Cerebras Systems builds the fastest AI accelerators in the industry. In this talk we will review how the size and scope of massive natural language processing (NLP) presents fundamental challenges to legacy compute and to traditional cloud providers. We will explore the importance of guaranteed node to node latency in large clusters, how that can’t be achieved in the cloud, and how it prevents linear and even deterministic scaling. We will examine the complexity of distributing NLP models over hundreds or thousands of GPUs and show how quickly and easily a cluster of Cerebras CS-2s is set up, and how linear scaling can be achieved over millions of compute cores with Cerebras technology. And finally, we will show how innovative customers are using clusters of Cerebras CS-2s to train large language models in order to solve both basic and applied scientific challenges, including understanding the COVID-19 replication mechanism, epigenetic language modelling for drug discovery, and in the development of clean energy. This enables researchers to test ideas that may otherwise languish for lack of resources and, ultimately, reduces the cost of curiosity.  ​