller files in formats like Parquet, with exabytes of Parquet data already stored in S3.

“The advantage of Iceberg is its ability to work with multiple Parquet files simultaneously, allowing them to be queried like a SQL table,” she explained. It’s the metadata that makes this possible: Iceberg uses a hierarchy of metadata and manifests to manage data snapshots and table updates.

Simplify data management

To address the challenges of managing Iceberg at scale, AWS introduced a new bucket type in S3 in late 2018, specifically designed to support Iceberg. This innovation allows users to manage and query data in Iceberg without additional maintenance tasks such as compacting files or deleting orphaned files.

Mai-Lan Bukovec highlighted the extensive capabilities this offers, such as trend analysis using Iceberg time travel queries and version learning pipelines for generative AI applications. “And that’s just the tip of the iceberg,” she added.

The Power of Metadata

Drawing on the importance of metadata in modern data architectures, she discussed the fact that many organisations currently operate complex and expensive metadata databases. These typically require creating a list of stored objects, creating and managing event processing pipelines that add metadata to objects, storing that metadata in databases, and developing code to synchronise metadata when adding and deleting objects.

To simplify this process, AWS introduced S3 Metadata last year. This automatically creates system metadata such as object size, storage class, and encryption status, which can be easily queried via SQL. Users can also add their own object tags to make it easier to query specific data.

“S3 metadata allows you to add your own tags, which helps you find the right data when querying .”

Convergence of data analytics and AI

Mai-Lan Bukovec noted a significant industry trend: the convergence of data analytics and AI workloads around similar data workflows. This convergence has made it increasingly important for organisations to access all their data and tools in one place to accelerate collaboration.

AWS has responded to this trend with the new version of Amazon SageMaker, which is now generally available. This updated version combines AWS’s best-in-class dedicated analytics services with the existing SageMaker service, creating a centralised platform for all data analytics and AI tasks.

She demonstrated the streamlined workflow: “To find data, just search. And if you don’t have access, simply subscribe to request it. When you’re ready to create, all the tools you need are at your fingertips. ”

Sagemaker Unified Studio supports different user profiles:

• Analysts working with SQL can use a new unified query editor to query from Redshift or Athena without leaving SageMaker
• Data scientists can work on EMR and Python from a single notebook, moving directly from data processing to model training without switching environments.

“Work that once required switching between tools and interfaces can now be done with a single, unified studio,” Bukovec said.

What makes SageMaker’s next generation particularly powerful is its Lake House data architecture, which is fully compatible with Apache Iceberg and allows seamless data flow between different users and tools.

All Lake House data can be managed through an integrated SageMaker catalogue that provides end-to-end governance, enabling data and AI teams to collaborate and share data, models, and generative AI assets across projects with trust and transparency.

To go further:

SageMaker Unified Studio: Centralized access to data analytics, AI and governance

Safran: When aeronautics takes flight in the cloud

“ Every two seconds, a commercial aircraft powered by Safran engines takes off. Every seven seconds, a commercial aircraft powered by Safran landing gear lands. ”

Frédéric Verger thus highlighted the group’s position as a world leader in several areas: aircraft engines (in a 50-50 partnership with GE Aerospace), helicopter engines, landing gear, electrical equipment, aircraft seats and interiors.

Two fundamental priorities for Safran:

1. “Safety First” – ensuring the safe operation of engines and equipment
2. Contribute to the decarbonisation of aviation, with the ambition of achieving net zero CO2 emissions by 2050

Frédéric Verger highlighted the progress already made, notably with their latest engine capable of running on 50% sustainable bio-sourced fuel, while the future generation will be able to use 100% sustainable fuel.

Among the operational challenges:

• Innovation, with 75% of R&D efforts devoted to decarbonisation
• Restoring supply chain performance after the Covid crisis
• The development of activities in the area of sovereignty (Safran notably produces the Rafale engine)
• Digital transformation, identified as a strategic axis to accelerate and industrialise company processes

“Among all the strategic levers to successfully achieve our digital transformation, we have decided to choose AWS as Safran’s preferred cloud service provider,” he explained, specifying that 80% of the group’s data is eligible for the cloud, the rest being regulated or particularly critical data.

Generative AI at the heart of aeronautical innovation

In 2024, the group completed around ten generative AI projects in a few months, using Amazon Bedrock and unified architectural patterns.

Concrete examples include:

• A generative AI assistance tool for aircraft engine engineers, capable of identifying regulatory requirements related to a particular topic in a corpus of more than 1200 documents
• A support tool for support engineering teams in charge of repairing nacelles (the engine casings), which extracts from various histories and databases the repair cases most similar to a customer need, saving engineers “ hours and hours of research.”

Of all the use cases implemented in 2024, Verger indicated an average savings calculation of one million euros over an 18-month period. In 2025, Safran is accelerating with 20 projects underway.

Towards a golden age of knowledge management

“Many of the use cases we address leverage Safran’s expertise in our data. We believe we are beginning a new golden age of knowledge management,” said Frédéric Verger.

The CDO specified that the implementation of these use cases is accompanied by project-by-project change management and a group-wide change plan, with technical training provided in partnership with AWS.

The Future of Inference with Amazon Bedrock

For Mai-Lan Bukovec, generative AI is a critical component that will soon become “ as fundamental as compute and storage ” for modern applications.

“The problems that generative AI can solve will make inference a part of every application .”

