In Germany, there is a large community of people who have concerns about committing to Anthropic. Building model-agnostic is standard here. Everyone does it, even the Anthropic fans. But I really cannot think of a reasonable argument for why I should use Anthropic infrastructure for business workflows that are tightly connected to my trade secrets, my data, my entire way of working, even if I can flexibly swap the models underneath. It would have to be a very clean abstraction layer, and I wonder whether that is even possible. Here in Germany, there are many companies that are genuinely good at workflows, software engineering, infrastructure, and security. They build all of this themselves. Smaller companies and government organizations do so anyway. I consider this the right approach. Why should Anthropic be better at this than people who have also been doing it for a lifetime? I see no advantage from Anthropic. As a company, I would only consider it if I did not trust my own people to do it, or if I needed to move faster than the competition. But as soon as I am not just a SaaS company, a bank, or an insurance company, but actually manufacture physical products, I have real trade secrets, workflows I would never reveal to my competitors. So why would I let an AI giant weave itself so deeply into my workflows? I simply lack the imagination for why I would want that.
@Bianca Schulz it’s a very important question to always ask ourselves!
For me, I see it as any build vs buy decision: We can often DIY something, so the question becomes one of cost-benefit(-risk) calculus.
At the end of the day, the US AI labs’ best models are ahead of what the rest of the world has to offer. But for many use cases, we don’t need Opus or GPT-5.5, and open models are good enough.
But for me this isn’t too different to thinking about whether to enter into an enterprise agreement with AWS for cloud infrastructure, or Microsoft for Teams+Sharepoint: We’re sharing trade secrets with them too, but bound by strict agreements around confidentiality and security.
Maybe this is an idea for another article: building an agentic AI infrastructure is a software engineering and software architecture discipline. Many companies can do that and to that already, so they would be competitors to Anthropic in this regards, because you do not have to build LLMs for building the infrastructure for AI agents.
Yes, but it’s a little bit different. Clouds have a kind of abstraction layer to my business workflows. For me the question is: how clean is the abstraction layer to agentic AI infrastructure. Just from small tests on my own machine: calling a model directly is very different of using an Anthropic product. In the product is “behaviour” built in. So when I want to switch for whatever reason, the behaviour will change. And as you said: when is it really cheaper to buy it. What do you think, will it be cheaper in future, for a manufacturer to use Anthropic agentic AI infrastructure instead of building their own. If yes, I guess many would use a solution like SAP or others instead (depending on what I use already), still with the latest Anthropic models when they need it.
I thought your concern was more around what Anthropic can access initially, but I get you now, you're talking about the entrenchment of their products in companies' stacks (same as what I was talking about in the article).
There was a great episode on Lenny's Podcast last year with Brian Balfour where Brian talks about how building on top of Open AI's platform (back when they were ahead in the race 😆) was both (a) a "trap", in the sense that they'd become more restrictive over time, and (b) at the same time the right move, because building on top of the platform would get you ahead compared to not doing so. That's more or less how I think about these tradeoffs too: We can try and build everything ourselves in order to not be vulnerable, but doing so comes at a big opportunity cost in terms of time and money. It doesn't mean it's never the right move, but it's ultimately a strategic move that needs to be considered deliberately.
Just like Anthropic have their own interests at heart, I would similarly be wary of e.g. the consultancy I hire to build my AI infrastructure, or even the Head of AI/Tech who may want to DIY it internally. Too many companies DIY everything not because that's what best for the business, but because the CIO/CDO wants to build their empire internally (and similarly, too many companies give way too much of their money and control to players like Accenture 😅)
Yes, you are right. I think it depends very much on the industry and which partner you already have. The agentic AI infrastructure is not only available from Anthropic, but also from SAP, Microsoft, IBM, ServiceNow, AWS, Google etc.
It depends on the industry, the company size, the use case, the workflow, your own employees, your partners, the laws you have to follow etc.
I think cloud infrastructure was an easier business model.
Thanks Darragh!! It was super fun putting it together. The only annoying thing was having to rewrite parts every few days as Anthropic released new products or revised policy choices 😂
Some of the best analysis I've read on the market!
very insightful and enjoyable read!
