QV Alert – Amazon Spends Big on AI

I don’t know if anyone has seen a picture of Jeff Bezos recently but the guy is rippling with muscles, like the company he founded. Amazon grows stronger and stronger. Some of the opportunities Amazon faces are incredible.

Internationally, we like the trajectory of our established countries, and see meaningful progress in our emerging geographies (e.g. India, Brazil, Australia, Mexico, Middle East, Africa, etc.) as they continue to expand selection and features, and move toward profitability (in Q4 2023, Mexico became our latest international Stores locale to turn profitable). We have high conviction that these new geographies will continue to grow and be profitable in the long run.

In October, we hit a major milestone in our journey to commercialize Project Kuiper when we launched two end-to-end prototype satellites into space, and successfully validated all key systems and sub-systems— rare in an initial launch like this. Kuiper is our low Earth orbit satellite initiative that aims to provide broadband connectivity to the 400-500m households who don’t have it today (as well as governments and enterprises seeking better connectivity and performance in more remote areas), and is a very large revenue opportunity for Amazon. We’re on track to launch our first production satellites in 2024. We’ve still got along way to go, but are encouraged by our progress.

[As a result of my passion for Brian Cox’s programmes on the solar system I know that the Kuiper Belt is a sphere of rubble and ice enclosing the solar system from which comets emerge.]

Shareholders letter, Amazon December 2023

Amazon is deeply involved with the latest iterations of AI.

Much of the early public attention has focused on GenAI applications, with the remarkable 2022 launch of ChatGPT. But, to our “primitive” way of thinking, there are three distinct layers in the GenAI stack, each of which is gigantic, and in each of which we’re deeply investing. The bottom layer is for developers and companies wanting to build foundation models (“FMs”). The primary primitives are the compute required to train models and generate inferences (or predictions), and the software that makes it easier to build these models. Starting with compute, the key is the chip inside it. To date, virtually all the leading FMs have been trained on Nvidia chips, and we continue to offer the broadest collection of Nvidia instances of any provider. That said, supply has been scarce and cost remains an issue as customers scale their models and applications. Customers have asked us to push the envelope on price performance for AI chips, just as we have with Graviton for generalized CPU chips. As a result, we’ve built custom AI training chips (named Trainium) and inference chips (named Inferentia). In 2023, we announced second versions of our Trainium and Inferentia chips, which are both meaningfully more price-performant than their first versions and other alternatives. This past fall, leading FM-maker, Anthropic, announced it would use Trainium and Inferentia to build, train, and deploy its future FMs. We already have several customers using our AI chips, including Anthropic, Airbnb, Hugging Face, Qualtrics, Ricoh, and Snap. Customers building their own FM must tackle several challenges in getting a model into production. Getting data organized and fine-tuned, building scalable and efficient training infrastructure, and then deploying models at scale in a low latency, cost-efficient manner is hard. It’s why we’ve built Amazon SageMaker, a managed, end-to-end service that’s been a game changer for developers in preparing their data for AI, managing experiments, training models faster (e.g. Perplexity AI trains models 40pc faster in SageMaker), lowering inference latency (e.g. Workday has reduced inference latency by 80pc with SageMaker), and improving developer productivity (e.g. NatWest reduced its time-to-value for AI from 12-18 months to under seven months using SageMaker).

