HappyRobot Just Raised $150M to Prove AI Workers Beat AI Tools. The Numbers Back Them Up.

Seven weeks after the HuggingFace model repository breach exposed embedded payloads in open-source AI models, Nvidia launched the Open Secure AI Alliance on July 27, 2026 with 37 founding members and a security framework designed to address exactly the vulnerabilities the breach revealed.

The founding membership list is a who's who of enterprise infrastructure: Microsoft, IBM, CrowdStrike, Dell, Lenovo, Cisco, KPMG, and 30 additional companies spanning cybersecurity, enterprise software, cloud infrastructure, and professional services. The alliance also released the NOOA framework — Nvidia Open Optimization Architecture — under the Apache 2.0 open-source license as its first technical deliverable.

Three companies were conspicuously absent from the announcement: OpenAI, Google, and Anthropic.

The three most-deployed AI model providers in enterprise environments — the companies whose APIs underlie most of the AI features that enterprise software buyers are evaluating, contracting, and deploying — did not join the alliance that the rest of the enterprise AI infrastructure industry just organized around.

That absence is not a footnote. It is the story.

What OSAIA Is, and What It Actually Covers

OSAIA's stated mission is establishing shared security standards for AI systems. Its initial technical output — the NOOA framework — addresses four threat categories that the HuggingFace breach made impossible to ignore:

Model integrity. Verifying that AI models have not been tampered with between training and deployment. The HuggingFace breach demonstrated that models in open repositories could contain embedded payloads that activate at inference time. NOOA establishes cryptographic signing standards for model artifacts and verification protocols for enterprise deployment pipelines.

Inference security. Protecting against prompt injection and inference manipulation attacks. As AI systems are integrated into enterprise workflows with access to sensitive data and operational systems, adversarial inputs that hijack inference behavior represent a significant and largely unaddressed threat vector.

Data provenance. Establishing chain-of-custody for training and fine-tuning datasets. Enterprises fine-tuning foundation models on proprietary data need to verify that training data has not been poisoned — either by adversarial injection into data pipelines or by contamination from third-party datasets with unknown provenance.

Agentic operation boundaries. Governance controls for AI agents that can take autonomous actions on enterprise systems. An AI agent with access to enterprise email, calendar, code repositories, and customer data requires a different security posture than a chatbot interface with read-only context access. NOOA provides a framework for defining, enforcing, and auditing the operational boundaries of agentic AI systems.

The framework is released under Apache 2.0, which is a meaningful governance decision: any enterprise can adopt NOOA without licensing fees, any vendor can build NOOA-compliant tooling without commercial restrictions, and any third-party auditor can assess NOOA compliance against the published standard. Open-source licensing for a security framework creates the conditions for independent auditability that proprietary security frameworks cannot offer.

The 37 Members and What They Signal

The founding membership of OSAIA is not random. Reading the list reveals deliberate strategic logic.

Microsoft is the most significant member. Microsoft is both a major Nvidia customer (Azure GPU infrastructure) and a major competitor (Azure OpenAI Service, Copilot). Its participation signals that enterprise AI security is worth organizing around even for a company that competes with Nvidia on other dimensions.

CrowdStrike, Palo Alto Networks, and Cisco represent the enterprise cybersecurity infrastructure that will implement AI security standards at the network and endpoint level. Their participation means NOOA can potentially be integrated into the security tooling that enterprise CISOs already manage, rather than requiring new standalone AI security products.

KPMG and Deloitte (not publicly confirmed but consistent with their enterprise AI advisory practices) represent the professional services layer that will help enterprises assess, implement, and audit OSAIA compliance. Their participation is necessary for the alliance's standards to reach regulated industries where compliance is assessed by third-party auditors, not self-attestation.

Dell, Lenovo, and HP represent the hardware OEM layer — the companies that ship the servers and workstations that run on-premises AI deployments. Their participation ensures that NOOA can be integrated into hardware provisioning workflows, not just software deployment pipelines.

The pattern is consistent: OSAIA is the alliance of the AI deployment supply chain. It is organized around the companies that build the infrastructure enterprises use to run AI, not the companies that sell the AI models themselves.

Why OpenAI, Google, and Anthropic Aren't There

The three absent companies have not publicly explained their non-participation. Enterprise security observers and industry analysts who spoke to StorageNewsletter following the announcement have offered three structural explanations.

