I think it’s part of what — if you look at our very significant growth rates right now in everyday essentials and consumables, a lot of it is when you — if you’re going to order something which you need in the same day or next day, you’re not going to consider it if it’s coming in 3 or 4 days. But when you’re consistently getting it in same day or the next day, just changes what you’re willing to do. I think the second question was on gen AI and the timing of monetization. What I would tell you is that we have been surprised at the pace of growth in generative AI. Our generative AI business is growing very, very quickly, as I mentioned earlier. And almost 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.
I mean now you have to get perspective. My perspective is that the cloud is still in the early stages. If you think about 90% plus of the global IT spend being on-premises, where I think that equation is going to flip in 10 years. I think cloud is early. So if you — with that lens on, I still think we’re very early in generative AI. And what’s interesting, too, around generative AI because we have so many companies who are doing all sorts of prototypes and it’s really accelerating very rapidly on the training side with Trainium and Inferentia and then on the application building and running side with Bedrock, is that companies are still trying to sort out for themselves what they’re going to run at large-scale production in all of these areas.
Because 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 — the 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. And so customers are experimenting with lots of different types of models and then different model sizes to get the cost and latency characteristics that they need for different use cases. And it’s one of the things that I think is so useful about Bedrock is that customers are trying so many variants right now but to have a service that not only lets you leverage lots of third party as well as Amazon large language miles but also lots of different sizes and then makes the transition of moving those workloads easy between them is very advantageous.
Operator: Our next question is from Brian Nowak with Morgan Stanley.
Brian Nowak: Thanks for taking my question, I have two, one on AWS, one on the retail business. On AWS AI Andy, I recognize you have a pretty multipronged AI approach. But just could you talk us through sort of one or two of the early generative AI products where you’re seeing the most early demand and interest? And as you talk to customers, are there still hurdles or pain points that you’re not quite serving in your product suite you look to solve and innovate on over the next couple of years in AI? And then the second one, you made a lot of steps on the regionalization of warehouses and making it more efficient. Where are you on robotics in the warehouses? And how should we think about the potential impact of that to drive profitability even higher?
Andrew Jassy: On the AI side, I think that if you’re looking for some of the products that we’re offering that are — that have a lot of early resonance and traction, I would — I’d start with Bedrock. These customers are — they’re very excited about Bedrock. And it’s making it so much easier to get applications, generative AI applications built. And again, it’s machine learning and AI has been something that people have been excited about for 25 years. In my opinion, about a half dozen years ago, it took a pretty significant leap forward where it was much easier given the economics and scalability of compute and storage and then some of the tools that we built like SageMaker, it was much easier for everyday developers to start to interact with AI.
But it just — it took another meaningful step forward with generative AI but still it’s complicated to actually figure out which models you want to work, you want to use and how you actually want to employ them and trying to make sure you have the right results, trying to make sure you get safe results, trying to make sure you end up with a cost structure and a customer experience that you want. And so it’s hard. And customers will like — there’s a certain number of customers who have very deep AI expert practitioners but most companies don’t. And so Bedrock just takes so much of the difficulty out of those decisions and those variables that people are very excited about Bedrock. They’re using it in a very broad way. They’re extremely excited about not just the set of models in there but if you look at a leader like Anthropic and the ability for our customers in Bedrock to have exclusive early access to models and customization tools and fine-tuning which gives them more control, there’s just — there’s a lot of buzz and a lot of usage and a lot of traction around Bedrock.
I would also say our chips, Trainium and Inferentia, as most people know, there’s a real shortage right now in the industry in chips. It’s really hard to get the amount of GPUs that everybody wants. And so it’s just another reason why Trainium and Inferentia are so attractive to people. They have better price performance characteristics than the other options out there but also the fact that you can get access to them. And we’ve done, I think, a pretty good job providing supply there and ordering meaningfully in advance as well. And so you’re seeing very large LLM providers make big bets on those chips. I think Anthropic deciding to train their future LLM model on Trainium and using Inferentia as well is really a statement. And then you look at the really hot start-up Perplexity.ai, who also just made a decision to do all their training and inference on top of Trainium and Inferentia.
So those are 2 examples. I’d say CodeWhisperer, too, is just, again, it’s just a game changer if you can allow your engineers not to have to do the more repetitive work of cutting and pasting and building certain functions that really, if somebody knew your code base better, could do. And so it’s a real — it’s a productivity game changer for developers. And then actually launching that customization vehicle so that they actually understand your own proprietary code base, that is something that customers are quite excited about. So those are ultimately early traction AI. There’s so much more to provide, Brian. I mean there’s — I think people — even though Bedrock is so much easier to use than people trying to build models themselves and build the applications, I think people are still looking to find ways to make it easier to look at a big corpus of data and run an agent on top of it or maybe they don’t have to do all that work themselves.
I think people are looking for automated ways to understand developer environments and be able to ask any question on the developer environment. So there’s a lot more. It’s going to be a long time before we run out of services. And yes, I think it’s a good thing to look toward the next few months in re:Invent to see the additional things that team launches. I think on the robotics piece, it’s a very significant investment for us. It has been for several years. It’s made a huge difference for us. It’s made a big difference for us both on the productivity side, on the cost side as well as importantly on the safety side where we can have our teammates working on things that are even safer than what they get to work on today. We have a very substantial investment of additional robotics initiatives, I would say many of which are coming to fruition in 2024 and 2025 that we think will make a further additional impact on the cost and productivity and safety and our fulfillment service.
Operator: Our next question comes from the line of Eric Sheridan with Goldman Sachs.
Eric Sheridan: Maybe one follow-up on AWS and one on the ads business. Andy, would love your perspective. The cloud optimization theme started in the second half of ’22 when there were a lot of macro concerns. And then the AI theme really only sort of came to the forefront in the last 8 or 9 months. What’s your perspective on how turning the calendar into 2024 and there being a new IT budget cycle could possibly lead us to put the optimization theme in the background and some of the AI theme come more to the forefront when there might be more distinct budgeting around AI as a theme? That would be number one. And then in your ads business, you’re approaching $50 billion run rate and it’s compounding in the mid-20s. What are you most excited about on the initiative front to continue to build scale both on Amazon properties and possibly off Amazon as a broader digital advertising player?
