Lambdalabs
Day 27 after signup
Most AI researchers get two accelerators or fewer
Lambda is now at lambda.com. Plus what 200+ researchers told us and model cards under load.
Why it works
- Lead with a concise preview of the issue so readers can quickly see the range of topics.
- Group related stories under clear section headings and give each a short summary with a consistent read-more call to action.
- Use specific findings and figures to make research summaries concrete without requiring readers to open every article.
How it is built
- Monthly roundup introduction
- Company news spotlight
- Research findings
- Product benchmarking update
Email copy
This month: Lambda has a new address, lambda.com. We also asked 200+ researchers in the Lambda Research Grant Program what's holding them back, and 62% named compute their #1 blocker. At NVIDIA GTC Berlin, we shared what we're building together on secure, efficient AI infrastructure. Plus: model cards that show how a deployment behaves under load, Open Jarvis closing 77% of the gap between a 9B model and a frontier model, and a CVPR 2026 workshop on turning world-model video into robot actions. Spotlight / Lambda has a new address Starting today, Lambda lives at lambda.com. Your services, account, contracts, and support stay the same, and existing links and email addresses keep working. Here's what changes and how to tell an email is really from us. Read more Lambda with NVIDIA at GTC Berlin At NVIDIA GTC Berlin, we shared how we're building with NVIDIA on secure, efficient AI infrastructure. Read more Hot off the command line / What 200+ AI researchers told us Seven in 10 AI researchers can reliably reach two accelerators or fewer for a typical project. We surveyed 200+ researchers in the Lambda Research Grant Program on what they're building, what stands in their way, and what comes next. Compute topped the list, with 62% ranking it their #1 blocker. Read more Model cards that show how a deployment behaves under load Our model cards now show benchmarks from NVIDIA's AIPerf package run on minimal Lambda hardware. You can see how a deployment behaves at different user loads, what its interactivity looks like, and when it's time to scale up to more replicas. Read more Knowledge drop / Open Jarvis: making local LLMs work as agents Drop a 9B open-weight model into an agent system built for a frontier cloud model, and accuracy on PinchBench, an agent benchmark, falls from 96.0% to 62.3%. Same task, same tools. Retune the system around the local model, and accuracy recovers to 88.4%, closing 77% of the gap. Read more Can a world model's video become a robot's actions? Does a realistic video give a robot enough to act on? That question drove the World Models Meet Active Sensing (WMAS) workshop at CVPR 2026, which Lambda co-organized and sponsored. RoboWM-Bench, the Best Paper, found that a video world model can produce convincing footage and still fail once its predicted behavior is turned into robot actions. GEM-4D raised real-world manipulation success from 61% to 81% by keeping geometry consistent across generated frames. Read more Event.log() / Lambda at STAC Summit NYC At STAC Summit NYC, the conference for trading technology, Jessica Nicholson, PhD candidate and machine learning engineer, and Devina Jain, research engineer, spoke on "FSI at Lambda: Research at the edge, performance in production." Lambda was a platinum sponsor, and we launched a new page for financial services teams to go with it. Read more Lambda Inc., 2510 Zanker Road, San Jose, CA 95131, United States Unsubscribe Manage preferences