First, proprietary safety frameworks as competitive assets. Anthropic has invested substantially in Constitutional AI methodology. OpenAI has built safety teams, alignment research programs, and internal security infrastructure. Google DeepMind's safety division is one of its most publicly prominent research areas. All three have reasons to view their safety and security approaches as competitive differentiators — and reasons to be cautious about standardization that might level those differentiators.

An industry standard for AI security is beneficial to the overall market but may be disadvantageous to individual vendors whose safety frameworks are perceived as superior to the standard. Participating in OSAIA potentially means accepting that an externally governed standard — not your proprietary methodology — is the benchmark against which enterprise buyers assess your security posture.

Second, competitive positioning with Nvidia. OpenAI, Google, and Anthropic all compete with Nvidia for enterprise AI budget, though at different layers of the stack. Nvidia sells infrastructure; these companies sell model access. But as the AI market matures, the infrastructure layer and the model layer are increasingly in competition for the same enterprise budget allocation decisions: do you build your AI capabilities on foundation model APIs, or do you build them on on-premises GPU infrastructure running open-source or fine-tuned models? Participating in a Nvidia-led consortium sends a signal about which side of that question you are on.

Third, different threat models for API-accessed AI. The NOOA framework is initially focused on threats specific to on-premises AI deployment: model tampering in deployment pipelines, data poisoning in fine-tuning workflows, and agentic operation on infrastructure the enterprise directly controls. OpenAI's, Google's, and Anthropic's enterprise customers primarily access AI via API — they do not deploy OpenAI models on their own infrastructure. The threat model for API-accessed AI is different: it centers on prompt injection in application integrations, data leakage through prompts, and trust boundaries at the API layer rather than the infrastructure layer.

This threat model difference may be the most honest explanation for non-participation. OSAIA addresses the security problems of on-premises AI deployment. If most of your enterprise customers access your AI via API, OSAIA is addressing a problem set that is largely not your customers' problem set.

The Governance Split That Every Enterprise CISO Sees Differently

The OSAIA membership list and the absence from it create a two-tier security landscape that enterprise security teams are actively mapping.

Tier 1: Independently auditable. Enterprises deploying AI on Nvidia GPU infrastructure — on-premises or in colocated data centers — running NOOA-compliant systems can have their AI security posture independently audited against a published open-source standard. Third-party auditors — the KPMG and Deloitte members of OSAIA — can assess and certify compliance. This is the standard that regulated industries require.

Tier 2: Vendor-attested. Enterprises whose AI deployments are primarily cloud-based, accessed via OpenAI, Google, or Anthropic APIs, operate under those providers' internal security frameworks. OpenAI publishes security documentation; Google and Anthropic have published technical papers on their safety methodologies. But there is no external auditor who can assess and certify an enterprise's OpenAI API integration against an independently governed standard.

For most enterprise AI use cases — internal productivity tools, customer service chatbots, developer assistance — this distinction does not matter. The risk exposure is low enough that vendor self-attestation is adequate.

For regulated industries, it matters significantly. A healthcare organization using AI for clinical decision support, a financial services firm using AI for compliance review, or a defense contractor using AI for logistics planning will face questions from their auditors that OSAIA participation directly addresses and API-only AI deployments cannot.

The enterprise AI agent ROI payback period analysis found that regulated industry enterprises report 40% longer AI deployment timelines than unregulated peers — and that compliance review of AI security posture is the most commonly cited cause of delay. OSAIA creates a path to accelerate that compliance review for on-premises deployments; it does not help API-centric deployments.

What the Absent Members Are Building Instead

The three absent companies are not passive observers of the AI security governance discussion. They are building parallel frameworks.

OpenAI has published its Model Spec and maintains its Preparedness Framework — internal governance documents that define how OpenAI evaluates and responds to AI risks. These are not enterprise security standards in the NOOA sense; they are vendor safety policies. The distinction matters: a vendor safety policy governs what the vendor does; an enterprise security standard governs what the enterprise deployer does.

Google DeepMind's safety frameworks are primarily focused on model behavior — alignment, honesty, and harm avoidance — rather than enterprise deployment security in the NOOA sense. Google Cloud has its own AI security capabilities, but they are product features rather than open standards.