Mai-Lan Bukovec methodically outlined the key requirements for building sustainable generative AI applications, starting with model selection. She highlighted Bedrock’s competitive advantage, offering a wide range of fully managed base models from leading vendors such as AI21 Labs, Anthropic, Cohere, Luma, Meta, Mistral AI, and Amazon, including Nova models.

A significant portion of her presentation focused on the challenges of data integration. While augmented data generation (RAG) was initially designed for unstructured data, organisations are using a variety of data types. She announced expanded capabilities that enable Bedrock knowledge bases to support structured data through SQL query generation, multimodal data processing through Bedrock Data Automation, and understanding complex relationships through GraphRAG.

Bukovec cited Trellix’s implementation of Bedrock for cybersecurity threat investigations as an example to illustrate these capabilities in action. By combining Nova with its proprietary investigation techniques and the billions of security records stored in Amazon OpenSearch, Trellix reportedly reduced its investigation costs by 65%.

Bedrock Guardrails’ automated reasoning capabilities

Also of note is Bedrock Guardrails’ automated reasoning capabilities to combat AI hallucinations, a significant concern for businesses in terms of adoption. She described this solution as “the first and only generative AI protection that prevents factual errors due to hallucinations through logically accurate and verifiable reasoning .”

Mai-Lan Bukovec also discussed cost optimization strategies, including model distillation and intelligent fast routing, which could potentially reduce costs by up to 75% and 30%, respectively, without significant compromises in accuracy.

To conclude this segment, she introduced AI agents as the next evolution beyond conversational AI, demonstrating how AWS Q Developer Agents can autonomously assist throughout the development lifecycle, from code generation to testing and documentation.

Qonto’s AI-powered financial innovation

Aymeric Augustin, Qonto’s CTO, then explained how the European fintech is leveraging AWS’s AI capabilities to transform banking services for SMEs and the self-employed.

Qonto has implemented a real-time AI monitoring system running on Amazon Bedrock that analyses transactions for suspicious patterns. Using advanced core models like Claude 3.7 in a regulatory-compliant environment, this system prevented €3 million in fraudulent transactions in just one month. This success is based on LLMs’ contextual understanding of the financial world, their ability to process unstructured data, and their understanding of human intentions, even in abbreviated transaction labels.

Amazon Q Migration Capabilities

Building on the theme of AI-powered transformation, Mai-Lan Bukovec again highlighted Amazon Q’s migration capabilities, highlighting AWS’s 19 years of experience in cloud migration.

She presented three critical migration paths that Amazon Q simplifies through intelligent automation :

For Windows .NET applications migrating to Linux, she explained how Q automatically detects incompatibilities, generates transformation plans, and refactors source code, accelerating modernisation up to four times faster while reducing Windows licensing costs by up to 40%.

For VMware migrations: Q’s transformation agents automate application discovery, wave planning, and end-to-end migration orchestration. She specifically highlighted its ability to convert complex virtual machine network configurations and firewall rules into native AWS structures at speeds up to 80 times faster than manual processes.

Q also facilitates the migration of mainframe applications, including IBM’s z/OS systems, by analysing, decomposing, planning, documenting, and even refactoring COBOL to Java, transforming a process that would normally take years into one that takes months. To illustrate this point, she shared a success story from Toyota North America, where Q condensed a months-long legacy mainframe documentation process into a single day, allowing migration planning to proceed with unprecedented efficiency.

Mistral AI

To conclude this morning, which was full of announcements, Julien Groues welcomed Arthur Mensch, CEO of Mistral AI, on stage for a discussion on the evolution of their strategic partnership, which was announced a year earlier.

Asked to share his assessment of the past year, Arthur Mensch explained how Mistral AI has evolved:

“We’ve pivoted from a company that only made models to a company that makes a platform that helps businesses gain value with generative artificial intelligence. ”

This transformation has manifested itself in the development of “ more efficient, cheaper, faster, more diversified ” models with multimodal capabilities, accompanied by a platform allowing companies to customise these models according to their specific needs.

Arthur Mensch cited Veolia as an example, which uses its technologies not only for knowledge management but also in water treatment plants, where operators interact conversationally with information systems. He also mentioned Veolia’s work with defense companies using SageMaker and its CodeMistral model, which has been customized through fine-tuning to better match these organisations’ specific needs.

Mistral Large availability

Regarding current value creation, he recalled the importance of document management, announcing on this occasion the availability of Mistral Large in several AWS regions, including in Europe.

“Mistral Large is a native multimodal model that can process documents and extract certain structures from them, ” he said, highlighting its usefulness in the financial sector for analysing structured documents like PDFs. He also mentioned Mistral OCR, available on SageMaker, which he called “ the best model on the market for processing documents quickly .”

The discussion was punctuated by a surprise announcement: Mistral Large has been available on Amazon Bedrock in Europe since 11:57 p.m. the day before, with an imminent deployment in the United States and on Trainium 2 in Paris.

Concluding on the future outlook, Arthur Mensch revealed that Mistral AI is working on a new generation of models “much more efficient in terms of reasoning capacity ”, with new modalities and a particular focus on customer service applications.

“For us, the challenge this year is to move from the model, which is just a small building block, to complete systems that connect to data, interact with users, and continuously train through these interactions. This is the new frontier, where we are investing a lot of resources, ” he concluded.