Closed AI tools blocked Hugging Face's own security team from finishing a forensic sweep of its own breach this month. The company switched to an open-weight model, GLM 5.2, running on its own infrastructure, and cleared 17,000 flagged actions in the time it took the closed system to keep asking for permission. NVIDIA turned that story into the founding argument for a new alliance, and persuaded 36 other companies to co-sign it.

The stack argument, not the model argument

NVIDIA's pitch to the 36 companies wasn't about which model is smarter. It was about which model a defender can actually open during an active breach. An AI agent is identity, permissions, a harness, guardrails, logs, and an evaluation loop wrapped around a model. Lock any one of those five layers behind a vendor's API and you've reintroduced the exact bottleneck that slowed Hugging Face down.

That framing produced a real list of named contributions alongside the member roster. HPE brought SPIFFE and SPIRE, zero-trust identity standards that cryptographically verify an AI agent before it touches enterprise data. Hugging Face handed Safetensors, its model-weight storage format built to block remote code execution, over to the PyTorch Foundation. Microsoft supplied MDASH, a multi-model scanning harness that sets several AI agents loose on a codebase to find and prove exploitable bugs against each other. IBM and Red Hat paired up on Lightwell, which signs open-source patches to secure the supply chain those patches move through.

NVIDIA is the one company here with two decades of receipts

NVIDIA's own contribution, the NVIDIA Labs Object-Oriented Agent framework, went up on GitHub under an Apache 2.0 license the same day as the announcement. NOOA is built to make agent behavior easier to trace and audit, which is a description that could apply to half the tools in this alliance. What sets NVIDIA apart is what it was already running before this week. Megatron Core and the NeMo framework, both open source, train large language models across thousand-node GPU clusters and already anchor a large share of the industry's own model-training pipelines. NCCL, NVIDIA's open communication library, moves the data between those GPUs. NemoClaw, released separately, adds privacy and security controls to OpenClaw, another alliance member, which tells you these companies were already building on each other before anyone called it an alliance.

NVIDIA's argument to policymakers rests on that history. A company can say open tools are safer in a press release. NVIDIA can point to a decade of GPU infrastructure that already runs that way.

What's new work and what's an existing project under a shared banner

Look closer at the contribution list and the pooling becomes concrete. Microsoft's MDASH first appeared on Microsoft's own security blog in May, two months before this alliance existed. SPIFFE and SPIRE have been Cloud Native Computing Foundation projects for years. Safetensors has been Hugging Face's standard model format since 2022. NOOA and Lightwell are the pieces built specifically for this launch.

That mix is the point. A defender doesn't need every tool in the stack invented from scratch this month. They need mature, battle-tested projects and brand-new ones sitting under one roof, backed by a shared argument to regulators that open tooling belongs in the defensive column.

The absent three

OpenAI, Anthropic, and Google are not on the member list. Each runs a closed frontier model business that competes directly with the open-weight argument this alliance is built on, so the absence tracks their commercial position more than any technical disagreement. Meta, which had just co-signed a separate open-weights advocacy letter alongside several of these same companies days earlier, is also missing from this specific roster. NVIDIA hasn't said why, and no member company has offered a public explanation either.

Every company on this list has a commercial stake in open tooling winning the argument. The case still has to hold up on its own technical merits.

What happens next

Two questions will separate a working alliance from a press release with staying power. Does the coalition publish shared governance, maintainers, and a release process for joint work, rather than each member continuing to ship independently under a common banner? And when the next AI-driven security incident hits a member company, does that company reach for an alliance tool, or does it quietly fall back to whatever closed vendor it was already using?