With the dust finally settling on 2024’s edition of AWS re:Invent, we highlight what the announcements mean for Amazon’s AI strategy.
With the evolution of AI in full swing, organisations are realising and grappling with its opportunities and challenges.
Against a backdrop of increasing regulatory scrutiny, with the EU AI Act coming into force and similar legislation being considered in the US, UK and Asia-Pacific regions, major technology companies are competing to establish their positions in this endlessly advancing market.
Where concerns about AI safety and ethics intersect with unprecedented technological advancement, several key trends have emerged.
First, there’s a growing emphasis on responsible AI development, particularly following the 2023 AI Safety Summit and subsequent international agreements.
Second, there’s increasing pressure for AI solutions that can operate within varying regulatory frameworks across different jurisdictions.
Third, businesses worldwide are seeking more cost-effective and efficient ways to implement AI technologies, especially amid economic uncertainties and rising computing costs.
Aiming to address these challenges with not just one but a multitude of new AI models, Amazon has announced its Nova family of AI models in the company’s AI strategy.
What sets the Nova announcement apart is its breadth and ambition with not just multiple models, but with each model being optimised for different use cases and computational requirements.
While earlier generations of foundation models demonstrated impressive capabilities, their practical implementation often faced challenges related to cost, latency and integration complexity.
Amazon’s Nova lineup appears specifically designed to address these pain points.
Amazon’s AI investment strategy
In the lead up to Amazon developing its own AI models with Nova, it has made significant investments in other AI companies.
Andy Jassy, Amazon’s CEO says that developers have a list of desires when it comes to improving AI: “They want better latency. They want lower cost. They want the ability to do fine-tuning.”
As a result, Amazon has invested billions in Anthropic, as well as launching Amazon Q, the AI-powered assistant for work that can be customised for specific company needs.
Amazon Q combines Amazon’s AI capabilities to help businesses boost productivity and transform their operations, designed to assist with tasks such as coding, testing, upgrading, troubleshooting and security scanning for developers, while business users can use it for tailored conversations, problem-solving, content generation and task streamlining.
Now, with external investments alongside Nova, Amazon’s AI investment strategy demonstrates an approach to AI development and implementation that combines in-house innovation with strategic partnerships to meet the diverse needs of businesses in the global AI market.
Nova lineup and capabilities
Announced on stage at re:Invent, the Amazon Nova lineup includes several models designed for different purposes.
Nova Micro is a text-only model focused on speed and low cost, whereas Nova Lite, Pro and Premier are multimodal models capable of processing text, images and videos.
Nova Canvas is an image generation model, while Nova Reel is designed for video generation.
According to Amazon, these models have been benchmarked against competitors and perform competitively or better in various tasks.
The company claims that its models offer superior price performance, with costs at least 75% lower than comparable models available through its Amazon Bedrock service.
Nova Micro, for instance, was found to be equal or better than both Meta’s LLaMa 3.1 8B and Google’s Gemini 1.5 Flash-8B on all applicable benchmarks.
With a speed of 210 output tokens per second, it is positioned as ideal for applications requiring fast responses.
Meanwhile, Nova Lite has shown competitive performance against other models in its intelligence class, performing equal or better on 17 of 19 benchmarks compared to OpenAI’s GPT-4o mini, and on 17 of 21 benchmarks compared to Google’s Gemini 1.5 Flash-8B.
It excels in understanding videos, charts and documents, as measured by benchmarks such as VATEX, ChartQA and DocVQA.
Nova Pro also demonstrated strong performance, matching or surpassing OpenAI’s GPT-4o on 17 of 20 benchmarks and Google’s Gemini 1.5 Pro on 16 of 21 benchmarks.
It particularly excels at instruction-following and multimodal agentic workflows.
Integration and customisation
Amazon is positioning its Nova models as highly integrable with existing systems.
The models are designed to work seamlessly with Amazon Bedrock, allowing customers to experiment with different models to find the best fit for their applications.
The company is also offering customisation options, including fine-tuning and distillation.
Fine-tuning allows customers to improve model accuracy using their own data, while distillation enables the creation of smaller, more efficient models based on larger, more capable ones.
The Nova models additionally support over 200 languages and offer extensive context length capabilities.
For example, Nova Micro supports a context length of 128K input tokens, while Nova Lite and Pro support 300K tokens or 30 minutes of video processing.
Amazon plans to support context lengths of over 2 million input tokens in early 2025.
Future developments and industry adoption
Amazon has outlined plans for future AI model releases.
“Our new Amazon Nova models are intended to help with these challenges for internal and external builders and provide compelling intelligence and content generation while also delivering meaningful progress on latency, cost-effectiveness, customisation, Retrieval Augmented Generation (RAG) and agentic capabilities”, Rohit Prasad, SVP of Amazon Artificial General Intelligence explains.
These include a speech-to-speech model scheduled for release in Q1 2025 and a multimodal-to-multimodal model capable of processing and generating content across various formats, expected in mid-2025.
The company is also developing a novel model that can take text, images, audio and video as input, and generate outputs in any of these modalities.
As Amazon dubs, this “any-to-any” modality model, is designed to simplify the development of applications that can perform a wide variety of tasks, such as translating content from one modality to another, editing content and powering AI agents that can understand and generate all modalities.
Rohit summarises: “Inside Amazon, we have about 1,000 Gen AI applications in motion and we’ve had a bird’s-eye view of what application builders are still grappling with.”
Several companies have already begun integrating Amazon Nova models into their operations.
SAP is incorporating the models into its AI Core Gen AI hub, while Deloitte plans to use the models to deliver AI services to its global clients.
In the creative industry, companies like Dentsu Digital, Musixmatch and 123RF are utilising Amazon Nova Reel and Canvas for video and image generation tasks. Caylent, a cloud services company, is utilising Amazon Nova models to provide video understanding capabilities to customers in media, sports and retail.
Palantir Technologies, meanwhile, is integrating Amazon Nova Pro with its AI Platform to enhance decision-making across various industries, including insurance and supply chain management.
“I have never been as excited today as I am about the future”, Matt Garman said on stage.
“We are at a seminal point. The amount of innovation that we’re seeing out there in the world is really exciting, and that’s not just innovation that’s coming from AWS.
“I’m really excited about the innovation that we’re seeing from customers, from partners, from enterprises, from incredible startups – there has never been a better time to be innovating.”
