AIArtificial Intelligence (AI) and ML implementations are limited by lack of data, high model complexity and available compute resources. The last two factors can be alleviated by using hardware acceleratorsHardware accelerator. These can be based on GPUsGPU, FPGAsFPGA or ASICASIC based designs. In future, the need for AI training and inference can be combined in a single piece of hardware. This will reduce both the capital expenditure (CAPEX) and operational expenditure (OPEX) in a data-center. To summarize, for training CPU based Cloud operations are good, for classificationClassification and predictionPrediction on real-time data GPUsGPU are good and for persistent data (like image archives) TPUsTPU are good.