It seems to be a &#34;laggard&#34; in the AI competition, but Amazon AWS has a clear strategic vision.
This Wednesday, Amazon held its re:Invent annual conference and launched a series of major products and plans, including the world's largest AI supercomputer, next-generation AI chips and models, and training for Apple AI...
Among them, the most noteworthy is Amazon's planning for future development, where AWS bets on &#34;generative AI productization,&#34; making products and services the core of development, building a more user-friendly and accessible platform, and lowering the threshold for user groups to use AI.
AWS believes that generative AI will become more prevalent in the future, not just an additional feature, but part of the computing infrastructure. The necessary path for productization is that the importance of large models will gradually decrease, and the differences between models for different purposes will not be significant for users.
<h2>Large models will never dominate alone; their importance will gradually decrease</h2>
In traditional AI development, large models can be said to be the core, requiring careful design and training to solve specific problems. However, AWS believes that in the productization process of generative AI, large models will no longer be the sole focus.
For the past two years, Amazon has relied on Anthropic's AI and Nvidia's hardware to drive its cloud sales growth, but at the re:Invent conference, Amazon CEO Andy Jassy stated:
<blockquote>
I believe almost everyone will eventually use Anthropic's Claude model, but they are also using the Llama model, the Mistral model, and some of our own models...
</blockquote>
This aligns with the views of AWS CEO Matt Garman, who previously stated that the AI competition has no finish line, and that in the future, large AI models will never dominate alone.
Some analysts suggest that AWS's bet is that a necessary component of generative AI productization is the gradual decrease in the importance of large models, which will lower the threshold for users to enter the field of generative AI. In other words, AI will be important enough that it will ultimately not be particularly special.
Garman also expressed in the No Priors podcast that generative AI will become a &#34;necessity&#34; in the future:
<blockquote>
From the development of technology, generative AI will not just be an additional feature, but will become part of the computing infrastructure, just like storage, computing, and databases, which are fundamental elements to consider when building applications.
In applications, if a large amount of data processing and decision-making (i.e., reasoning) is required, then generative AI will be a key component. This indicates that the reasoning capabilities of AI will be widely integrated into various applications to meet these needs.
</blockquote>
<h2>Products and services are the core, making progress in cost reduction</h2>
As the importance of large models decreases, AWS focuses on building AI products and services. AWS is working to integrate generative AI into its platform, allowing developers to leverage the AI services provided by AWS when building applications, without having to start from scratch with complex AI systems.
When discussing the company's development focus, Garman stated:
<blockquote>
Amazon now has a Bedrock (pre-trained model family) platform, training chips, and inference chips, along with a range of other features and models, both proprietary and open-source, as well as open weights.
Enterprises prefer to build on this platform, and we see that companies are really inclined to and want to build in this area because it gives them a lot of autonomy, which is exactly what they want when building applications.
</blockquote>
Garman believes:
<blockquote>
AWS's job is to make it increasingly easier for users to build AI applications in a tightly coupled manner, making it easier to use different components, innovate quickly, and build proprietary data from the AWS data lake. Frankly, if you don't have interesting data, most of these generative AI systems won't be very useful.
</blockquote>
When using generative AI, developers need to make trade-offs in terms of cost and latency capabilities. Large models may offer more complex functionalities but could be more expensive and have higher latency; whereas smaller models may be more economical and respond faster, but their functionalities may be limited. By providing multiple model options, AWS allows users to choose the most suitable model based on specific needs, rather than relying on a single model.
Amazon has made some progress in reducing model and chip costs, Jassy stated that Amazon's cutting-edge new models are basically as powerful as the latest models from Anthropic (and OpenAI), and notably, they are priced more than one-third lower than these models.
Garman mentioned that Amazon's new AI chip Trainium 2 offers far greater value to AI developers than Nvidia's flagship H100 chip, and he also noted that Anthropic will use a large number of Trainium chips to develop future models

AMZN

AWS bets on "generative AI productization," making products and services the core of development, building a more user-friendly and accessible platform, lowering the barriers for users to utilize AI. In the future, the importance of large models will gradually decrease, and the differences between models for different purposes will be minimal for users

- Amazon AWS is positioning itself strategically in the AI competition, focusing on the productization of generative AI to lower user barriers.  
- The company aims to integrate generative AI into its platform, allowing developers to utilize AWS services without building complex systems from scratch, while emphasizing the importance of cost-effective and efficient models.  
- AWS has made progress in reducing costs for models and chips, with new AI models performing comparably to leading competitors at significantly lower prices.

"Cloud's largest factory" clearly articulated the AI strategy: whether it's large models or chips, they will be options rather than the core