12 real-world examples of companies getting results from AI delivering cost savings: the numbers, the companies, and the tools behind each win. Updated daily.
Most-used tools here: NVIDIA Cosmos Data Factory, NVIDIA Nemotron, Voxel51, Phaidra, NVIDIA DSX, Sidekick
Porsche is leading a shared autonomous vehicle data flywheel across Volkswagen Group brands. It uses NVIDIA Cosmos Data Factory blueprints, run through Voxel51 on Microsoft Azure, to generate and process training data. The effort is expected to sharply…
Phaidra applied self-learning AI agents built on NVIDIA DSX Flex, DSX Max-Q and DSX Exchange to tune cooling and power allocation in data centers. The result is roughly 10% more usable compute under the same power budget, plus better grid flexibility…
Shopify post-trained an open NVIDIA Nemotron model to run the GraphQL subagent behind its Sidekick merchant assistant. The model matched a leading frontier model's accuracy on an internal benchmark while cutting serving costs 90% and responding seven times…
Researchers at Perplexity and NVIDIA introduced JOLT, a post-training method that trains one language model as both a privileged-information teacher and a student. On the MATH benchmark it matched GRPO's accuracy while processing 6.5 times fewer training…
Google CEO Sundar Pichai says he now relies on AI agents that condense the previous 24 hours of internal activity, follow threads across chat groups, and monitor compute usage by product area. He says the agents save him roughly an hour a day, while humans…
Google Cloud's new A5X instances, built on NVIDIA Vera Rubin NVL72, are generally available next month after early access. In WPP's initial tests, creative agent workflows that normally needed a multi-node cluster ran on a single A5X instance.
European AI cloud provider Nscale tested NVIDIA DSX MaxLPS on GB300 NVL72 systems to match power allocation to workloads. The tests showed 37% more GPUs and nearly 50% higher aggregate inference throughput within the same provisioned power. Nscale plans Vera…
Siemens trained its circuit-board inspection system at an Erlangen electronics factory using synthetic defect images created with NVIDIA Cosmos Data Factory tools. The approach reduced inspection calibration time by 40%. Siemens also estimates 20% cost…
Google Cloud and partners released llm-d 1.0, a stable Kubernetes-native layer that coordinates vLLM or SGLang model servers into one pool. A vendor-reported Ironwood TPU coding benchmark showed median time-to-first-token cut by more than 85% and 2.4x…
i14, working with AWS's GenAI Innovation Center, rebuilt its Laplace 3B linear attention model for AWS Trainium2 chips using native PyTorch via TorchNeuron Native. The design uses a fixed-size recurrent state, so serving memory stays flat as context length…
Cloudflare tested its Clairic extraction model on Workers AI in a damaged-product claims pipeline, chaining it with its Clef decision model and GLM-5.3 for drafting replies. The cascade ran about 3.2x faster than GLM-5.3 alone and lowered cost per million…
Microsoft says its in-house Maia 200 accelerator is now running in production across the US, Europe and Australia, serving Copilot and built-in AI features in Teams. The company reports up to 1.7x more tokens per watt than leading GPU infrastructure and…