{"slug": "cea-leti-ceo-ais-real-bottleneck-is-architecture", "title": "CEA-Leti CEO: AI’s Real Bottleneck Is Architecture", "summary": "CEA-Leti CEO Sébastien Dauvé told EE Times that AI's primary bottleneck is architectural, not just energy, as the industry must integrate memory, photonics, communications, and sensing into complex systems. He highlighted the emergence of physical AI interacting with the real world through vehicles and robots, demanding specialized, energy-efficient architectures. The comments came ahead of Leti Innovation Days 2026, which will focus on advanced packaging, optical interconnects, and memory-centric designs.", "body_md": "Speaking in an exclusive interview with EE Times ahead of Leti Innovation Days 2026 in Grenoble, CEA-Leti CEO Sébastien Dauvé argued that while the next phase of AI will be constrained by energy availability, the first challenge is how to integrate memory, photonics, communications, and sensing into increasingly complex systems.\n\n“The bottleneck is really architectural,” he said. As AI workloads grow in scale and complexity, the challenge is increasingly about moving, storing, and managing data efficiently rather than simply adding more compute.\n\nThe theme echoes throughout this year’s Leti Innovation Days agenda, which covers topics ranging from advanced packaging and optical interconnects to memory-centric architectures and Europe’s semiconductor strategy.\n\n**AI moves into the physical world**\n\nAccording to Dauvé, the past year has marked a turning point for AI. “What has changed is the scale and speed of investment worldwide, across the entire value chain—compute, memory, photonics, packaging,” he said. “We now clearly see that AI is moving beyond experimentation into real applications.”\n\n[View All](https://www.eetimes.com/category/sponsored-content/)\n\nWhile much of the semiconductor industry has recently focused on training and inference inside large data centers, Dauvé sees a second wave coming. “We also see the emergence of physical AI,” he said.\n\nUnlike [generative AI](https://www.embedded.com/generative-ai-at-the-edge-unlocks-speech-control-for-robots/) systems confined to data centers, physical AI interacts directly with the real world through vehicles, robots, industrial equipment, healthcare devices, sensors, and other connected systems. That changes semiconductor design priorities, pushing intelligence closer to where data is generated and creating demand for more specialized, energy-efficient architectures.\n\nThese systems operate under strict constraints on power consumption, latency, reliability, and safety. They also require processing closer to where data is generated, driving demand for edge computing and increasingly specialized architectures.\n\nThat shift is reflected throughout the Leti Innovation Days agenda, which highlights applications spanning automotive systems, healthcare, defense, industrial automation, communications infrastructure, and next-generation sensing technologies. In Dauvé’s view, these emerging markets will increasingly shape semiconductor roadmaps alongside traditional data center requirements.\n\n**The bottleneck is architectural**\n\nFor decades, semiconductor progress was driven primarily by transistor scaling. While process technology remains important, Dauvé said he believes future gains will increasingly come from how systems are organized and integrated. “There is no single technology,” he said. “It’s really about combining several building blocks.”\n\nThose building blocks include FD-SOI platforms, advanced memories, silicon photonics, optical interconnects, advanced packaging, and heterogeneous integration. “The key point is integration,” Dauvé said. “We need to bring compute, memory, communication, and sensing together.”\n\nThe conference program reinforces that view. Sessions throughout the event focus on memory-bandwidth-centric computing, advanced 3D integration, chiplet-based architectures, optical-electronic convergence, and new approaches to system design. The common thread is a growing recognition that moving data efficiently has become as important as processing it.\n\nMemory sits at the center of that discussion. Increasingly, the industry is treating memory not as a supporting component but as a primary determinant of AI system performance.\n\nDauvé argued that conventional architectures, which separate memory and compute into distinct domains, are becoming increasingly inefficient for AI workloads. “There is a lot of innovation in new types of memory—spintronics, FeRAM, and others,” he said. “But more importantly, memory and compute must be more tightly coupled.”\n\nBringing the two closer together reduces the energy and latency penalties associated with moving data across increasingly large AI systems.\n\nOptical technologies are also becoming increasingly important. As AI clusters grow larger and more distributed, electrical interconnects face mounting challenges in bandwidth, latency, and energy consumption. Reflecting that trend, Leti Innovation Days will devote an entire track to optical interconnects, examining how silicon photonics could improve data movement both within and between future AI systems.\n\nTogether, these developments point toward a broader architectural shift in which memory, communication, and compute are increasingly co-designed rather than optimized independently.