Welcome to the Agent Platform Research Briefing for Tuesday, March 17th, 2026. Three big stories today — NVIDIA just delivered its GTC keynote, the Anthropic Pentagon fight is gaining new allies, and the MCP team finally put out a formal 2026 roadmap.

**NVIDIA GTC 2026: Vera Rubin, Groq 3 LPX, and NemoClaw** — Jensen Huang's GTC keynote on Monday was the year's most consequential hardware event for AI infrastructure. NVIDIA officially launched the Vera Rubin platform — the successor to Blackwell — comprising seven chips and five rack-scale systems. The flagship is the NVL72 rack, packing 72 Rubin GPUs and 36 Vera CPUs, delivering up to 10x higher inference throughput per watt compared to Blackwell. The new Vera CPU alone is twice as efficient and 50% faster than conventional server CPUs. Alongside Vera Rubin, NVIDIA formally introduced the Groq 3 LPX inference rack — 256 Groq LPUs with 128 gigabytes of on-chip SRAM and 40 petabytes per second of memory bandwidth, connected to Vera Rubin NVL72 racks via a dedicated 640 terabyte-per-second Spectrum-X interconnect. The design targets the decode bottleneck in agentic token workloads, where GPUs have always struggled. And in the biggest move for the OpenClaw ecosystem specifically: NVIDIA announced NemoClaw, an enterprise-grade OpenClaw stack that installs with a single command and adds OpenShell — a secure sandboxed runtime with policy-based guardrails for autonomous agents. Jensen framed OpenClaw as the "operating system of agentic computers" and said that — internally — 100% of NVIDIA engineers now use Claude Code, citing it as evidence of the agentic AI inflection point. Looking further out, NVIDIA teased the Feynman architecture targeting 2028, which will add a new GPU, a new LPU, and a new CPU called Rosa — named after Rosalind Franklin.

**LangChain and NVIDIA announce enterprise agent platform at GTC** — Timed to the keynote, LangChain announced a comprehensive enterprise partnership with NVIDIA, combining LangSmith, Deep Agents, and LangGraph with NVIDIA's NIM microservices, Nemotron open models, Dynamo 1.0 production inference, and OpenShell for sandboxed agent execution. LangChain also joined NVIDIA's Nemotron Coalition — a global initiative to advance frontier open models through shared data and compute. The pitch is direct: development teams have been spending months building custom infrastructure instead of shipping agents; this stack is meant to close that gap with a single integrated platform. Deep Agents specifically adds planning, virtual filesystem, context management, and persistent memory on top of LangGraph. LangChain says its open-source frameworks have surpassed one billion downloads. The combination of LangSmith observability and NVIDIA's hardware-level security layer is arguably the first serious enterprise answer to the "how do we actually deploy autonomous agents in production" question.

**Anthropic DoD fight: tech industry groups join the amicus filings** — The DC Circuit case is still grinding, but yesterday saw a notable escalation: Axios reports that tech industry groups representing hundreds of companies filed briefs urging the court to pause the Pentagon's supply-chain blacklisting of Anthropic. This is on top of the individual amicus filings from Microsoft, Google, Amazon, Apple, and OpenAI reported last week. No court ruling yet. On the commercial side, Anthropic appears to be turning the political adversity into a brand moment — the company extended its off-peak usage doubling promotion through March 27th, framed explicitly as a thank-you to customers. Ramp's March AI spending index, released last week, showed Anthropic overtaking OpenAI in enterprise spend with 70% head-to-head win rates — a direct effect of the Pentagon backlash driving a Claude loyalty wave.

**MCP publishes formal 2026 roadmap — shifts from release milestones to Working Groups** — The Model Context Protocol team published an official 2026 roadmap on March 9th that's worth flagging now that it's gotten wider attention. The headline change is structural: the roadmap is no longer organized around versioned releases with dates, but around priority areas driven by Working Groups. Lead maintainer David Soria Parra was explicit that release-milestone framing implied a predictability that open-standards work rarely has in practice. Priority areas include hardening production deployments, formalized authentication and authorization, better multi-agent composition patterns, and tooling for observability. The shift reflects MCP's maturation from a clever local-tools connector to infrastructure running at scale inside enterprises. The MCP spec itself hasn't had a new version since November 2025 — the working group model is meant to let protocol improvements flow faster and from more contributors simultaneously.

That's the briefing for Tuesday, March 17th. Big hardware week with GTC driving most of the signal.