I think AI kills cloud.

I want to be very specific about this prediction. To me, a/the “cloud” means:

  • technically, computers that someone else owns that you rent
  • colloquially, the big 3 public clouds: AWS, Azure, GCP

I don’t think the service of renting servers will cease, but I suspect the big public clouds in their current form may for a few reasons:

  • data processing and storage efficiency over past decades has gone up orders of magnitude
  • the markup of SaaS provided by the big 3 public clouds doesn’t make sense with AI
  • economics

the prediction #

This is unscientific, I’m not about to do math.

background #

I started my career in Azure ML – you can think of it as a competitor to Databricks and a host of startups like DataRobot, Domino Data Lab, and more. It was (is) a product for data scientists to train and deploy machine learning models and would go on to power infrastructure for GitHub Copilot, ChatGPT, and more.

But fundamentally it’s a first-party offering – i.e. built by Microsoft for the Microsoft cloud – to compete with third-party products like Databricks that also run on Azure (in addition to AWS SageMaker and GCP Vertex AI equivalents for cross-cloud competition). You can think of this as a hedge by Microsoft: if Databricks doesn’t play nice with Azure, use the threat of your sales org pushing Azure ML (or Synapse or Fabric or …). Use your field engineers to prototype migrations. And so on. More on this later.

Once that organization (Azure ML, or the broader AI platform org) is profitable, as Microsoft you’re in a good position as the marketplace for Azure (your own public cloud) to sell your own services against third-party. A balance of competition and financial optimization.

Azure ML is one of dozens if not hundreds of these first-party services Microsoft staffs teams for – all wrappers around the underlying compute, storage, and networking the cloud offers. While a bit juvenile to dismiss anything as simply a wrapper, all SaaS is fundamentally a database with UIs around it and the infrastructure you need to run it at the scale you need. The public clouds offer infinite scale overnight (though this is increasingly scarce) as the cost of significant markup.

technical skills compression #

LLMs (in some sense quite literally) compress technical skills & best practices into a neat package through AI agents. I have seen firsthand previously non-technical – but determined & curious – people do quite technical things. At a very junior level, which makes sense – and yet they learn just like anyone else does! That is, many people who were previously non-technical, or were just early in their journey, are now starting that journey with an accelerated learning curve due to new technology. And it tends to look similar: write some code to do something, lose code at some point, learn about version control, learn about website hosting, learn about…learn about static typing…

More technical folks can also do more – and more easily cross domains horizontally while quickly upskilling on best practices. While there are some risks, I think people tend to understimate how smart people can be in adopting new technology.

So how does this relate to the cloud? Well at some point on that journey, you realize every cloud service (or web services more generally) boils down to at the infrastructure layer:

  • compute
  • storage
  • networking

Once you understand this, and that your agent can easily write code targeting basically any infrastructure to do anything, what do you need SaaS for?

outline #

  • most companies just need a mac mini (data evidence)
  • OSS for everything but AI (and maybe that to?)
  • licensing model weights

what I’m not predicting #

I’m not predicting any of these companies will die. They may adapt and evolve.

acknowledgements #

I did at some point read: https://a16z.com/the-cost-of-cloud-a-trillion-dollar-paradox