The product, unveiled alongside AMD Advancing AI 2026, targets governed agentic AI deployment with AMD EPYC CPUs, AMD Instinct GPUs and Red Hat OpenShift validation.
Embedded LLM has launched TokenVisor Spaces, a new product for AMD-powered AI clouds and enterprise deployments that packages agent execution, model inference, policy controls and auditable workspaces into a governed agentic AI service. Introduced in step with AMD Advancing AI 2026, the offering is built to let AI cloud operators and enterprises run long-lived AI agents on AMD server infrastructure, with execution on AMD EPYC x86 CPUs and inference on AMD Instinct GPUs. The company says the product is meant to address a practical gap in enterprise AI deployment: agents need persistent workspaces, tool access, approvals, policy enforcement, metering and audit trails, not just raw token generation. TokenVisor Spaces provides isolated containers for shell access, files, code, browser automation and tools, along with human approval workflows, replayable event histories and governed model access through TokenVisor. Embedded LLM said the product has been validated on Red Hat OpenShift (Kubernetes-based container platform), where OpenShift manages containers and GPUs while TokenVisor handles entitlements, routing, budgets, rate limits, metering and auditability. Embedded LLM also highlighted the economics of KV-cache reuse (reusing stored model context to avoid recomputing prompts) for long-running agents. Citing SemiAnalysis, it said that across 1.5M+ Claude Code requests, roughly 95% of all tokens were cache reads, reducing the prompt-token bill by about 84%. In a separate benchmark on AMD Instinct MI355X GPUs using vLLM, LMCache ConnectorV1, AMD hipFile/GDS and native local NVMe, the company reported 3.31x lower warm-turn median latency and 2.23x faster total wall-clock time versus a matched vLLM HBM prefix-cache baseline. The launch comes with collaborations involving VAST Data and Tensormesh around platform-scale KV-cache reuse, agent state, trace capture, replay and evaluation, and RL (reinforcement learning) data. TokenVisor Spaces is available for partner deployment and evaluation by AI cloud operators, private AI environments and on-premises enterprises.