AWS wants 99% of the AI market
Add Axios as your preferred source to
see more of our stories on Google.
Amazon Web Services (AWS) CEO Adam Selipsky. Photo: Noah Berger/Getty Images for AWS
Amazon Web Services CEO Adam Selipsky has more than 10,000 organizations using his Bedrock service to build their AI applications — but that's only around 1% of the potential market, he told Axios.
Why it matters: AWS aims to expand from cloud computing market leader to running the world's biggest AI playground.
- Amazon's strategy is to bring millions of companies into that ecosystem while it works to improve its own AI models and chips, and increase its market share.
Catch up quick: Selipsky's AI tool chest runs from data centers and AWS purpose-built chips to helping Bedrock customers finetune AI models built outside AWS.
Driving the news: AWS recently expanded its collaboration with Anthropic March 4, when the model developer launched its "Claude 3" family, which includes Opus Sonnet and Haiku.
- The Claude 3 family will be available through Bedrock, while Anthropic will use AWS Trainium and Inferentia chips to build, train, and deploy its future foundation models.
- AWS announced around 40 generative AI "competency partners" — AWS-approved service and software partners including Nvidia, Hugging Face, Anthropic, Deloitte, and MongoDB, which AWS will help pair with their customers to build AI applications quicker.
- Mistral, another hot AI startup was added to Bedrock on March 1.
Intrigue: Selipsky says AWS is "best place in the world" to both house and rent GPUs — the most sought after AI chips — but the company is also working furiously to provide alternatives to GPUs.
- The AWS play is to offer better "price performance" and lower energy consumption with its chips, which it thinks are good enough to use for most tasks.
What's happening: Selipsky says AWS is collaborating with Anthropic to "help make their models better, " and in return Anthropic helps improve AWS's chips.
- Data from Anthropic shows Opus outperforming OpenAI's GPT-4—in reasoning, math, and coding, with Selipsky calling them the "leading models in the market right now" for those specializations.
- The 10,000 Bedrock users include golf's PGA Tour for fan content and Genomics England, for spotting previously unnoticed associations between genetic variants and medical conditions in peer-reviewed papers.
- Selipsky thinks his customers are getting smarter about model choices — seeing "faster and less costly" smaller models as useful for low-risk tasks like meeting summaries where some loss of accuracy is acceptable.
- Those prizing accuracy need to think about bigger models, and techniques such as retrieval-augmented generation (RAG) which enable models to cite their sources and provide more accurate answers, he said.
Reality check: AWS is still playing catch-up in several key markets including chasing OpenAI and Microsoft's models, and Nvidia's GPU chips.
What they're saying: Selipsky called Bedrock "essentially a safety harness" around AI models, and urged AI leaders to be "humble about our state of knowledge" and about "how much more we need to collaborate" to manage AI risks.
- Sheila Gulati, co-founder and managing director at Tola Capital, told Axios "It's great to see mega caps having to get aligned with newer companies, and to have multiple models succeed as we go through this massive, massive societal shift."
What's next: Selipsky promised the next generation of Trainium chips will be four times faster than the current version, and is looking to continue to expand the number of models available in Bedrock.
- He predicts 2024 will bring more AI applications running on computing edge infrastructure and more demand from companies for customized AI models.
