tensorfolk

Oracle makes Nvidia Vera Rubin NVL72 and standalone Vera CPU instances generally available on OCI

At Oracle AI World, Oracle and Nvidia said Vera Rubin NVL72 racks and standalone Vera CPU compute are generally available on OCI. The CPU instances target agent work such as tool calls, code execution and data retrieval, and they scale separately from GPUs through OCI's Kubernetes engine. Multi-gigawatt superclusters and HGX Rubin NVL8 servers have no dates yet.

Why it matters

Separating CPU-heavy agent orchestration from GPU inference can improve GPU utilization and cut costs for agentic workloads.

Source: Read the full story on Oracle

ocinvidiavera-rubincpuagents

More in Infrastructure & LLMOps

Infrastructure & LLMOps / Cloud AI Platforms (Bedrock, Azure, GCP)

Oracle will offer Google TPUs on OCI from 2027, the first hyperscaler outside Google to do so

Announced at Oracle AI World, the deal puts Google's Tensor Processing Units into Oracle Cloud Infrastructure next to its Nvidia, AMD and other accelerators, with capacity ramping over several years from 2027. Oracle supplies the power, cooling and networking; neither company has said which TPU generation or regions are included.

Why it matters: TPUs become a choice outside Google Cloud, giving teams another way to source training and inference compute.

Read on Google Cloud / Oracle
Infrastructure & LLMOps / Cloud AI Platforms (Bedrock, Azure, GCP)

Akamai opens betas for a managed inference service and a WebAssembly agent sandbox

Akamai launched Akamai Inference in beta. It serves 12 open-weight models, from 8B up to 2.8T parameters and including Nemotron 3 Ultra, gpt-oss-120b and Gemma 4, through OpenAI-compatible APIs on serverless or dedicated deployments. Alongside it, Akamai Agent Sandbox uses the open-source Spin framework's WebAssembly isolation to limit what agent-written code can access.

Why it matters: Adds an edge-network option for hosting open models, plus a lightweight way to contain untrusted code written by agents.

Read on Akamai
Infrastructure & LLMOps / Cloud AI Platforms (Bedrock, Azure, GCP)

Google Cloud announces Agent Compute Engine, a microVM per agent session

Google unveiled a service that gives every agent session its own hardware-isolated microVM on Titanium, with quick checkpoint and restore and credentials kept outside the guest. It also open-sourced Agent Substrate for GKE, usable now, and extended its agent gateway and identity controls beyond Google Cloud. The managed service reaches preview in the coming months.

Why it matters: Agent sandboxing is becoming a first-class cloud primitive with stronger isolation than containers.

Read on Google Cloud