日本語 
Datadog just called out a problem most teams don’t want to admit. AI is getting expensive not because GPUs are rare, but because nobody really knows how they’re being used.
Their new GPU Monitoring product is basically about visibility. Not surface-level metrics, but actual clarity on which workloads are using which GPUs, who owns them, and whether they’re doing anything useful. One screen, full stack view. That sounds simple, but this is exactly where most teams struggle today.
Right now, companies over-allocate GPUs as a safety net. They lack the ability to distinguish between two different states of inactivity and two different states of equipment failure. They use additional hardware to solve their problem, which they describe as a method of scaling. That’s where the money burns.
Also Read: Mitsubishi Electric DI and Nutanix team up on private cloud
This shifts that equation. You can trace performance issues faster, spot underutilized resources, and decide whether you actually need more GPUs or just better allocation. It also pulls platform and ML teams onto the same page, which usually doesn’t happen.
The larger view shows this development fits into an established pattern. The production of artificial intelligence systems demands both operational discipline and cost control as essential requirements for successful implementation. Model accuracy is not the only requirement.
English