CoreWeave just made multi-cloud AI training work the way enterprises always wished it would

Running AI training jobs across multiple cloud providers has never been clean work. Different scheduling systems, unpredictable data movement costs, and operational models that shift depending on where the compute sits have made multi-cloud feel more like a compromise than a strategy. CoreWeave just shipped a set of updates that push back against all of that at once.

SUNK Anywhere headlines the release, expanding the company’s Slurm-on-Kubernetes platform across CoreWeave, Google Cloud, AWS, and Azure through one unified control plane. The practical outcome is straightforward: engineering teams stop managing separate scheduling and scaling systems for each cloud and start working from a single consistent model instead. For teams running large distributed training jobs, that shift alone removes a category of operational headaches that tends to quietly drain time and introduce avoidable mistakes.

The platform also tackles data movement. CoreWeave introduced LOTA Cross-Cloud, a caching extension that lets data stay wherever it already lives while still delivering near-local throughput of around 7 GB per second per GPU to compute resources in other environments. Egress fees have long made genuine multi-cloud setups economically messy at training scale. Paired with CoreWeave’s Zero Egress Migration program, this approach takes direct aim at those costs rather than working around them.

The Weights and Biases integration also got meaningful attention. New additions include support for Google’s Gemini CLI, access to Gemma models through W&B Inference, and expanded telemetry now covering TPU utilization. Developers running workflows across several providers have historically struggled with visibility gaps that make performance tuning harder than it should be. That added instrumentation fills in a real gap.

Observers in the industry point to Google Cloud’s involvement as a signal worth noting separately. A hyperscaler actively partnering on multi-cloud tooling rather than competing against it suggests the market is settling around interoperability as a feature rather than a threat. For enterprise AI teams that have spent years managing the messiness of distributed infrastructure, that shift in posture from major providers changes what becomes possible going forward.