Sep 30, 2026 · 14 min
AI’s next bottleneck is moving data, not building GPUs
IMEC Says Today’s AI Will Look Ancient in 10 Years
As transistor scaling slows, the future of AI may depend on scarce memory, power, cooling, manufacturing capacity, and faster connections between components.
- 1AI infrastructure increasingly strains memory bandwidth, power supply, cooling systems, and data movement rather than GPU availability alone.
- 2Photonics could make AI systems faster and more efficient by replacing copper connections, but moving data remains the central challenge.
- 3Steven Latré expects today’s AI to look primitive within five to ten years as hardware advances and software becomes more efficient.
Don't miss
Steven Latré predicts that today’s AI systems will look primitive within five to ten years as hardware and software evolve.
The brief
Steven Latré of imec and Adam Chambers argue that AI’s next constraints may come from memory, power, cooling, manufacturing capacity, and data movement rather than GPUs alone.
As Moore’s Law slows, chip development becomes a long research and manufacturing process, while larger AI models expose shortages in memory bandwidth, regional capacity, and energy.
Photonics offers a possible alternative to copper interconnects, promising faster, cooler, more efficient data transfer—but the need to move data will survive any winning AI architecture.
Latré’s closing forecast is stark: within five to ten years, today’s AI may look primitive, with a new software revolution following advances in hardware.
Featuring
Steven Latré
IMECListen to the full episode and explore every guest, topic, and moment on PodLume.