The short answer is that DataRobot has invested significant R&D dollars into allowing neural networks to run performantly on commodity CPUs. For example, DataRobot has created a unique, patented method of transfer learning for computer vision that does not require GPUs. We've also found that for many real world tabular datasets, neural networks on CPUs perform just fine and allow us to avoid passing along the additional costs of GPUs to our customers.
The longer answer is that clearly some tasks, e.g. fine-tuning pre-trained neural net models for computer vision or NLP, do require GPUs. That's why we're actively researching ways to incorporate GPUs into our platform that allow users to utilize them when required but don't dramatically increase the cost.
Is there a specific use case you were curious about that requires GPUs?