MinIO and NVIDIA Forge a Secure Data Path for AI Factories

REDWOOD CITY, CA – June 01, 2026 – In a move set to redefine the data foundation for enterprise artificial intelligence, MinIO today announced a landmark collaboration with NVIDIA. Revealed at COMPUTEX 2026 in Taipei, MinIO’s AIStor and MemKV solutions will now natively support the NVIDIA Vera BlueField-4 STX security architecture, creating what the companies describe as a "secure-by-design" data path for the burgeoning era of agentic AI.

The partnership aims to address one of the most critical and complex challenges in modern enterprise technology: ensuring the data that fuels autonomous AI systems is as secure and performant as the silicon it runs on.

The New Security Imperative in the Age of Agentic AI

The concept of the "AI factory" has rapidly moved from theory to reality. In these advanced environments, autonomous AI agents are designed to continuously reason, retrieve data, and act on behalf of the enterprise, often without direct human intervention. While this promises unprecedented efficiency and innovation, it also introduces a dramatically expanded attack surface. Data is no longer a passive asset in a repository; it is the active, life-sustaining fuel for these intelligent agents.

This shift makes the storage layer the most consequential and exposed component in the entire AI data path. Critical assets—including model weights, vast training datasets, inference context, and the very memory of the AI agents—all flow through the ubiquitous S3 object storage interface. Industry experts have warned that a security compromise at this layer would not remain isolated. Instead, it could propagate silently and catastrophically through autonomous agent actions, degrading model reliability, corrupting business decisions, and eroding operational trust across the entire organization.

Enterprises building these AI factories face a daunting task: implementing a zero-trust architecture not just for human users, but for the AI agents themselves. This requires continuous, inline security enforcement that can validate every data access request at line speed, a challenge that traditional, CPU-bound security solutions struggle to meet.

Forging a Silicon-to-Storage Shield

The collaboration between MinIO and NVIDIA directly confronts this security challenge by embedding trust at the lowest level of the stack. The solution combines MinIO’s high-performance object storage, AIStor, with NVIDIA's Vera BlueField-4 STX, a data processing unit (DPU) based architecture, and its associated NVIDIA DOCA Vault software.

NVIDIA’s BlueField-4 STX is engineered to offload, accelerate, and isolate data center infrastructure services from the main server CPU. The DOCA Vault microservice, running on the BlueField-4 silicon, acts as an inline security sentinel. It enforces granular, zero-trust file access policies directly at the data path, independent of the host operating system. This means that even if a host server is compromised, DOCA Vault can continue to enforce rules, ensuring only authorized AI workloads can access the correct files with the right permissions.

By integrating AIStor with this architecture, MinIO extends that silicon-level enforcement to the S3 object layer where enterprise AI data actually lives.

“The AI factory runs on data, and that data lives in object storage,” said AB Periasamy, co-founder and CEO of MinIO. “AIStor and MemKV form the only complete, secure data path from persistent object storage to in-context GPU memory. With NVIDIA DOCA and NVIDIA Vera BlueField-4 STX, every byte moving through that path is protected in silicon.”

This creates a unified secure data fabric, establishing a chain of trust that extends from the hardware all the way to the application data.

“In the agentic AI factory, data is the foundation of intelligence, making security within object storage essential to protecting the integrity, privacy, and value of AI-driven outcomes,” said Jason Hardy, vice president, Storage Technologies at NVIDIA. “MinIO AIStor, NVIDIA DOCA Vault, and NVIDIA Vera BlueField-4 STX together establish a secure data fabric for agentic AI factories, enabling enterprises and sovereign clouds to safeguard critical data while scaling trusted AI with performance and efficiency.”

Eliminating the GPU Performance Bottleneck

While security is paramount, the defining performance challenge of the AI factory is keeping its enormously expensive GPUs fully utilized. "GPU starvation"—a state where GPUs sit idle waiting for data—is a primary cause of inefficiency and a major barrier to AI ROI. As network speeds advance to 800GbE and beyond, the traditional data path, which relies on the host CPU to process and move data, becomes a significant bottleneck.

The integrated MinIO and NVIDIA solution is engineered to dismantle this bottleneck. By leveraging the BlueField-4 STX DPU, object data can be delivered from MinIO’s storage directly to the GPU over RDMA (Remote Direct Memory Access) at wire speed, completely bypassing the host CPU. This frees up valuable CPU cycles for critical tasks like orchestration, scheduling, and inference logic, while ensuring the data pipeline can keep pace with the GPUs' voracious appetite.

Further enhancing performance is MemKV, MinIO's purpose-built context memory store. In production inference, especially with large language models that rely on long context windows, GPUs can spend significant time on the "recompute tax"—recalculating or re-fetching contextual data from slower storage tiers. MemKV addresses this by turning petabytes of NVMe flash into a shared, high-performance context tier, likely leveraging the NVIDIA CMX (Context Memory Storage) platform. This ensures that context is readily available, dramatically reducing GPU idle time and maximizing the efficiency of inference operations.

Unifying Data for Coherent, Multi-Agent Systems

The complexity of modern AI doesn't stop at security and performance. Sophisticated multi-agent systems must reason over vast quantities of unstructured data, like documents and images, while simultaneously grounding their decisions in structured data that reflects business logic and operational state. Historically, this has required siloed implementations—an object store for unstructured data and a separate database for structured data—which expands the attack surface and complicates data governance.

MinIO is tackling this challenge with AIStor Tables, a feature built on the open-source Apache Iceberg V3 table format. This allows AIStor to natively handle both structured and unstructured data through a single, unified S3 interface and governance model. By providing transactional consistency, schema evolution, and time-travel capabilities for massive datasets, Apache Iceberg V3 brings database-like reliability to the data lake.

This unification allows enterprises to build more coherent and robust AI systems. Autonomous agents can seamlessly access and reason across all relevant data types through a single, secure, and performant interface, eliminating the architectural complexity and security risks of bolting separate systems together. This integrated approach is a critical step toward building AI factories that are not only powerful and secure but also manageable and scalable.

The joint solution from MinIO and NVIDIA represents a significant step forward in building the foundational infrastructure for the next generation of AI. By addressing the core challenges of security, performance, and data complexity in a single, cohesive architecture, they are providing enterprises with the tools needed to build and scale trusted, efficient, and intelligent AI factories.