An LLM subscription is like an OS

Why you cannot simply swap one LLM provider for another and expect it to work, and why employers should let employees choose among a few, as they already do for the OS.

Why is an LLM not a commodity? Why can’t you easily swap between different ones and expect them to work? I think this is part of what people on HN are debating when they ask whether AI labs have a moat.¹

An AI subscription goes beyond a model. It comes with:

• memory;
• projects;
• a coding harness, e.g. Claude Code;
• infrastructure, such as cloud instances and scheduled tasks;
• “connectivity” and personalization. MCP is one way, but the point is being connected to more personal stuff such as your email or even your social network (see Meta’s personal superintelligence), to provide a personalized AI that can automate things you otherwise can’t;
• model personality. Not to claim this is correct, but there is an association: compared to ChatGPT, Claude is flexible, smart at filling in the gaps in an ambiguous task, and obeys instructions in spirit more than literally.

One way to see why switching between LLMs is not as easy as it might seem, especially for those who use only 1% of the features they offer, is to think about Office applications. You would think switching from Microsoft Office to, say, LibreOffice or OpenOffice would be easy. After all, they are just there for typing stuff into a file, and can even save to the same file formats. However, the city of Munich thought exactly that. In its LiMux project it replaced Microsoft Office with OpenOffice (later LibreOffice), and Windows with Linux, and by 2012 had moved 12,600 of its 15,500 desktops. In 2017 the city council voted to go back to Windows and Microsoft Office by 2020, citing users unhappy with the lack of software and compatibility issues with Office documents.² If an application as simple as typing text into a file has so much sticking power, how much more would an LLM have, even if you think it is just a chatting interface (it is not)?

Now compare an LLM subscription to an OS. It is a platform, and a lot of things come as a bundled deal: OS-wide shortcuts, platform-specific behaviours (such as how the menu is searchable in macOS), bundled applications (such as Mail, Calendar, a browser), the ability to run UNIX tools and respect POSIX (macOS has an edge over Windows here, which is why WSL2 came along), etc.

Which leads to a question: why do companies offer employees a choice of OS? It is standard practice that when you are onboarded you can choose between at least macOS and Windows, and sometimes even Linux.

A friend of mine, when he worked at Sony, was pressured to use a laptop produced by the company, but they still offered him a choice of macOS.

Another example is something I read on HN or Reddit (I couldn’t find it again, so take it as hearsay). Someone insisted on using macOS for work. The CTO, maybe a new one, resisted and said it needed to be justified, i.e. there had to be a business case for it. Then he resigned, specifically mentioning this as the reason. In the end the resignation was the business case for the CTO to offer that choice, and a colleague even thanked him for “taking one for the team”.

Switching between LLM providers is like that: in principle you can, just like you can switch between OSes. The fallacy is: given an OS X, employee A is able to do their job, so employee B should be able to do their job with OS X (hey, X might be macOS!). Similarly, the productivity of an employee nowadays depends heavily on how effectively they can use an LLM provider, much as it does on how effectively they use an OS. Therefore employers should have the incentive to give employees the freedom to choose among LLM providers, as they do for the OS: a limited set for sure, but a set that covers most cases.

P.S. A corollary is that winning the education market is important. A college student who used OS X and software Y for their studies will continue to use them in their future job. Around May 2025, the top three all had some form of offer in the US. OpenAI gave college students in the US and Canada two months of ChatGPT Plus for free, claimable until the end of May. Anthropic offered students at a number of US universities Claude Pro at $1/month for 3 months, until mid-May. Google gave US college students Gemini Advanced for free until June 2026, more than a year. That may explain why Google is so aggressive in giving free subscriptions to the education market, much more generous than the others.

Footnotes

1. The question goes back at least to the leaked Google memo We Have No Moat, And Neither Does OpenAI (2023).

2. Whether those were the real reasons is disputed. The head of LiMux called it a political decision, Microsoft had moved its German headquarters to Munich the year before, and in 2020 a new coalition turned the city back towards open source.