My original notes: the seed that eventually became this article.
AI is here to stay — there are no doubts about it. And there are plenty of projects out there, and organisations as well, trying to get it adopted at enterprise scale. One common theme that Sequoia Capital recently identified was the dependence on foundation model providers like OpenAI and Anthropic. What Sequoia was essentially speaking about was sovereignty as an important consideration for enterprises, to do two things: one, keep the data in; and two, reduce the dependency on the foundation model providers.
Each one of these two things is extremely important by itself, and when done properly together they would set an enterprise up for being AI-ready as well as being protected from the very volatile AI landscape out there. Of course, this is not to say that you need to start buying your own compute, your own chips and your own GPUs. It just means that we need to look at your entire AI solutioning space in a strategic way — one that is sustainable as a business and also gives you the correct results.
Take, for example, what OpenAI and Anthropic charge you for a seat licence to use their frontier models: anywhere between $100 and $200. There is real cost involved in operating AI at scale, especially when there are millions of users on it concurrently. It is difficult to imagine that, for $100, the volume of tokens you consume is actually worth only $100 or $200. And one wonders when these prices will start meeting the real cost, since we know none of these organisations are actually making a profit — right now they are trying to capture as much of the market as possible. So when they decide to charge the user what it really costs, and to make a profit over it, you might just find enterprises paying upwards of $1,000 a month, or even $2,000 a month, which makes this whole thing very untenable — especially when you are trying to enable an entire workforce on AI tools. What revenue plays do these vendors have to then make money or profit?
One revenue option for these AI companies is to climb up the value ladder. One example was how Anthropic released multiple plugins for Claude, each specialised in a certain domain — law, for instance. This, if you recall, created a fairly big ripple in the US markets. This was an AI provider stepping up the value chain from being just a model provider to being a domain-expertise provider. This was a year ago.
Today there is news that they Anthropic might be writing their own billing software to manage customer payments. They depend on Stripe today, but pretty soon they could replace that with an in-house system - even if they say "not really". So practically anything that Anthropic currently uses third-party vendors or software for, they are in a position to build themselves.
Use an HCM? They can build their own. Use an ERP? They can build their own.
As long as they have access to the information and data they get from various businesses, they have the knowledge and the know-how to do this. Today they might say "no, we won't use the data" — but when there are billions of dollars at stake, and it becomes impossible to serve profitably, or impossible to keep prices that low, it is quite likely they will use your data and information to come up with solutions: either to sell to you, sell as you, or replace you.
This no longer seems like a matter of if — just a matter of when.
So whatever approach we take has to essentially help enterprises do two things: one, keep the data in and not feed some other entity that will at some point be able to use it to create competition; and second, reduce the dependency on external providers, because you never know when the prices are going to jump.
And the answer to all of this is sovereign AI — which should not come as a surprise. In the cloud era we called this private cloud. In the AI era we are calling it sovereign AI: AI that is set up by you, on your premises or in your cloud, and under your control.
Recent developments in the open-source community with open-weight models have shown significant gains, especially in terms of running larger, more complex and more useful models.
But sovereign AI is not a project. It is not something you get in six months or a year — it is a transformation of everything you do.
You need good clean data, and for that you need good clean data practices and governance processes in place.
You need an AI adoption framework — one that is goal-driven and that builds out your entire enterprise landscape, rather than living in small pockets.