Inference

Redis announced Redis Feature Form, a managed enterprise feature store platform for defining, orchestrating, versioning, and serving ML features across training and inference. The release adds multi-tenant workspaces, fine-grained job control, atomic DAG updates, enhanced RBAC/security, simplified deployment, and a redesigned dashboard.

SK hynix began mass production of 192GB SOCAMM2, an LPDDR5X low-power DRAM server memory module built on its 1cnm process. The company claims over double bandwidth and over 75% improved power efficiency versus RDIMM, and positions the module for NVIDIA Vera Rubin–based next-generation AI servers.

AI workloads require coordinating GPUs, storage, and apps, making networking a core factor in AI scalability and ROI via openness and standardization.

The post Platform Engineering for AI: Why Platform Teams Are the New Product Teams appeared first on Itential.

Dell’Oro says AI RAN is expanding across the RAN stack, expects $10B+ by 2029, and cites GPU-RAN constraints and GPU interest.

Teleport launched Beams, a runtime that runs AI agents in isolated Firecracker VMs with identity, policy-controlled access, and audit trails.

NVIDIA launched the Vera Rubin platform with seven new chips to support scalable AI infrastructure across all phases of AI processing.

NVIDIA introduced the BlueField-4 STX, a modular storage architecture enhancing token throughput and efficiency for agentic AI.

Crusoe expanded support for NVIDIA Vera CPUs and Nemotron models, integrating with Omniverse DSX and contributing to the NVIDIA Dynamo framework.

NVIDIA released Dynamo 1.0, an open source software designed for AI inference at scale, integrating with frameworks and supported by major cloud providers and enterprises.

Amazon Web Services will deploy Cerebras wafer-scale inference processors in its data centers under a multiyear agreement.

SONiC adoption reduces per-gigabit costs and vendor lock-in as refresh cycles shorten for AI and GPU workloads.