The next AI infrastructure bottleneck is observability

Thesis

The binding constraint on agentic workloads is increasingly not raw compute but whether their behavior can be observed, measured, and improved.

01
Beyond compute
Why non-compute constraints dominate as agents scale.
02
Measuring output quality
What 'useful output' means under latency, reliability, and safety pressure.
03
Agentic workloads
How acting systems stress observability differently from chat.
04
Hardware/software boundary
Where silicon observability lessons transfer to the platform layer.
05
Research direction
Concrete instrumentation patterns worth building first.

Planned

Planned framework, not yet published.

Public framing

This roadmap page frames a planned piece and remains intentionally non-confidential and public-source oriented.