Interview, Fireside Chat
IMEC Says Today’s AI Will Look Ancient in 10 Years
- Hardware development involving unique clean room facilities for next-generation chips is projected as a five-to-ten-year process, with new chips expected to emerge in seven to ten years.
- The semiconductor industry faces the end of physics regarding transistor scaling, necessitating significantly more inventive innovation than achieved in the preceding 10 to 20 years to maintain the roadmap.
- Value is shifting from software to hardware as the software market remains immature and relies on brute force approaches, with software value anticipated to increase only after closer integration with hardware.
- Memory shortages are expected to persist and drive industry-wide price increases due to insufficient production capacity and high inventory absorption by hyperscalers.
- Chip manufacturing is identified as the primary bottleneck for industry growth, particularly concerning the production of specific component types across different global regions.
- The era of building ever-larger AI models is predicted to cease as companies abandon the scaling hypothesis to focus on resolving memory bottlenecks.
- Data transfer architecture is expected to transition to copackaged optics within the next five years, relocating optical components closer to the GPU.
- Photonics is predicted to become the critical technology for future chips by addressing energy, cooling, and data transfer constraints.
- Current AI technology is expected to appear obsolete within five to ten years, comparable to the dial-up internet era.
- A software revolution is expected to occur within 10 years, moving beyond current large language model brute force methods.
- Companies combining European academic or technical backgrounds with US business acumen are identified as likely to succeed, reflecting a trend among successful founders.