← Blog

Software IP Just Became Worthless

Summary

A Photoshop clone hit 46,000 GitHub stars in eleven days. What that does to Adobe's $17bn subscription business, and why the moat was switching cost, not code.

RB
Roozbeh Bandpey

Co-founder and CEO of Bandpey GmbH

Outcome
A star-backed scenario model for Adobe's subscription exposure, a plain-language account of why clean-room clones and agents dissolve switching costs, and a renewal test any software vendor can apply to its own customers.
Software IP Just Became Worthless

A Photoshop clone went from zero to 46,000 GitHub stars in eleven days. The code was never what Adobe was selling.

Mario, running inside Minecraft

Someone wired Super Mario 64 into Minecraft. Not a video of Mario: the real thing, his movement, his jumps, his momentum, running around a blocky world. The developer of Mario-in-Minecraft credits Claude with writing the mod. The engine underneath comes from years of community work that turned Nintendo's shipped binary back into readable source. You bring the ROM. The AI stitches it together.

That reconstruction work used to be a monk's job. Maurice Heumann, a reverse engineer, reported in October that he rebuilt 99% of a commercial game's functions in three months, 83% of them byte for byte, with AI doing the grind. Three months. A complete game. One person.

Read that again, then ask what it means for a business whose entire value is a compiled binary nobody is supposed to be able to read.

I did not have to wait long for the answer.

Six days

The claim arrived the way most AI claims arrive now: someone had rebuilt Adobe's tools and put them online for free. I downloaded it expecting a convincing demo that falls over the moment you touch it.

It did not fall over. It had bugs. I finished my job anyway, faster than I would have in Photoshop, because I never opened a menu. I told the agent what I wanted and it found the controls for me.

Then I checked the repository. PhotoCraft: first commit 30 September, signed release 5 October. Six days. The developers rate their own work at 76% feature breadth and 45% ready for real work. They call it an alpha. For my needs, 45% was enough.

The office apps followed the same week. WordCraft and DeckCraft: started 5 October, released 7 October. GridCraft: repository created 5 October, major implementation pass the next day. Brandon Thomas and his contributors now ship a catalogue that covers Photoshop, Lightroom, Premiere, Acrobat, Illustrator, Word, Excel, PowerPoint, AutoCAD and Pro Tools.

Eleven days after the first commit, the Craft clones hold 109,000 GitHub stars

Stars per repository on 11 October 2026 · each bar is a clone of a named commercial product

Source: github.com/storytold, read on 11 October 2026. PhotoCraft repository created 30 September 2026; a GitHub API snapshot from early October showed 77 stars. Stars count accounts that bookmarked a repository, not installs or paying users.

Eleven days. More than 109,000 stars across the organisation. Photoshop's clone alone holds 46,000. A GitHub snapshot I captured in early October showed 77.

As a user: impressed. As someone who builds software for a living: uncomfortable. Both reactions are correct.

Adobe is making the wrong comparison

Here is how an incumbent reads a clone. It puts the clone next to the full product, counts the missing features and relaxes. 76% breadth. 45% readiness. Nothing to see.

Here is how I read it. I put the clone next to Tuesday's task. Tuesday's task needed 5% of Photoshop. The clone had that 5%.

A customer does not cancel a subscription because the alternative matches the product. A customer cancels because the alternative does the job. Adobe sells to professionals handling complex production files, and it sells to people like me who need one thing done by Thursday. The second group is larger. The second group just got an exit.

What the moat actually was

The lazy conclusion is that thirty years of Adobe code became worthless this month. Wrong. The code was never the moat. The law settled that years ago.

In SAS Institute v World Programming, the EU's highest court ruled that what a program does, the language it uses and the file formats it reads are not protected by copyright. World Programming watched SAS, tested it and built the same behaviour. Legal.

ArtCraft calls its work clean-room reimplementation: look at what the app does, build it again from scratch, never touch the vendor's code. Decompilation is the other route: take the shipped binary and work backwards to source, which the EU's Software Directive permits for interoperability. Both techniques are decades old. What changed is that an AI now does the boring 90% of either one in days instead of years.

"Clean room" is a claim about method, and Adobe can challenge it. I found no public action by Adobe or Microsoft against these projects as of 11 October. I expect letters. I do not expect them to matter, because the thing Adobe should be defending was never in the code.

The moat was your habits. Eight years of knowing where the menus are. Colleagues who use the same shortcuts. Files that open in one place. Leaving cost you weeks of relearning, so you paid the renewal and grumbled.

An agent does not need to learn the menus. It reads the controls, you describe the result, done. The GridCraft developers expose an MCP connection: read ranges, write ranges, evaluate formulas, build charts. Ask for a three-scenario forecast, review the workbook. You never see the ribbon. Which ribbon it was stops mattering.

That is what died this month. Not IP. Switching cost.

The arithmetic

Adobe reported $23.769 billion in revenue for the year to 28 November 2025. Its Form 10-K attributes $22.904 billion of that to subscriptions. 96.4%.

