Lambda

Lambda has closed its second institutional debt facility and its first $1 billion-plus GPU debt financing. The facility was oversubscribed and rated A (low) by Morningstar DBRS and Baa1 by Moody's. Proceeds will fund the purchase and development of GPU infrastructure supporting three committed customer deployments with two investment-grade offtakers across multiple data centers. Read the announcement: https://lnkd.in/gtgdpsTX

A realistic video is not enough for a robot to act reliably. That was the central question at CVPR 2026’s WMAS workshop. RoboWM-Bench, the Best Paper, found that a video world model can produce convincing footage while still failing when its predicted behavior is turned into robot actions. The failures came from problems such as spatial reasoning and unstable contact. GEM-4D improved real-world manipulation success from 61% to 81% by adding geometric consistency across generated frames. SAW-Bench found a 37.66-point gap between humans and the best multimodal model on situated-awareness tasks. Lambda co-organized and sponsored the workshop. The program included 20 accepted papers and invited talks from Nicholas Roy, Alan Yuille, Yiannis Aloimonos, and Chelsea Finn. Lambda also provided compute credits for the awards. Read the workshop insights: https://lnkd.in/e589mxvm

Drop a 9B open-weight model into an agent system built for a frontier cloud model and accuracy falls from 96.0% to 62.3% on PinchBench. Same task. Same tools. Retune the system around the local model. Change the quantization, reasoning loop, tool access, and memory. Accuracy recovers to 88.4%, closing 77% of the gap. The configuration was the limiting factor. Open source, evaluated on Lambda GPUs. The original setup was optimized for a different model. Open Jarvis separates the stack into five parts. Intelligence covers the model and quantization. Engine covers the runtime and hardware. Agent logic governs prompting and tool use. Tools and memory cover what the system can access and retain. Learning covers how it improves over time. Across eight benchmarks, the best local setup comes within 3.2 points of Claude Opus 4.6. It runs at roughly 800× lower per-query cost and 4× lower latency on local hardware. A frontier model can act as a one-time teacher during configuration search. That costs roughly $15. It does not need to handle every query. Tool changes produced the largest share of improvements on research and tool-calling tasks. Full results and setup are here: https://lnkd.in/e2wdZikp Repo: https://lnkd.in/duSZ5SFs

The organizations developing AI have the deepest understanding of their technology, and we are pleased to support companies like NVIDIA that are taking a leadership position to facilitate responsible use and development.  The NVIDIA Open Agent Safety Platform is a concrete step toward setting safer boundaries for AI agents, enabling teams across industries to run frontier training and inference safely.

Lambda is coming to Mayes County, Oklahoma. The new data center will deploy closed-loop cooling to reduce water use, and Lambda will pay 100% of its energy costs. It's expected to generate $500M in tax revenue over the next decade and create up to 1,000 construction jobs: https://lnkd.in/g4GSrVF2 We’re excited to join the Mayes County community and look forward to listening, learning, and working alongside local leaders and residents. We’re committed to being a good neighbor and long-term partner as this project moves forward. Read more in Tulsa World Media Company: https://lnkd.in/gDYvs49f

AI compute is not becoming a commodity. The industry is building vertically integrated infrastructure, reaching from software and networking to the physical sites and power that make large-scale compute possible. The stakes are physical. Lambda Co-Founder & CTO Stephen Balaban makes that case on the Main Stage at Yotta 2026. He will also discuss our target of 3 GW of AI compute capacity by 2030. Session details: “Building a Cloud for the Age of Superintelligence” Tuesday, September 29, 2026 2:00–2:25 p.m., Main Stage Caesars Forum, Las Vegas Lambda Co-Founder Stephen Balaban will discuss the infrastructure buildout driven by frontier-model revenue, the shift from traditional data centers to gigawatt-scale AI campuses, and Lambda’s roadmap toward three gigawatts of AI compute capacity by 2030. Session page at https://lnkd.in/es2MBaWx The argument in one line. AI infrastructure is moving from traditional data centers to gigawatt-scale campuses, and the companies building it are changing with it. Day 2 raises a different question. If you are building an AI-native enterprise, when do you rent compute, when do you reserve dedicated capacity, and when does owning it become the competitive advantage? Lambda President of Cloud Services David Ward joins a panel on exactly that. It takes place Wednesday, September 30, 11:40 a.m.–12:25 p.m., in the Future of Compute track. Panel page at https://lnkd.in/eiXFyWAM

Stephen Balaban has been through 5 pivots in 14 years. He started with facial recognition and AI filters before building the workstation and GPU compute business. The interview is a look at the long, non-linear path behind Lambda, and the decisions that shaped it along the way. Listen to the full conversation with Lambda's CTO below on the Founders In Arms podcast with Immad Akhund and Rajat Suri: https://lnkd.in/ecaxTSBw

Ask a 3D vision-language model what's near the table and in front of the curtain, and it might guess "sewing machine." The right answer is a tray rack. CVP (UC San Diego + Lambda, WACV 2026) fixes this with a target-affinity token for task-relevant objects and an allocentric grid for global context. Against Video-3D-LLM: • SQA3D EM: 58.6 → 62.3 • Scan2Cap CIDEr: 83.8 → 90.5 • Better on all 5 benchmarks tested Full results across ScanQA, SQA3D, ScanRefer, Multi3DRefer, and Scan2Cap, plus how the central/peripheral split works: https://lnkd.in/ebntU5HH

New STAC Research report: benchmarking the infrastructure for agentic quantitative research. Given a goal, agentic systems will propose an approach, write the code, train a model, evaluate it, and decide what to try next. Using a workload built from a recorded research session contributed by a large US market maker, our latest report leverages the SwarmOne simulator and toolkit to measure a Lambda 1-Click Cluster of NVIDIA B200 GPUs at four levels of load and shows the performance of the platform as more agents run on the cluster. STAC is forming a working group under the STAC Benchmark Council to develop a rigorous benchmark for this class of workload. If you're interested in helping shape it, contact us at https://lnkd.in/dd-MEsfa Read the report: https://lnkd.in/dBzxzYUZ Subscribe to STAC Insights to read the accompanying STAC Configuration Disclosure with exact product versions and tuning detail: https://lnkd.in/d6S3vAXB

AI for molecular dynamics has a data problem. The trajectories you need to train on are expensive. EGInterpolator (ICLR 2026, with Stanford) learns molecular structure first from abundant conformer data, then uses scarce MD data to learn motion. On the DRUGS benchmark, it reduced the gap to reference simulations by 73% for bond angles, 78% for bond lengths, and 24% for torsional motion. The structure-first ablation also matters. Removing pretraining increased mean JSD from 0.173 to 0.332 for bond angles and from 0.142 to 0.386 for bond lengths. Training and evaluation ran on Lambda GPU infrastructure Paper: https://lnkd.in/eGmqiXTA Blog: https://lnkd.in/eM6Q52nC