SAN JOSE, CALIFORNIA — NVIDIA formally launched NemoClaw, its open-source enterprise AI agent platform, on Monday, March 16, 2026, during CEO Jensen Huang‘s keynote at the SAP Center in San Jose, positioning the product as a direct, security-hardened alternative to OpenClaw — the viral autonomous agent project that OpenAI acquired earlier this year, according to Wired’s reporting and confirmed live at GTC.
The launch signals something bigger than a product release. NVIDIA — for decades the dominant seller of chips that power AI — is now building software infrastructure designed to run on any hardware at all, including chips made by rivals AMD and Intel. The company that built its empire on silicon is now betting it can lock in enterprises through software instead.
NemoClaw Runs on Any Chip — Including Rivals
What mainstream coverage mostly skipped: NemoClaw is explicitly hardware-agnostic by design. Sources familiar with the platform confirmed to Wired that access will not be restricted to products using NVIDIA’s own chips — a deliberate positioning that trades short-term hardware revenue for long-term ecosystem control.
This directly mirrors the playbook that made CUDA an unbreakable developer moat in the GPU era. Data examined by reporters from pre-GTC technical briefings indicates NemoClaw layers atop NVIDIA’s existing NeMo models and NIM microservices, but the runtime itself remains accelerator-neutral. That distinction is critical for CIOs evaluating multi-cloud or hybrid infrastructure — and it is one that most enterprise software comparisons have so far buried or ignored entirely.
OpenClaw’s Security Gaps Drive Enterprise Demand
NemoClaw did not emerge in a vacuum. OpenClaw, originally known as Clawdbot and later Moltbot, became GitHub‘s fastest-growing open-source project in history, surging from 9,000 to 210,000+ stars within weeks before OpenAI acquired it and hired its creator. The speed of adoption, however, outpaced its security architecture.
Security researchers raised concerns about OpenClaw‘s lack of provenance controls, absent audit trails, and agent scripts that could bypass enterprise access policies — a set of vulnerabilities that made corporate IT teams deeply reluctant to deploy it at scale. NemoClaw directly addresses this: official records reviewed by reporters confirm the platform embeds signed skills, role-based access controls, and activity logging baked into every agent workflow.
CrowdStrike and NVIDIA have already demonstrated persistent security agents built on the platform that operate inside zero-trust architectures — a showcase that reportedly accelerated partnership discussions with Cisco, Salesforce, Adobe, and Google. Whether any of those partnerships have been formally signed remains unconfirmed; none of the companies responded to requests for comment before publication.
What NemoClaw Actually Does
The platform enables enterprises to deploy always-on AI agents that autonomously execute complex, multi-step tasks — not just respond to single prompts. At GTC 2026‘s Build-a-Claw workshops, running March 16–19 from 1 PM to 5 PM PT, attendees built functional agents in under an hour, naming them, assigning tool access, and deploying them through messaging apps.
NemoClaw‘s core capabilities, as confirmed through official GTC documentation reviewed by reporters:
- Always-on autonomous execution of multi-step enterprise tasks
- Enterprise system integration with internal files, tools, and databases
- Built-in governance frameworks, including policy checks and observability hooks before each action step
- Local deployment option on NVIDIA DGX Spark hardware, eliminating cloud dependency for sensitive workloads
- Deep customization of agent personality, name, and permitted tool access
The Fine Print Nobody Is Discussing
Here is what the launch headlines missed. NemoClaw is open-source — meaning partner companies like Salesforce and Cisco get access for free, in exchange for development contributions. But official documentation on independent security auditing of the platform’s encryption protocols, secrets rotation, and vulnerability disclosure policies has not yet been published.
Sources familiar with implementation said that CUDA optimizations will almost certainly favor NVIDIA hardware in real-world performance benchmarks, even if the runtime itself remains technically chip-agnostic. That gap between marketing language and operational reality is exactly the conversation enterprise procurement teams will be forced to have — and one that NVIDIA has not yet publicly addressed.
Jensen Huang Calls It “The Most Important Software Release Probably Ever”
Huang did not understate the moment. Speaking from the SAP Center stage before an audience of 30,000 attendees from 190 countries, he called the broader OpenClaw ecosystem — which NemoClaw extends — “the most important software release probably ever,” according to NVIDIA‘s live GTC 2026 updates blog.
That framing is aggressive, even for Huang. But the underlying logic holds: agentic AI represents a fundamental shift from AI that answers questions to AI that completes workflows. Every enterprise productivity stack — from Salesforce CRM to Adobe Creative Cloud — becomes a deployment surface. NVIDIA wants the agent runtime layer to be as dominant as its GPUs were to model training.
Whether enterprise security teams, still scarred by OpenClaw‘s audit failures, will trust a platform that is simultaneously open-source and governed by a chip company with obvious infrastructure incentives — that question remains loudly unanswered. At least one source familiar with the partner conversations went silent when asked about contractual data governance terms.
What Happens Next
GTC 2026 runs through Thursday, March 19. The Build-a-Claw developer workshops continue daily at GTC Park, San Jose, through the close of the conference. NVIDIA has not confirmed a standalone NemoClaw developer documentation release date, nor has it published a formal security disclosure timeline. The company’s next confirmed public milestone is the Vera Rubin GPU production ramp, slated for later in 2026 — the hardware layer that NemoClaw agents will ultimately run fastest on, hardware-agnostic positioning aside.
