Amazon Launches Bedrock to Bring Generative AI to AWS

Amazon today unveiled Amazon Bedrock, a new AWS service that lets businesses add generative artificial intelligence models to their applications through an API. The news is not the arrival of a single model, but of a platform: AWS wants to become the intermediary that makes it easier to use, adapt and deploy several of the market’s leading models.

Bedrock is initially being offered in limited preview. It includes Amazon’s own models, grouped under the Titan brand, as well as models from AI21 Labs, Anthropic and Stability AI. It is a direct response to the race led by OpenAI, Microsoft and Google to move generative AI advances from the lab into enterprise products.

One API for multiple models

So-called foundation models are systems trained on enormous amounts of data that can handle different tasks from the same underlying base: drafting text, summarizing documents, answering questions, generating code or creating images. Their power comes with a trade-off: training them from scratch requires computing capacity, data and budgets available to very few companies.

Bedrock’s proposition is to hide much of that complexity. Instead of installing models, reserving specialized servers and managing their scaling, a company will be able to send requests through an application programming interface, or API, and pay for the service based on usage once it is generally available.

Through Bedrock, AWS will offer AI21 Labs’ Jurassic-2 models, designed for text generation and comprehension; Anthropic’s Claude models; and Stable Diffusion from Stability AI, which creates images from written prompts. The company is also adding Titan, its own family of foundation models.

This combination matters because each model has different capabilities, limitations and costs. A company that needs to search its documentation may prioritize a model that produces numerical representations—also known as embeddings—while a marketing team may need one that drafts copy or generates images. AWS wants that choice to happen within its cloud, without forcing customers to rebuild their entire technology stack.

Titan, Amazon’s own bet

Amazon is not merely hosting other companies’ software. Titan marks its entry into a market that has so far been shaped by models such as OpenAI’s GPT-3 and GPT-4, Google’s PaLM and the systems developed by Meta.

The company has announced two main uses for Titan: text generation and the creation of embeddings. Embeddings turn words, phrases or documents into numbers that make it possible to compare their meanings. They are a common building block in internal search engines, recommendation systems and applications that answer questions using a document repository.

AWS also says customers will be able to customize some models with their own data. That capability is central to businesses: a customer-service assistant does not need to know everything about the internet; it needs to answer accurately about products, contracts or internal procedures. Adapting a model to that context can make it more useful, although it does not by itself eliminate generative AI’s errors.

The cloud becomes the battleground

The launch comes as the major cloud-computing platforms try to determine who will control the infrastructure behind the new wave of AI. Microsoft has integrated OpenAI’s models into Azure and products such as Bing. Google announced new generative tools for Vertex AI, its enterprise platform, in March. Amazon, whose AWS division is the global leader in cloud services, needed a visible response.

Bedrock also reflects a shift in the business model. Until recently, many organizations looking to experiment with advanced AI relied on highly specialized providers or in-house teams capable of operating complex models. Cloud platforms want to package that work as a service: API access, integration with existing tools and usage-based billing.

That could speed adoption, but it does not resolve the most difficult decisions. Companies will have to determine what data they send to the provider, how it is stored, whether the model reproduces incorrect information and who is responsible for a problematic output. A model that writes fluently is not necessarily a reliable system for every commercial, legal or financial decision.

The value lies in integration, not just the model

Amazon has presented Bedrock as a way to bring generative AI closer to developers and businesses already working with AWS. Its potential advantage lies in connecting models to the rest of the cloud catalog: storage, databases, access controls and development tools.

The question is whether companies will favor that integration or go directly to the model makers. Bedrock does not decide which system will be best for each task, but it turns that comparison into a platform decision. For AWS, getting generative applications to originate and run on its infrastructure could be as important as having a competitive model of its own.