\n\n**Energy becomes the constraint**\n\nArchitecture is only part of the challenge. The other is energy. According to Dauvé, industry must push digital technologies forward while accounting for environmental constraints, particularly energy consumption and materials use. “Data centers consume a huge amount of energy, and it is increasing rapidly,” he said.\n\nYet Dauvé argued that the challenge extends beyond improving chip efficiency. “The real issue is the availability of reliable electricity,” he said. “In some regions, projects are already being delayed—or even canceled—because the energy infrastructure cannot support them.”\n\nThe implication is that future AI growth may depend not only on semiconductor innovation but also on power generation, grid capacity, and infrastructure investment.\n\nMany sessions at Leti Innovation Days address this challenge directly. Topics range from energy-efficient AI factories and sustainable semiconductor manufacturing to optical interconnects designed to reduce data-center power consumption. The RESOLVE initiative, which will be presented during the conference, is specifically examining how Europe can address what organizers describe as a “1,000× energy challenge” while maintaining competitiveness.\n\nFor Dauvé, solving these problems requires a system-level approach. “It means we cannot design technology in isolation anymore,” he said. “We need to co-design the entire stack—materials, devices, circuits, systems, and applications.”\n\nThe growing complexity of AI infrastructure also increases the need for collaboration. “The complexity of modern systems forces us to work across disciplines,” Dauvé said. “That’s why collaboration is essential.”\n\n**From research breakthroughs to industrial ecosystems**\n\nWhile much of the discussion around AI focuses on technology, Dauvé believes Europe faces a broader challenge: turning research leadership into industrial impact.\n\nThe emphasis on industrialization is reflected across initiatives such as FAMES and RESOLVE, which aim to strengthen Europe’s ability to move semiconductor innovations from research into production. The goal is not simply to invent new technologies, but to accelerate their path into production.\n\n“We work not only on technology but also very closely with industrial partners and end users,” Dauvé said. “This allows us to connect research with real applications and industrial deployment.”\n\nThe same thinking shapes Dauvé’s view of Europe’s semiconductor strategy. “The first Chips Act was very positive,” he said. “It created strong momentum in terms of investment, collaboration, and engagement with industry.”\n\nHowever, he said the next phase should focus more heavily on applications and markets. He also argued that Europe must strengthen capabilities across the entire value chain, including advanced packaging and heterogeneous integration. That objective is reflected in his vision for the end of the decade. “For me, success would be to have built a strong European value chain around advanced 3D integration and packaging,” Dauvé said.\n\nSuch an ecosystem would extend beyond packaging itself to include photonics, heterogeneous integration, manufacturing capabilities, and collaboration across research organizations and industry. “It does not have to be entirely European,” he said. “But it should be a strategic asset for Europe.”\n\nIn Dauvé’s view, the future of AI will not be defined by a single breakthrough device or process technology. Instead, it will depend on the industry’s ability to rethink architectures, manage energy constraints, integrate previously separate technologies, and accelerate the journey from laboratory innovation to industrial deployment.\n\nFor Dauvé, success by 2030 will not be measured by a single breakthrough device or process node. It will be measured by whether Europe succeeds in building a competitive ecosystem around advanced packaging, photonics, memory, and heterogeneous integration. In the AI era, he argued, the real challenge is no longer inventing individual technologies but integrating them into systems that can scale from laboratory breakthroughs to global deployment.\n\n##### See also:\n\n**Lam Research, CEA-Leti Open Fast Lane for Specialty Technologies**\n\n**Inside FAMES: How Europe Depends on RTO Collaboration**\n\n**Applied Materials, CEA-Leti Deepen Ties with Specialty Chip Lab**", "url": "https://wpnews.pro/news/cea-leti-ceo-ais-real-bottleneck-is-architecture", "canonical_source": "https://www.eetimes.com/cea-leti-ceo-ais-real-bottleneck-is-architecture/", "published_at": "2026-06-23 07:00:00+00:00", "updated_at": "2026-06-24 00:34:27.597066+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-chips", "ai-ethics", "ai-research"], "entities": ["CEA-Leti", "Sébastien Dauvé", "EE Times", "Leti Innovation Days", "FD-SOI", "FeRAM", "spintronics"], "alternates": {"html": "https://wpnews.pro/news/cea-leti-ceo-ais-real-bottleneck-is-architecture", "markdown": "https://wpnews.pro/news/cea-leti-ceo-ais-real-bottleneck-is-architecture.md", "text": "https://wpnews.pro/news/cea-leti-ceo-ais-real-bottleneck-is-architecture.txt", "jsonld": "https://wpnews.pro/news/cea-leti-ceo-ais-real-bottleneck-is-architecture.jsonld"}}