Many companies that begin their AI projects in the cloud often reach a point when cost and time variables become issues. That\u2019s typically due to the exponential growth in dataset size and complexity of AI models.\n\n\u201cIn an early phase, you might submit a job to the cloud where a training run would execute and the AI model would converge quickly,\u201d says Tony Paikeday, senior director of AI systems at NVIDIA. \u201cBut as models and datasets grow, there\u2019s a stifling effect associated with the escalating compute cost and time. Developers find that a training job now takes many hours or even days, and in the case of some language models, it could take many weeks. What used to be fast, iterative model prototyping, grinds to a halt and creative exploration starts to get stifled.\u201d\n\nThis inflection point related to the increasing amount of time needed for AI model training \u2014 as well as increasing costs around data gravity and compute cycles \u2014 spurs many companies to adopt a hybridized approach and move their AI projects from the cloud back to an on-premises infrastructure or one that\u2019s colocated with their data lake.\n\nBut there\u2019s an additional trap that many companies might encounter. Paikeday says it occurs if they choose to build such infrastructure themselves or repurpose existing IT infrastructure instead of going to a purpose-built architecture designed specifically for AI.\n\n\u201cThe IT team might say, \u2018We have lots of servers, let's just configure them with GPUs and throw these jobs at them\u2019,\u201d he says. \u201cBut then they realize it\u2019s not the same as a system that is designed specifically to train AI models at scale, across a cluster that\u2019s optimized to deliver results in minutes instead of weeks.\u201d\n\nWith AI development, companies need fast ROI, by ensuring data scientists are working on the right things. \u201cYou\u2019re paying a lot of money for data-science talent,\u201d Paikeday says. \u201cThe more time they spend not doing data science \u2014 like waiting on a training run, troubleshooting software, or talking to network, storage or server vendors to solve an issue \u2014 that\u2019s lost money and a lot of sweat equity that has nothing to do with creating models that deliver business value.\u201d\n\nThat\u2019s a significant benefit of a purpose-built appliance for AI models that can be installed on premises or in a colocation facility. For example, NVIDIA\u2019s DGX A100 is meant to be unpacked, plugged in and powered-up enabling data scientists to be productive within hours, instead of weeks. The DGX system offers companies five key benefits to scale AI development:\n\nWhen it\u2019s time to move an AI project from exploration to a production application, the right choice can speed and scale the ROI of your AI investment.\n\nDiscover how NVIDIA DGX A100, powered by NVIDIA A100 Tensor Core GPUs and AMD EPYC CPUs, meets the unique demands of AI.