NVIDIA AI Updates: June 10, 2026

1. NVIDIA Confidential Computing to Help Expand Apple’s Private Cloud Compute

NVIDIA. NVIDIA announced that its Confidential Computing technology will support an expansion of Apple’s Private Cloud Compute beyond Apple’s own data centers. The deployment uses NVIDIA Blackwell GPUs with integrated security features to handle server-side inference for Apple Intelligence while maintaining privacy and security guarantees. The effort targets high-performance cloud inference for AI systems that combine on-device and cloud-based processing. Source

2. NVIDIA Adds Enterprise Manageability for DGX Spark Lifecycle Control

NVIDIA. NVIDIA introduced Enterprise Manageability capabilities for DGX Spark and GB10 systems, providing an operational framework that spans provisioning through end-of-life retirement. The solution uses agentless SSH execution with standardized JSON outputs and integrates with existing tools such as Ansible, Chef, Puppet, and Canonical Landscape rather than requiring dedicated management infrastructure. It supports air-gapped deployments and security compliance for managing AI infrastructure at scale. Source

3. NVIDIA Details Converting FP8 Checkpoints into TensorRT Inference Engines

NVIDIA. NVIDIA published a workflow for converting FP8-quantized model checkpoints into optimized TensorRT inference engines using ModelOpt and ONNX. The process exports quantized checkpoints to ONNX with QuantizeLinear and DequantizeLinear nodes, then compiles them into TensorRT engines that fuse those nodes into specialized FP8 kernels. On an RTX 6000 Ada GPU, the approach delivered roughly 1.39x to 1.45x lower inference latency versus FP16 while reducing model size by 34 to 50 percent. Source

4. NVIDIA FLARE Auto-FL Uses AI Agents to Accelerate Federated Learning Research

NVIDIA. NVIDIA introduced FLARE Auto-FL, an automated research framework that uses AI agents to accelerate federated learning experimentation. The system constrains agent actions through a control plane and fixed benchmark contracts, allowing agents to autonomously test federated learning strategies while preserving experimental reproducibility. It combines bounded mutation surfaces, experiment ledgers, and literature-grounded recovery mechanisms to help researchers evaluate more ideas more quickly. Source

5. NVIDIA Speeds Clinical ASR Evaluation with Agent Skills and Nemotron Speech

NVIDIA. NVIDIA described a workflow that pairs agent skills with NeMo Data Designer and Nemotron Speech to build pronunciation-aware synthetic clinical audio benchmarks without real patient data. The approach combines profile-driven benchmarks, pronunciation-aware text-to-speech, explicit review gates, and entity-level evaluation to stress-test speech recognition systems on medical terminology such as drug names and procedures. It aims to shorten clinical ASR validation cycles while maintaining phonetic accuracy through human review of pronunciation gaps. Source