The disconnect is wild: only about 21% of these orgs have AI running in production at scale, but nearly half (45%) are already looking to move into AI-specialized clouds. It feels like a massive spending spree without a dashboard. Even worse, fewer than half can actually tell you what their compute is costing them on a unit basis.
A few technical takeaways that stood out to me:
Vendor Churn: 64% plan to switch or add providers within a year. This is huge for something as foundational as infra.
Decision Drivers: Only 8% care about the headline price per million tokens. Most are prioritizing integration (41%) and TCO (35%).
The Blind Spot: Memory bandwidth is becoming the real bottleneck for inference scaling, but about 20% of enterprises aren't even tracking it yet.
I'm trying to figure out how to better implement a real-world AI workflow that doesn't just burn credits or leave GPUs idling. If you've managed a deployment, how are you actually tracking unit economics? Are you using specific observability tools, or is it just a guessing game based on the monthly cloud bill?
For anyone starting from scratch, focusing on TCO over token price seems to be the move, but the lack of visibility in the industry is a red flag.
[Next Multimodal RAG chunking is a mess i →](/en/threads/1128/)