Amazon is leaning into generative AI with several new projects and partnerships. The results are more impressive than the executive team expected.
"We have been surprised at the pace of growth in generative AI," said Brian Olsavsky, CFO of Amazon during the company's third quarter earnings call, as transcribed by The Motley Fool. "Our generative AI business is growing very, very quickly…and by any measure, it's a pretty significant business for us already. And yet, I would also say that companies are still in the relatively early stages."
Amazon categorizes generative AI into three macro layers, and is investing heavily in each. The lowest layer is compute to train large language models that can make predictions; the second layer is large language models as a service; the third layer is the applications running the large language models, which have gotten "a lot more powerful recently" after launching a new customization capability, Andy Jassy, CEO of Amazon, said during the earnings call.
"Every one of our businesses is building generative AI applications to change what's possible for customers, and we have a lot more to come," Mr. Jassy said.
Anthropic, a large language model maker, chose Amazon Web Services as the primary cloud provider recently and aims to use Trainium and Inferentia to build, train and deploy large language models in the future, according to Mr. Jassy. The company also introduced Amazon Bedrock so customers can access large language models from Anthropic, Cohere, Stability AI and other third-party providers.
Right now, more than 90% of global IT spend is on-premise, Mr. Olsavsky said, but in the next decade he believes that will drop to 10% on-premise, 90% in the cloud. He has seen companies, both established and startups, using Trainium and Inferentia to accelerate large language model training as the companies figure out which models will be most beneficial to produce at the large scale.
"What happens is you try a model, you test the model, you like the results of the model and then you plug it into your application, and what a lot of companies figure out quickly is that using the really large models and the large sizes ends up often being more expensive than what they anticipated and what they want to spend on that application, and sometimes too much latency in getting the answers as it shovels through the really large models," said Mr. Olsavsky. "Customers are experimenting with lots of different types of models and then different model sizes to get the cost latency characteristics that they need for different use cases."