September 2026 Summaries

Lightning’s September 2026 update highlighted Pangram Labs’ use of its infrastructure to develop the Pangram 4 AI-content detection model, which reportedly saved three engineering years, while also previewing the company’s attendance at NVIDIA GTC in Berlin in October. The company described ongoing datacenter upgrades from NVIDIA H100 to H200 systems, support for Therna Biosciences’ Chronos RNA dataset and AI RNA-Biologist platform through a ready-to-use Lightning Studio, and a Business Insider discussion about specialized AI models trained on proprietary data. Lightning also announced the appointments of Betty Junod as chief marketing officer and Heather Paduck as chief financial officer, reported 16-fold quarter-over-quarter cloud growth and 240% growth among early customers, promoted cloud features such as immediate H100 access and multi-node placement, and invited readers to its San Francisco Tech Week event.

Pangram, an AI detection startup founded shortly after ChatGPT’s release, develops transformer-based models designed to distinguish human-written text from machine-generated content and has expanded into AI image detection. Initially, its three-person team managed AWS GPU infrastructure with Slurm, but quota delays, file-system problems, and configuration complexity consumed engineering time, prompting a shift to Lightning AI Studios for on-demand and reserved GPU capacity. The company reports scaling from training on one A10G GPU to clusters of eight H100 GPUs and leasing dozens more, while growing to 25 employees without creating a dedicated infrastructure team. Pangram says Lightning helped save an estimated two to three engineering years and reduced compute costs by 50–60% compared with equivalent AWS capacity, while supporting data preparation, experimentation, evaluation, and internal demos. In July 2026, Pangram raised $9 million led by Menlo Ventures, bringing total funding to nearly $13 million, alongside the release of Pangram 4 and a research-preview image detector.

DeepSeek V4 Pro serverless Model APIs will be retired and disabled at 5 p.m. PT on September 23, after which requests will return HTTP 400 errors. Users are advised to update integrations before the deadline to use DeepSeek V4.1 Flash, identified as offering stronger performance at a lower cost, or GLM 5.3 for frontier-level performance. Organizations requiring continued access to DeepSeek V4 Pro can contact the provider to discuss a dedicated deployment.

PyTorch Lightning has improved its cloud-storage performance through fsspec, the common filesystem interface used across providers such as Google Cloud Storage, AWS, and Azure, meaning the underlying fixes can benefit multiple storage backends. In collaboration with Google Cloud, the project integrated Rapid Buckets through GCSFS, replacing legacy REST-based behavior with persistent bidirectional gRPC streams and requiring users only to change their bucket target. The update changes checkpoint writing from a buffer-copy-heavy write() approach to pipe(), which reads file chunks directly and reportedly reduces checkpoint write times by up to 81% on standard Google Cloud Storage buckets and 95% on Rapid Buckets. It also adds object-storage support for Fully Sharded Data Parallel checkpointing through PyTorch Distributed Checkpoint’s fsspec readers and writers, eliminating the need to stage checkpoints on local disks, while new profiler instrumentation exposes training dataloader initialization and prefetch delays. Joint testing cited a 23% reduction in data-loading time on Rapid Buckets and an additional 70% reduction in checkpoint writing beyond the pipe() improvements, with the relevant features available in GCSFS 2026.8.0 or later and PyTorch Lightning 2.6.5 or later.

GPT-OSS 20B and GPT-OSS 120B serverless Model APIs will be retired and disabled at 5 p.m. PT on September 15, after which requests to either model will return HTTP 400 errors. Users are advised to update integrations before the deadline to use lightning-ai/Qwen3.8-27B, a replacement model positioned for strong agentic coding and tool-use performance. Organizations requiring continued access to the retired GPT-OSS models can contact the provider to discuss a dedicated deployment.