June 2026 Summaries

Therna Biosciences uses its RNA-Logix platform to apply AI-driven iterative design to RNA therapeutics, generating and ranking large numbers of mRNA, ASO, and siRNA candidates before laboratory testing feeds results back into subsequent model training. The company reported that its early single-cloud infrastructure created delays, high H100 GPU costs, unreliable notebook sessions, and low GPU utilization, while its AI lead also had to manage both research and infrastructure. By adopting Lightning AI, Therna says it gained access to multi-cloud GPU capacity within seconds, reduced compute costs by 60 to 80 percent, preserved development environments across sessions, and improved GPU utilization through optimization support. A notable example was the development of a Boltz-2-based application for structure prediction and ligand screening, which the company says was built and deployed in one day rather than the weeks such work previously required. Therna emphasizes that faster, cheaper, and more reproducible computing supports more rapid experimentation, although researchers continue to validate model outputs and oversee therapeutic development.

Shadow AI refers to AI compute, tools, and development environments used outside organizational visibility or control, often because approved infrastructure is too slow, inaccessible, nonpersistent, or incompatible with researchers’ workflows. While informal workarounds can initially help small teams move quickly, they become significant enterprise risks by obscuring costs, weakening data governance and security, fragmenting knowledge, and limiting collaboration. The piece argues that policies and stricter enforcement address symptoms rather than the underlying cause, potentially driving unauthorized activity further out of view. It presents faster, researcher-friendly approved infrastructure as the preferred solution and promotes Lightning AI as a platform offering rapid GPU access, persistent collaborative workspaces, lifecycle support, centralized cost visibility, and compliance-oriented controls, alongside proof-of-value trials for organizations evaluating whether it can replace unmanaged tools.

Kimi K2.5 and GLM 5 serverless Model APIs will be retired and disabled at 5 p.m. PT on June 11, 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, which is recommended for agentic coding and tool-use performance. Organizations requiring continued access to the retiring models can inquire about dedicated deployments.

Lightning AI has launched support for NVIDIA Nemotron 3 Ultra as a Day 0 partner, positioning the open-weight model as a reasoning and orchestration backbone for long-running, multi-agent systems. Nemotron 3 Ultra uses a 550-billion-parameter mixture-of-experts architecture with 55 billion active parameters and supports context windows of up to one million tokens, enabling planning, tool use, validation, and adaptation across complex workflows. The announcement argues that the model can reduce reliance on costly proprietary frontier models by offering lower agent-task costs, higher throughput than comparable open models, and control over model customization, data, and deployment. Through Lightning’s platform, developers can combine Ultra with specialized execution, speech, and safety models, optimize long-context inference, deploy across cloud, on-premises, or hybrid environments, and use enterprise governance and monitoring capabilities.