NVIDIA
·
5 days ago
NVIDIA introduced the Jetson T3000 and T2000 modules based on its Thor architecture to enable robotics and edge AI applications, with the T3000 delivering 865 FP4 teraflops of compute in a compact form factor half the size of the T5000. The T3000 combines a Blackwell GPU with 32GB memory and 273GB/s bandwidth, while the T2000 offers 400 FP4 teraflops with 16GB memory, available in Q1 2027. Companies like Boston Dynamics and Amazon Robotics can now deploy foundation models on robots more efficiently, with customers like Agile Robots achieving up to 15GB memory savings through software optimization.
The New Stack
·
6 days ago
● 2 sources
Google released GKE Agent Sandbox to general availability in May 2026 and introduced Agent Substrate, a separate scheduler designed for AI agents because Kubernetes was built for long-running services rather than the bursty, mostly-idle workloads that agents represent. Agent Substrate achieves 30x or more oversubscription with sub-second activation by snapshotting idle sessions to storage and multiplexing them onto pre-warmed worker pods, while Agent Sandbox provides kernel isolation via gVisor for running untrusted model-generated code at 300 sandboxes per second with 90 percent finishing in 200 milliseconds. The emergence of this dedicated agent runtime as a fourth compute offering alongside virtual machines, containers, and serverless indicates that the infrastructure layer for managing agents is consolidating into specialized systems separate from general container orchestration.
The New Stack
·
6 days ago
● 2 sources
Meta is building a cloud business to sell excess GPU capacity through its Meta Compute initiative, choosing between offering hosted AI models or raw compute rental, while the shoe company Allbirds pivoted to selling GPU-as-a-Service after a $39 million sale of its footwear brand. Meta reported the compute business on July 1, and CoreWeave and Nebius stocks dropped double digits after the announcement, with Nebius holding a $27 billion contract with Meta. As GPU supply fragments across multiple providers including hyperscalers, neoclouds, and accidental clouds, enterprises will gain leverage through abstraction layers that treat suppliers as interchangeable provisioning targets rather than betting on single vendors.
Ars Technica
·
6 days ago
SpaceX plans to deploy a constellation of 1 million satellites in orbit to serve as data centers, shifting its business focus away from rockets toward hosting artificial intelligence computing infrastructure. The company aims to power up to 100 million frontier-class GPUs across these orbital facilities, with individual AI1 satellites designed to generate 120 GW of power. This orbital data center approach would allow SpaceX to offer computing services directly from space rather than relying on ground-based infrastructure.
Together AI
·
6 days ago
Together AI released updates to its GPU Clusters platform including passive health checks that monitor running workloads for failures like GPU bus drops and thermal throttling, auto node repair with human approval, and a rebuilt Slurm-on-Kubernetes stack addressing daemon crashes and process cleanup. The platform added operational features including a redesigned cluster overview dashboard showing health and utilization, external OIDC authentication for per-user Kubernetes access, and startup scripts for self-serve node customization. These changes reduce incident resolution time from hours to minutes and enable teams to manage clusters at scale without sharing admin credentials or performing manual node setup.
