Lambda

GPU utilization increased from ~20% to 43% on a reservation of 96 NVIDIA H100 GPUs, while cutting queue starvation by 74%, with blocked jobs falling from roughly 10 per day to around 4. That’s the concrete result SPREEAI saw after fixing their orchestration. When a unified diffusion model requires 80–100 GB of memory, you can’t simply throw workloads at a cluster and expect to use those GPUs efficiently. SPREEAI was dealing with workload fragmentation, ad-hoc submissions, and storage I/O blocking that left expensive GPUs idle. Working with Lambda’s ML engineering team, they implemented MLflow-based experiment orchestration with structured queuing and workload matching. They also connected Lambda’s Prometheus APIs to Grafana for real-time visibility into utilization gaps. The video testimonial covers how they diagnosed the bottlenecks and what the remediation looked like.

Evaluating image editing models with a single score hides the nuance behind a lower score. EdiVal-Agent (ICLR 2026) splits evaluation across instruction following, content consistency, and visual quality. Its agentic judge reached 81.3% agreement with human judgments, compared with 75.2% for a VLM-only evaluator and 68.9% for CLIP_dir. The failure that stands out is how quickly multi-turn editing exposes weak preservation. Each new instruction has to land without undoing earlier edits or changing unrelated content. Errors compound even when single-turn results look strong. EdiVal-IF measures instruction following. EdiVal-CC covers content consistency, while EdiVal-VQ checks visual quality. https://lnkd.in/eVidPybM

Build or buy AI? Wrong question. Lambda's Robert Brooks IV is on the main stage at AI Infra Summit 2026 with JLL CTO Yao Morin, U.S. Bank EVP & Chief AI Officer Prashant Mehrotra, and Carrier Chief Data & AI Officer Arun Nandi, and the room's landing on the same answer: it's build AND buy. New term coined: valuemaxxing. Using value per task as a lens to govern resource allocation, while architecting around agility (easy model swaps) and availability (SLAs and model lifecycle).

MLPerf Inference v6.1 is out, and we have two firsts: the first agentic inference workload on datacenter hardware, and the first MLPerf deployment of a model over a trillion parameters. On 4x NVIDIA Blackwell Ultra GPUs, we posted the leading Offline throughput on GPT-OSS 120B among all Blackwell Ultra GPU submissions, plus an 8.85% throughput gain over v6.0 on identical hardware. That's six months of pure software optimization. On NVIDIA HGX B200, we swapped in Kimi K2.6 for the open-division agentic benchmark, running a 1T+ parameter model where the reference workload expects 27B. Same harness, no memory ceiling. VLM inference made its debut in our lineup too. Qwen3-VL on NVIDIA HGX B200 came in 28.5% faster (Offline) than the fastest v6.0 submission on comparable hardware. The agentic number that matters most: 1,007 of 1,007 replay turns completed, zero failures, 86.83% BFCL v4 accuracy, at a fraction of the latency of the 27B reference deployments. Full results and methodology in the blog. Link below. https://lnkd.in/e8udxxeM