The moving bottleneck: Why scaling AI is harder than before
About this session
Artificial intelligence appears to scale almost magically: add more data, more computing power and more sophisticated models, and intelligence should keep improving. But in the real world, scaling AI rarely works that way.
Every time we remove one constraint, another emerges. A faster GPU exposes a memory bottleneck. More GPUs create a networking bottleneck. Faster infrastructure reveals problems in power, cooling, data movement, software orchestration or reliability. And when all of those work, the bottleneck may move again—to cost, human trust or our ability to use the technology responsibly.
This is the moving bottleneck.
Drawing on decades of experience scaling complex technology and infrastructure systems, Subha Shrinivasan explains why the next era of AI will not be won simply by building bigger models or buying more hardware. It will be won by people who can see the system as a whole, recognize where the constraint has moved and redesign the system around it.
Using accessible examples from factories, traffic and modern AI data centers, this talk reveals a counterintuitive principle: progress does not eliminate bottlenecks—it relocates them.
The lesson reaches far beyond AI. Whether we are scaling a company, a city or our own ambitions, the greatest constraint is often not the problem we have already solved. It is the one our success has just created.
Speaker
Key takeaways
- Scaling AI, Startup & Products