Disaggregated Inference Added a Latency Tax. More Bandwidth Was Not the Fix.
A push-based connector in vLLM fixes a latency tax in disaggregated inference, where KV transfer setup stalls the decode worker.
2,100 GPU Servers Expose Their Metrics, and One Bug Can Switch Them Off
Researchers found 2,100 GPU servers exposing NVIDIA DCGM Exporter to the internet, plus a high-severity flaw that can crash GPU monitoring.
The Agent Sandbox Moved Out of the Model and Into the Operating System
Agent containment is now a Windows feature. MXC ships generally available with four isolation levels, from a process sandbox to a MicroVM.
The cgroup v2 Cutover Is Quiet Until a Limit Reads Wrong
In cgroup v2 in Kubernetes the limit still applies, but Java, Node and monitoring agents may read the host total instead of the pod.
Kubernetes Node Swap Triples Agent Density When Memory Is the Constraint
Kubernetes node swap hit GA in v1.34. Fresh benchmarks show it triples agent sandbox density on one node while memory prices keep climbing.
The Cold Start Was Never the Container. It Was the Model.
GKE Pod snapshots restore a loaded model in seconds instead of reloading weights. The catch is the compatibility key that decides which pod can restore.
Agents Broke Out of Their Own Sandboxes. DeepSeek Published the Catalog.
DeepSeek runs 3 million agent sandbox environments a day. Its new paper shows what breaks at that scale, and why an agent sandbox fleet is infrastructure.
NVIDIA Open Sources NodeWright to Patch GPU Fleets Without Killing Training Runs
NVIDIA open sourced NodeWright, a Kubernetes native tool that patches GPU node fleets without killing long training runs.
Docker Put an Agent’s Permissions Inside the Image It Runs
Docker's Sandbox Kit Spec packs an agent, its tools and a typed list of what it may reach into one OCI image. CNCF will govern the format.
The GPU Cluster Now Spans Two Data Centers, Not One
WhiteFiber Continuum joins two data centers 83 kilometers apart into one distributed GPU supercluster, pushing inference past a single campus.