Inference

Alluxio introduced a distributed caching solution for AI workloads on Oracle Cloud Infrastructure, citing sub-millisecond access and up to 1 TB/s throughput.

Weekly roundup covering exploit telemetry automation, AI governance integrations, telco network modernization, and microsegmentation forecasts.

ABI Research forecasted AI inference workloads to surpass training by 2033 and reach more than 46 GW by 2035.

Nokia launched its AI Networking Innovation Lab in Sunnyvale to co-develop and validate AI-ready data center networking with partners.

MinIO introduced MemKV, a context memory store for AI inference on NVIDIA BlueField-4 STX to reduce context loss.

The author previews a book and blog series arguing AI performance depends on infrastructure layers like data movement, network design, and telemetry.

Rackspace Technology and AMD signed an MOU outlining a multiyear partnership to build a governed Enterprise AI Cloud for regulated and sovereign workloads. The proposal integrates AMD Instinct GPUs and EPYC CPUs into a fully managed stack, covering private/hybrid deployment, inference runtime, and managed inference services with defined SLAs.

Aviz Podcast Episode 2 discusses AI factories and AI fabrics, arguing networks must support multiple fabrics for training and inference with cost models like token serving.

Hedgehog contributed OCP Accepted AI training and inference fabric reference architectures to the OCP Marketplace, available for deployment with Ethernet-based disaggregated networks.

Aviz Podcast Episode 2 discusses building AI factories and AI fabrics, arguing networking must support multiple workloads and open standards.

Quantum Computing Inc. said NeuraWave reached deployment-ready status for real-time AI inference at the edge, using a server PCIe plug-in card.

AI traffic is increasing demand on switching fabrics and shortening network refresh cycles from 5–7 years to about 3.