In Germany, there is a large community of people who have concerns about committing to Anthropic. Building model-agnostic is standard here. Everyone does it, even the Anthropic fans. But I really cannot think of a reasonable argument for why I should use Anthropic infrastructure for business workflows that are tightly connected to my trade secrets, my data, my entire way of working, even if I can flexibly swap the models underneath. It would have to be a very clean abstraction layer, and I wonder whether that is even possible. Here in Germany, there are many companies that are genuinely good at workflows, software engineering, infrastructure, and security. They build all of this themselves. Smaller companies and government organizations do so anyway. I consider this the right approach. Why should Anthropic be better at this than people who have also been doing it for a lifetime? I see no advantage from Anthropic. As a company, I would only consider it if I did not trust my own people to do it, or if I needed to move faster than the competition. But as soon as I am not just a SaaS company, a bank, or an insurance company, but actually manufacture physical products, I have real trade secrets, workflows I would never reveal to my competitors. So why would I let an AI giant weave itself so deeply into my workflows? I simply lack the imagination for why I would want that.
@Bianca Schulz it’s a very important question to always ask ourselves!
For me, I see it as any build vs buy decision: We can often DIY something, so the question becomes one of cost-benefit(-risk) calculus.
At the end of the day, the US AI labs’ best models are ahead of what the rest of the world has to offer. But for many use cases, we don’t need Opus or GPT-5.5, and open models are good enough.
But for me this isn’t too different to thinking about whether to enter into an enterprise agreement with AWS for cloud infrastructure, or Microsoft for Teams+Sharepoint: We’re sharing trade secrets with them too, but bound by strict agreements around confidentiality and security.
Maybe this is an idea for another article: building an agentic AI infrastructure is a software engineering and software architecture discipline. Many companies can do that and to that already, so they would be competitors to Anthropic in this regards, because you do not have to build LLMs for building the infrastructure for AI agents.
Yes, but it’s a little bit different. Clouds have a kind of abstraction layer to my business workflows. For me the question is: how clean is the abstraction layer to agentic AI infrastructure. Just from small tests on my own machine: calling a model directly is very different of using an Anthropic product. In the product is “behaviour” built in. So when I want to switch for whatever reason, the behaviour will change. And as you said: when is it really cheaper to buy it. What do you think, will it be cheaper in future, for a manufacturer to use Anthropic agentic AI infrastructure instead of building their own. If yes, I guess many would use a solution like SAP or others instead (depending on what I use already), still with the latest Anthropic models when they need it.
Great builds Bianca, I love it.
I thought your concern was more around what Anthropic can access initially, but I get you now, you're talking about the entrenchment of their products in companies' stacks (same as what I was talking about in the article).
There was a great episode on Lenny's Podcast last year with Brian Balfour where Brian talks about how building on top of Open AI's platform (back when they were ahead in the race 😆) was both (a) a "trap", in the sense that they'd become more restrictive over time, and (b) at the same time the right move, because building on top of the platform would get you ahead compared to not doing so. That's more or less how I think about these tradeoffs too: We can try and build everything ourselves in order to not be vulnerable, but doing so comes at a big opportunity cost in terms of time and money. It doesn't mean it's never the right move, but it's ultimately a strategic move that needs to be considered deliberately.
Link to the episode: https://www.lennysnewsletter.com/p/why-chatgpt-will-be-the-next-big-growth-channel-brian-balfour
Just like Anthropic have their own interests at heart, I would similarly be wary of e.g. the consultancy I hire to build my AI infrastructure, or even the Head of AI/Tech who may want to DIY it internally. Too many companies DIY everything not because that's what best for the business, but because the CIO/CDO wants to build their empire internally (and similarly, too many companies give way too much of their money and control to players like Accenture 😅)
Yes, you are right. I think it depends very much on the industry and which partner you already have. The agentic AI infrastructure is not only available from Anthropic, but also from SAP, Microsoft, IBM, ServiceNow, AWS, Google etc.
It depends on the industry, the company size, the use case, the workflow, your own employees, your partners, the laws you have to follow etc.
I think cloud infrastructure was an easier business model.
Woah this is an incredible essay @Nick Zervoudis.
I’m going to have to re read it again to make sure I absorb every detail.
Thanks Darragh!! It was super fun putting it together. The only annoying thing was having to rewrite parts every few days as Anthropic released new products or revised policy choices 😂
Oh look, Microsoft is jumping to token-based pricing instead of per-seat pricing too now: https://www.thurrott.com/a-i/github-copilot/335125/report-microsoft-to-bring-token-based-billing-to-github-copilot