A certain characteristic of legacy desktop apps, like Microsoft Office, Autodesk AutoCAD, Adobe Photoshop and so on, are that they have crappy proprietary file formats. In 2025, we barely have reliable, fully-supported open source libraries to read and write .DOCX, .XLSX, .PPTX,1 .DWG, .PSD and so on, even though related products keep making billions in revenue.

The reason is simple: Moat through obfuscation.

The business model for these products when they first appeared in the 1980s and 1990s was to sell the compiled binaries for a one-time fee. This was pre-internet, before Software-as-a-Service (SaaS) could provide a reliable revenue stream. Having a standardized file format would have meant giving competitors a chance to develop a superior product and take over the market. So they went the other way and made sure their file formats would only be read by their own products, for example by changing the specifications in each new version. To keep their businesses safe, they prevented interoperability of entire modalities of human work, and by doing so, they harmed the entire world’s economy for decades.2

Can you blame them? The only thing they could monetize was the editor. Office 365 and Adobe Creative Cloud has since implemented a SaaS model to capitalize even more, but the file formats are still crap—a vestige of the old business model.3

But finally, a revolution is underway. This might all change.

None of these products were designed to be used by developers. They were designed to be used by the “End User”. According to Microsoft, the End User does not care about elegance or consistency in design.4 The End User could never understand version control. The End User sends emails back and forth with extensions such as v1.final.docx, v1.final.final.docx. Until recently, the End User was the main customer of software.

However, we have a new customer in the market: AI. The average AI model is very different than Microsoft’s stereotypical End User. They can code. In fact, models have to code, or at least encode structured data like a function call JSON, in order to have agency. Yes, we will also have AIs using computers directly like OpenAI’s Operator, but it is generally more straightforward to use an API for an AI model than to use an emulated desktop.

We will soon witness AI models surpass the human End User in terms of economic production. Tyler Cowen5, Andrej Karpathy6 and others are convinced that we should plan for a future where AIs are major economic actors.

“The models, they just want to learn”. The models also want intuitive APIs and simple file formats. The models abhor unnecessary complexity. If you have developed a RAG pipeline for Excel files, you know what I mean.

If AI creates pressure to replace legacy file formats, then what can companies monetize if not the editor? The answer is the AI itself. Serve a proprietary model, serve an open source model, charge per tokens, charge for inference, charge for kilowatt-hours, charge for agent-hours/days. The business model will differ from industry to industry, but the trend is clear: value will be more and more linked to AI compute, and less and less to Software 1.07.

There is now a huge opportunity in the market to create better software, that follow the File over App philosophy:

if you want to create digital artifacts that last, they must be files you can control, in formats that are easy to retrieve and read. Use tools that give you this freedom.

We already observe that AI systems work drastically more efficiently if they are granted such freedom. There is a reason why OpenAI based ChatGPT’s Code Interpreter on Python and not on Visual Basic, or why it chose to render equations using LaTeX instead of Office Math Markup Language (OMML)8. Open and widespread formats are more represented in the datasets, and the models can output them more correctly.

There is going to be an AI-native “Microsoft Office”, and it will not be created by Microsoft. Copilot is not it, and Microsoft knows it. Boiling tar won’t turn it into sugar. Same for other Adobe, Autodesk and other creators of clutter.

Internet Explorer’s 2009 YouTube moment is coming for legacy desktop apps, and it will be glorious.

  1. Yes, Microsoft’s newer Office formats .DOCX, .XLSX, .PPTX are built on OOXML (Office Open XML), an ISO standard. But can all of these formats be rendered by open source libraries exactly as they appear in Microsoft Office, in an efficient way? Can I use anything other than Microsoft Office to convert these into PDF, with 100% guarantee that the formatting will be preserved? The answer is no, there will still be inconsistencies here and there. This was intentional. A moment of silence for the poor souls in late 2000s Google who were tasked with rendering Office files in Gmail and Google Docs. ↩

  2. For a recent example of how monopolies create inferior products, imagine the efficiency increase and surprise when Apple Silicon (M1) first came out, and how ARM is now the norm for all new laptops. We could have had such efficiency a decade before, if not for Intel. ↩

  3. On the other end of the spectrum, we have companies that are valued in the billions, despite using standardized open source standards: MongoDB uses Binary JSON (BSON), Elasticsearch uses JSON, Wordpress (Automattic) uses MySQL/PHP/HTML,CSS, and so on. ↩

  4. Companies like Notion beg to differ: Software should be beautiful. People apparently have a pocket for beauty. ↩

  5. “Should you be writing for the AIs?” ↩

  6. “Be good. Future AIs are watching.” ↩

  7. Traditional pre-AI software, as opposed to Software 2.0. ↩

  8. Long forgotten format for Microsoft Equation Editor. ↩