96% of Adobe's revenue is subscriptions: that is what a free clone attacks

Adobe FY2025 · USD billions · year ended 28 November 2025

Source: Adobe FY2025 Form 10-K, consolidated income statement, p. 48. GAAP figures, not cash flows. Operating income excludes non-operating income and tax. Adobe does not report revenue per application.

Adobe does not publish revenue per application. The closest line is Digital Media subscriptions: $17.389 billion, covering Creative Cloud, Acrobat, Express and the rest. That is the number a free clone attacks.

The usual move here is to assume a 10% decline and show the damage. I would rather start from the one adoption signal we have, the stars, and ask what they would have to become before a 10% decline is plausible.

Stars are bookmarks, not installs. For developer tools the users-to-stars ratio runs well past ten to one; for a consumer app with signed builds and a launcher it runs higher. Not every user was paying Adobe. So I model paying Adobe seats cancelled per star, from 1 to 10, and I price a lost seat at $420 a year, between the $19.99 Photography plan and the $69.99 Creative Cloud Pro plan.

What the stars would have to become before Adobe loses 10% of its creative subscription revenue

Share of Digital Media subscription revenue ($17.389bn) lost · stars across all Craft repositories × paying Adobe seats cancelled per star · $420 a year per seat

Source: Adobe FY2025 Form 10-K, Note 2 (Digital Media subscription revenue $17.389bn); github.com/storytold star counts on 11 October 2026; author assumptions. $420 a year sits between Adobe's $19.99 Photography plan and its $69.99 Creative Cloud Pro plan. Revenue loss = stars × seats per star × $420 ÷ $17.389bn. Costs and Adobe's responses are ignored.

Read the bottom row first. At today's 109,000 stars, the damage sits between 0.3% and 2.6% of Digital Media subscription revenue. Real money. Not a crisis.

Now read the diagonal. A 10% loss needs roughly 4.1 million cancelled seats. At five seats per star, that is 830,000 stars: 7.6 times today. PhotoCraft collected 46,000 in eleven days. Star growth decays after a launch, and I know it. I still think the Craft repositories cross a million stars within a year, and I think five seats per star is conservative once the agent does the switching for you.

At that point Adobe loses $1.7 billion of revenue and, with costs held flat, 20% of its $8.7 billion operating income. Adobe stays profitable. Adobe also loses the ability to raise prices at renewal, the lever it has leaned on for years.

My prediction: Adobe cuts list prices or bundles harder within eighteen months, and the earnings call will call it "customer value".

The people with permission to ship

A detail from the roadmaps bothers me more than the star count. The ArtCraft developers hear a complaint, change the application and publish the fix. Same day.

I know enterprise developers with the same AI tools who cannot ship one feature in a quarter. The feature needs a product manager, an architecture board and a dependency another team will release next sprint. Every person in that chain has a reasonable concern. Nobody owns the delay.

The incentive is inverted. The independent developer earns a reputation by shipping. The enterprise manager is punished for approving a change that fails and forgiven for postponing one the customer needs. Same model. Same week. One ships, the other schedules a review.

Giving that organisation a better model changes nothing. The bottleneck was never the typing.

Where the money goes

Here is the part the "free software wins" crowd misses.

If most work starts in an assistant from two or three model providers, those providers decide which application opens. You download a free tool and pay Anthropic, OpenAI or Google to choose and operate it. Fewer application subscriptions. One more dependency, and a bigger one.

So follow the whole bill, not the cancellation. A business that drops Adobe spends part of the saving on inference and integration, and it may spend more if its people can now do work they could not do before. Easy money does not disappear in this story. It changes address, from the application vendor to the model vendor.

Salesforce has understood this. Its Headless 360 exposes 60-plus MCP tools so agents can work the records directly, and its AIforce notice sets out credit billing for every successful agent call. Salesforce owns the system of record and the rules around it. A clone of the screen gives you nothing. That survives a change of interface. A vendor whose customer needs five reproducible functions has no such line to stand behind.

What I expect from here: harder renewals, discount requests, and executives who discover how much of their margin was customers not bothering to leave. Some vendors will earn their price with work nobody can reproduce. The rest priced the inconvenience of switching, and a small team writing Rust just removed it.

Thirty years of code did not become worthless. It became free to copy. For a software vendor, those are the same thing.

👉🏻

Research note: sources checked on 11 October 2026. The hands-on experience is my own. Project timelines and readiness estimates come from the developers' own roadmaps and parity assessments; star counts from github.com/storytold on 11 October. The Adobe figures are FY2025 GAAP accounts. The scenario grid uses my assumptions for seats per star and revenue per seat, holds costs constant and ignores Adobe's response. The predictions are mine.

About the author

Roozbeh Bandpey

Co-founder and CEO of Bandpey GmbH

Roozbeh Bandpey is an engineer and entrepreneur in Munich with more than 20 years of hands-on software engineering. He is the co-founder and CEO of Bandpey GmbH, which builds Facthory, Emolysis, Copolly and Inteliwert, and a Databricks Solutions Architect Champion.

Related insights and work