The middle layer is for customers seeking to leverage an existing FM, customize it with their own data, and leverage a leading cloud provider’s security and features to build a GenAI application—all as a managed service. Amazon Bedrock invented this layer and provides customers with the easiest way to build and scale GenAI applications with the broadest selection of first- and third-party FMs, as well as leading ease-of-use capabilities that allow GenAI builders to get higher quality model outputs more quickly. Bedrock is off to a very strong start with tens of thousands of active customers after just a few months. The team continues to iterate rapidly on Bedrock, recently delivering Guardrails (to safeguard what questions applications will answer), Knowledge Bases (to expand models’ knowledge base with Retrieval Augmented Generation—or RAG—and real-time queries), Agents (to complete multi-step tasks), and Fine-Tuning (to keep teaching and refining models), all of which improve customers’ application quality. We also just added new models from Anthropic (their newly-released Claude 3 is the best performing large language model in the world), Meta (with Llama 2), Mistral, Stability AI, Cohere, and our own Amazon Titan family of FMs. What customers have learned at this early stage of GenAI is that there’s meaningful iteration required to build a production GenAI application with the requisite enterprise quality at the cost and latency needed. Customers don’t want only one model. They want access to various models and model sizes for different types of applications. Customers want a service that makes this experimenting and iterating simple, and this is what Bedrock does, which is why customers are so excited about it. Customers using Bedrock already include ADP, Amdocs, Bridgewater Associates, Broadridge, Clariant, Dana-Farber Cancer Institute, Delta Air Lines, Druva, Genesys, Genomics England, GoDaddy, Intuit, KT, Lonely Planet, LexisNexis, Netsmart, Perplexity AI, Pfizer, PGA TOUR, Ricoh, Rocket Companies, and Siemens.

The top layer of this stack is the application layer. We’re building a substantial number of GenAI applications across every Amazon consumer business. These range from Rufus (our new, AI-powered shopping assistant), to an even more intelligent and capable Alexa, to advertising capabilities (making it simple with natural language prompts to generate, customize, and edit high-quality images, advertising copy, and videos), to customer and seller service productivity apps, to dozens of others. We’re also building several apps in AWS, including arguably the most compelling early GenAI use case—a coding companion. We recently launched Amazon Q, an expert on AWS that writes, debugs, tests, and implements code, while also doing transformations (like moving from an old version of Java to a new one), and querying customers’ various data repositories (e.g. Intranets, wikis, Salesforce, Amazon S3, ServiceNow, Slack, Atlassian, etc.) to answer questions, summarize data, carry on coherent conversation, and take action. Q is the most capable work assistant available today and evolving fast.
While we’re building a substantial number of GenAI applications ourselves, the vast majority will ultimately be built by other companies. However, what we’re building in AWS is not just a compelling app or foundation model. These AWS services, at all three layers of the stack, comprise a set of primitives that democratize this next seminal phase of AI, and will empower internal and external builders to transform virtually every customer experience that we know (and invent altogether new ones as well). We’re optimistic that much of this world-changing AI will be built on top of AWS.

Generative AI may be the largest technology transformation since the cloud (which itself, is still in the early stages), and perhaps since the Internet. Unlike the mass modernization of onpremises infrastructure to the cloud, where there’s work required to migrate, this GenAI revolution will be built from the start on top of the cloud. The amount of societal and business benefit from the solutions that will be possible will astound us all.

Annual Shareholders Letter, Amazon, 2023

It is helpful to know what Amazon means by ‘primitives’.

The best way we know how to do this [rapidly improve the customer experience] is by building primitive services. Think of them as discrete, foundational building blocks that builders can weave together in whatever combination they desire. Here’s how we described primitives in our 2003 AWS Vision document:
“Primitives are the raw parts or the most foundational-level building blocks for software developers. They’re indivisible (if they can be functionally split into two they must) and they do one thing really well. They’re meant to be used together rather than as solutions in and of themselves. And, we’ll build them for maximum developer flexibility. We won’t put a bunch of constraints on primitives to guard against developers hurting themselves. Rather, we’ll optimise for developer freedom and innovation.”

Of course, this concept of primitives can be applied to more than software development, but they’re especially relevant in technology. And, over the last 20 years, primitives have been at the heart of how we’ve innovated quickly.

Annual Shareholders Letter, Amazon, 2023

Share Recommendations (8 January 2025)

Amazon.com. AMZN

Morgan Stanley expects Amazon to spend almost $100bn on capital expenditure in 2025. This is a staggering sum and vividly demonstrates why in terms of sheer computing power AI is a game only the biggest boys can play.

On top of that Amazon is an innovation machine. Its consumer-facing businesses generate less than 40pc of profits. Amazon is a technology giant more than an online retailer.