Anthropic has been the most active in the space closest to OSAIA's concerns: its Constitutional AI methodology addresses model behavior at the training level, and its published research on AI safety is substantial. But Anthropic has not, as of the OSAIA launch, published an enterprise AI security deployment standard analogous to NOOA.

The white house voluntary AI hacking tests framework, which OpenAI, Google, and Anthropic did participate in, covers a different dimension: red-teaming of AI models for dangerous capabilities (bioweapons assistance, CBRN information, cyberweapons generation). That framework addresses what the AI model can be made to do. OSAIA addresses how the AI deployment is secured. Both are necessary; they do not substitute for each other.

The Enterprise Buyer's Security Evaluation Framework

For enterprise security and procurement teams evaluating AI vendor commitments to security governance, the OSAIA launch creates a useful evaluation lens.

Question 1: Which security framework governs your enterprise AI deployment, and who can independently audit it?

If the vendor's answer is "our internal security policies," ask for documentation and references from enterprise customers who have completed third-party security audits of their deployment. If the vendor cannot provide references, the security posture is self-attested.

Question 2: What is your threat model for agentic AI specifically?

As AI agents that can take autonomous actions on enterprise infrastructure become more common, the security questions shift from "what data can the AI access" to "what actions can the AI take." Ask vendors how they enforce and audit the operational boundaries of agentic AI systems — and whether those boundaries are externally auditable.

Question 3: How does your security framework address the AI supply chain specifically?

The HuggingFace breach demonstrated that AI-specific supply chain attacks — model tampering, embedded payloads, data poisoning — require different defenses than traditional software supply chain security. Ask vendors which specific AI supply chain threat vectors their framework addresses.

Question 4: What is your OSAIA participation status or equivalent?

This question will reveal whether the vendor has joined OSAIA, is building a comparable framework, or is operating solely on internal security policies. None of these is automatically disqualifying — but the answer reveals the vendor's security governance posture in a way that is difficult to obscure.

The Standards War Coming

The OSAIA launch sets up a governance battle that the enterprise AI market will navigate for the next two to three years.

On one side: OSAIA and the Nvidia-led infrastructure ecosystem, pushing for open, independently auditable AI security standards. On the other side: the three largest model providers, operating under proprietary frameworks that have substantial safety investment behind them but are not governed by external industry standards bodies.

In mature technology markets, this dynamic typically resolves in one of three ways: the proprietary vendors join or align with the open standard (as happened with web standards, with software security standards, and with networking protocols), the proprietary vendors create a competing open standard that fragments the market for several years before one consolidates, or regulatory mandate forces convergence on an external standard that all parties must comply with.

The regulatory path is increasingly plausible. The EU AI Act's security requirements for high-risk AI systems create exactly the kind of third-party audit requirement that OSAIA is positioned to satisfy. If EU regulators explicitly reference OSAIA compliance as a pathway to AI Act conformity — a development that several OSAIA member companies are actively pursuing with Brussels — the three absent companies will face regulatory pressure to participate in or align with the standard.

The Rust LLM policy and open source AI governance debate illustrates how quickly open-source governance structures can become de facto standards when enterprise procurement requirements align with open-source auditability. The OSAIA-NOOA combination has the same structural logic: open-source licensing plus industry consortium governance plus regulated industry compliance requirements is the pattern that has successfully produced widely adopted standards in every previous enterprise software security context.

The enterprise buyer's position in this standards war is straightforward: favor deployments that can be independently audited, and build security evaluation criteria now that will remain durable regardless of which governance standard ultimately prevails.

Takeaway: Nvidia's OSAIA launch is not primarily a security story — it is a governance strategy story. The 37 founding members cover the enterprise AI deployment supply chain comprehensively; the three most-deployed AI model providers are not among them. That absence creates a two-tier security governance landscape that enterprise buyers must navigate: independently auditable on-premises deployments versus vendor-attested cloud API deployments. For most use cases, this distinction is academic. For regulated industries where third-party security audits are required, it is immediately actionable. Enterprise security teams should begin asking every AI vendor which security standard their deployment can be audited against — because the vendors who cannot answer that question clearly are the ones whose security posture is most difficult to evaluate.