14 real-world examples of companies getting results from AI with machine learning: the numbers, the companies, and the tools behind each win. Updated daily.
Most-used tools here: PRAGMA, NVIDIA Dynamo-Triton, NVIDIA Cosmos Data Factory, Voxel51, Phaidra, NVIDIA DSX
Revolut is rebuilding parts of its technology stack around PRAGMA, a single foundation model running on NVIDIA Blackwell Ultra GPUs with cuDF and Dynamo-Triton. The system already serves an initial 80 million customers and is engineered to handle three times…
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…
AXA is using NVIDIA's open Earth-2 weather models to simulate regional cyclone risk across more than 1,000 seasons. Vaisala Xweather is adding lightning data to its StormScope nowcasting tool for short-term warnings, and TotalEnergies is testing subhourly…
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…
The nonprofit U.S. Hunger uses Amazon Quick, connected nightly to its SQL Server data via SPICE, so leaders can explore program and financial data without waiting on new data pulls. CEO Rick Whitted now follows his own questions through Full Cart…
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…
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…
AWS engineers fine-tuned the open-weight GPT-OSS-20B model on Amazon SageMaker HyperPod using supervised fine-tuning, with GRPO added in some runs. The model's reasoning matched the requested language in French, German, Spanish and Italian far more often…
Spotify says it used its recommendation system to filter out more than 50 million tracks it classed as AI-generated slop or spam over the past year. It also shifted recommendations toward human-made music and reports a marked drop in listener complaints…
Three startups from the Barcelona Supercomputing Center's AI Factory Accelerator Programme have moved AI into production or deployment. Qatium uses its water-network AI in early production, OneCareAI built a tool using NVIDIA NIM to flag hidden atrial…
Deepgram, which builds speech-to-text, text-to-speech, and voice agent models, moved its workloads to AWS Trainium to run its voice AI at scale. The company now handles quadrillions of audio tokens a year for more than 200,000 developers and 1,400…
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…