[hnlmorg]

I’d wager AWS makes more money from GenAI than any other domain. So it makes financial sense for them to sell that part of the business hard at the moment.

This opinion is based on admittedly anecdotal experience, but I’ve worked in a large range of domains on AWS over the years and by far the biggest AWS bills were for startups specialising in GenAI.

[jen20]

This is almost certainly untrue - the vast majority of most bills for AWS (especially large ones) are EC2, EBS, Network Traffic and S3.

Startups doing machine learning on EC2 might have large bills, but that isn't what the AWS focus on generative AI is about!

[hnlmorg]

AWS focus on GenAI is about using AWS specific tools that lock you into AWSs ecosystem.

Stuff like SageMaker et al still require compute, storage, network traffic, etc.

[jen20]

Any AI startup using that kind of thing instead of their own stack is already doomed before cost.

[hnlmorg]

It’s not as simple as that. Otherwise nobody would use AWS specific tooling to begin with.

For example, some instance types are only available on SageMaker.

AWS aren’t amateurs, they know how to promote their platform effective.

[AgentOrange1234]

Interesting. I figured all their AI efforts were motivated by FOMO rather than actual returns. Why is AI stuff making so much money? Wouldn’t a new area like AI be a loss leader as they try to get market share?

[CapeTheory]

The only way cloud providers are making money on "AI" is via obscene mark-ups on access to GPUs used by companies who are deluded into thinking that (re-)training their own LLM is what their own shareholders want to see.

[hnlmorg]

The problem is that legal world are still undecided about the safety of public models.

Plus often businesses need GenAI to be trained on their own IP (ie stuff sensitive to their own business that they don’t want in the public domain).

—-

Point 1 will be decided over the next few years as creators take companies to court (or “ethical AI” starts to displace the current models trained on unlicensed content)

Point 2 cannot be resolved without training your own models.

——

Let’s also not forget that LLMs are just one part of the GenAI movement. There’s audio and image generation too (Plus video, but that’s more an extension of image). In fact it was the latter that I worked on.

And then you have other areas of AI outside of the generative space too. From hundreds of different applications of image recognition to sound processing to searching for other kinds of bespoke patterns. These are all areas I’ve worked in too.

Often a GenAI product will require multiple different “AIs” to function, as part of a larger pipeline that appears like a single opaque box to the customer. And most of those models in the pipeline likely aren’t generative, let alone LLMs.