OpenSharing Lands at Linux Foundation, Nemotron 3 Ultra Goes Open

<p><strong>Databricks and the Linux Foundation launched OpenSharing</strong> on June 10 — the first open protocol to cover agent skills, AI models, and unstructured data across platforms, potentially making agent skills as portable as Delta tables. Meanwhile, <strong>NVIDIA's Nemotron 3 Ultra</strong> (550B open-weight MoE) ships as the first US open mode...

Highlights

  • Databricks and the Linux Foundation launched OpenSharing on June 10 — the first vendor-neutral open protocol to cover agent skills, AI models, and unstructured data, replacing bespoke cross-org integrations with a single standardized API surface. (Databricks)
  • NVIDIA released Nemotron 3 Ultra (550B open-weight MoE, June 4) at Computex — built explicitly for long-running agentic workflows, delivering 300+ tokens/sec at just 55B active parameters per token. (NVIDIA GTC Blog)
  • The Databricks Data+AI Summit opens June 15 in San Francisco; pre-summit coverage confirms Lakebase, Agent Bricks MCP server support, and MLflow 3.0 agent tracing headlining the keynote, with OpenAI and Microsoft on stage. (ChatForest)
  • NVIDIA shipped AI Enterprise 5.0 alongside Computex, adding Blackwell Ultra support, agent-layer security controls, and updated MLOps governance tooling to the enterprise stack. (NVIDIA GTC Blog)

Key Signals

  1. 1 OpenSharing donates agent-skill portability to a neutral foundation June 10, 2026

    Before OpenSharing, sharing a fine-tuned agent skill or model artifact across organizations meant one-off pipelines or single-vendor marketplaces. The new protocol — donated to the Linux Foundation the same day Databricks announced it — standardizes discovery, authorization, and access APIs regardless of platform, including on-premises targets via storage partners MinIO and Qumulo. If adoption tracks the path Delta Sharing carved for tabular data, teams will publish a skill once and expose it to any compliant runtime. (Databricks / Linux Foundation)

  2. 2 Nemotron 3 Ultra raises the bar for open-weight agent models June 4, 2026

    The 550B-parameter hybrid Mamba-Transformer MoE runs at over 300 tokens per second with 55B active parameters per token — a cost profile that makes it the first US open model credibly worth benchmarking as an orchestrator layer. NVIDIA positioned it directly at workloads where open weights have historically fallen short: enterprise document RAG, research automation, multi-step planning, and iterative code execution. Open weights mean teams can self-host or fine-tune without committing to a proprietary inference endpoint. (Crypto Briefing)

  3. 3 Summit week will surface the next enterprise agent platform bets June 11 preview

    The Data+AI Summit (June 15–18, Moscone Center) brings Agent Bricks MCP integration, Lakebase, and MLflow 3.0 native agent tracing to 30,000+ attendees, with guest keynotes from OpenAI and Microsoft joining LangChain, LlamaIndex, Anthropic, and CrewAI in the session agenda. The lineup signals that Databricks is converging on the governed data and execution layer for the full agentic ecosystem, not just the analytics stack. (ChatForest / Databricks Blog)

Why It Matters / What To Watch

  1. 1 OpenSharing could make agent skills as portable as Delta tables
    • Watch whether AWS Bedrock, Google Vertex AI, and Azure AI Foundry announce OpenSharing support or route around it with proprietary registries — adoption speed here determines whether this becomes genuine interop or another preferred-vendor format. (The New Stack)
    • Teams with multi-vendor agent stacks should evaluate the protocol now: if skill discovery APIs land in connector ecosystems you already use, governance overhead drops considerably and cross-org skill reuse becomes practical. (Databricks)
  2. 2 Open frontier weights shift the inference cost and control calculus
    • Nemotron 3 Ultra is the first US open-weight model worth including in your multi-agent topology evaluation alongside proprietary frontier APIs — benchmark inference cost per agentic task, not just per token, before committing to an endpoint model for orchestration work. (Crypto Briefing)
    • NVIDIA AI Enterprise 5.0 bundles policy controls and audit tooling directly with the model serving stack — relevant for regulated industries that need governance at the inference layer, not just the application layer. (NVIDIA GTC Blog)

Quick Links (sources)