July 2026 Summaries

In July, Lightning announced the launch and expansion of Lightning Cloud, a managed AI training platform offering on-demand H100 access, high-performance storage, optimized data loading, and multi-node training in early access. The company added smaller GPU options including T4s, L4s, and RTX Pro 6000s for testing, reported interconnect performance exceeding bare-metal speeds, and began onboarding enterprise and research customers. Its infrastructure footprint also grew through a GB300 cluster deployed in Chicago for Reflection.ai and planned capacity in Colorado Springs and Vancouver, including facilities designed for NVIDIA’s Vera Rubin platform. The update emphasizes the value of self-hosted open-weight models for handling sensitive enterprise data, citing an incident in which Hugging Face reportedly used a self-hosted GLM 5.2 model to investigate a security breach without sending evidence externally. Lightning also highlighted customer use by Archetype AI, new data-center capacity, and support for running Anthropic’s Claude Fable 5 through its platform.

Following the merger of Lightning AI and Voltage Park, the combined organization describes building a vertically integrated GPU cloud by controlling its hardware, virtualization, networking, and scheduling layers together. Its approach standardizes H100 nodes through fixed images and detailed validation, hardens Cloud Hypervisor through internal testing and upstream open-source fixes, and exposes real NUMA, CPU, GPU, and NIC topology to guest VMs to improve workload placement and GPUDirect RDMA performance. The company reports that these changes raised host-to-device bandwidth from uneven results of roughly 29–43 GB/s to about 48–49 GB/s across GPUs, compared with about 26 GB/s in its cited GCP test. Its scheduler uses known physical network topology to place multi-node jobs on well-connected machines, preserve contiguous capacity for large training runs, reclaim preemptible spot capacity when needed, and handle concurrent requests efficiently. Lightning AI says the resulting platform operates more than 36,000 AI-focused GPUs, provides managed VMs with near-bare-metal NCCL performance, supports multiple external cloud providers, and benefits from a small cross-disciplinary engineering team working across data centers, hypervisors, schedulers, and developer environments.

Cursor partnered with Lightning AI, NVIDIA, and Dell to design customized NVIDIA HGX B200 SuperPODs for the reinforcement-learning workloads behind Composer, Cursor’s agentic coding model. The collaboration began nearly two years earlier with 128 HGX H100 GPUs and expanded into multiple InfiniBand-connected SuperPODs spanning Hopper and Blackwell hardware generations. Cursor sought bare-metal infrastructure that enables rapid experimentation, direct control over scheduling and runtime behavior, detailed debugging, and freedom from provider-defined software stacks that can restrict observability or performance. Lightning AI owns and operates the systems, providing monitoring, engineering support, and security practices supported by certifications including ISO 27001, SOC 2 Type II, and HIPAA. According to the account, the teams jointly designed and deployed the B200 training system in under three months, establishing a model for customer-specific AI infrastructure intended to support evolving reinforcement-learning research.

Lightning AI has launched the Lightning Cloud, a vertically integrated AI infrastructure platform built on company-owned H100 GPUs and designed to provide applied AI teams with guaranteed on-demand and spot capacity for experimentation and training. The company argues that owning and operating the full stack, including data center hardware, networking, virtualization, scheduling, and development tools, avoids the provisioning, topology, performance, and visibility issues associated with platforms that layer software over rented infrastructure. Lightning Cloud offers instant H100 access, topology-aware placement for multi-node workloads, near-bare-metal NCCL performance, and enterprise-oriented reliability, while its fleet is also expanding to B200 and GB300 systems. The company positions the service as an alternative to long-term reservations and costly hyperscaler offerings, with pricing beginning at $4.50 per hour under a one-year commitment, compared with a cited $7.90 per hour elsewhere. Archetype AI is among the early users, training its Newton foundation model for understanding physical-world data on the platform, and Lightning AI is onboarding additional teams seeking dependable AI compute.

MiniMax M2.5 and Nemotron 3 Ultra serverless Model APIs will be retired and disabled at 5 p.m. PT on July 13, after which requests to either model will return HTTP 400 errors. Users are advised to update integrations before the deadline to use lightning-ai/deepseek-v4-pro, or DeepSeek V4 Pro, which is recommended for agentic coding and tool-use workloads. Organizations requiring continued access to the retired models may contact the provider to discuss dedicated deployments.