One developer and seven Adobe apps
Adobe spent about three decades building Creative Cloud. In early October a developer in Atlanta named Brandon Thomas put out free, open-source versions of seven of its flagship apps. PhotoCraft stands in for Photoshop, FilmCraft for Premiere, LightCraft for Lightroom, and so on through Illustrator, After Effects, InDesign and Acrobat. He wrote them in Rust with Claude Opus 5.5 doing most of the typing, published them under MIT and Apache licenses, and calls them clean-room reimplementations built from public file-format specs and watching how the originals behave rather than from Adobe’s code. He says Adobe charged him its cancellation fee one too many times, so he built his own.
Two weeks later he did the same thing to Microsoft Office. WordCraft, GridCraft and DeckCraft showed up on GitHub with their own parity scorecards: GridCraft claims 513 of Excel’s 523 worksheet functions, WordCraft sits at a self-reported 62 percent of real Word parity, DeckCraft at 79 percent of PowerPoint. The ArtCraft organization passed 82,000 GitHub stars in about two weeks.
To be fair, these are alphas. PhotoCraft’s issue tracker has close to a hundred open bugs mentioning PSD files and about fifty mentioning crashes. The Register tried WordCraft on a Surface, said it worked pretty well, and told readers to keep it away from anything that matters. Nobody is cutting a feature film in FilmCraft this quarter.
The market noticed anyway. HSBC cut Adobe from Buy to Hold on October 13 and dropped its target from $310 to $267, and the analyst named ArtCraft as the reason. The stock barely moved that day, down less than one percent, but it was already sitting at $236 against a 52-week high of $364. Adobe declined to comment on the clones. Microsoft acknowledged the question and said nothing.
The moat was always “this is hard to build”
For about twenty years the real competitive advantage in software was that software was expensive to write. Nobody put it that plainly in an investor deck. They said switching costs, workflow embedding, network effects, proprietary code. Underneath all of that, rebuilding Photoshop meant hiring a few hundred engineers for a few years, so nobody did.
That assumption broke this year, and it broke months before ArtCraft showed up. Software stocks shed more than a trillion dollars in market value in the first few months of 2026, with the IGV software ETF down about 25 percent and Salesforce down close to 30. The trade press called it the SaaSpocalypse, which is a dumb word for a real repricing. Nobody had heard of Brandon Thomas in February. What investors were pricing was the idea that a competent engineer and a frontier model could rebuild most seat-based application software in weeks instead of years.
ArtCraft is the proof of concept, which is why a buggy alpha got an HSBC downgrade. HSBC is pricing what a clone costs now, and that number is heading toward the price of tokens. Claude Opus 5.5 runs $4 per million input tokens and $20 per million output. Thomas built seven applications on that. Adobe spends billions a year on R&D.
The copying is happening at every layer, including the models themselves
The AI labs are getting the same treatment they’re enabling. In February, Anthropic accused DeepSeek, Moonshot AI and MiniMax of running an industrial-scale distillation campaign against Claude: about 24,000 fake accounts and 16 million exchanges aimed at its most differentiated capabilities, with the outputs used to train cheaper rivals. OpenAI had sent a memo to the House Select Committee on China eleven days earlier saying DeepSeek was doing the same thing with new, obfuscated methods. In June Anthropic named Alibaba. On September 8, US officials accused six Chinese labs of copying models from Anthropic, OpenAI, Google and SpaceX at industrial scale. Nobody stole any weights. They asked the model a lot of questions and trained on the answers, which is distillation, which is what the labs do to their own models every day.
I keep coming back to that one. The tool that makes Adobe copyable got copied the same way, and the people selling the tool are now filing complaints with Congress about it.
The gap between open and closed models tells you how fast that copying works. Epoch AI measured it in May: since January, the best open-weight models have trailed the closed frontier by an average of four months, or 8 points on their capability index, which is about the distance between GPT-5 and GPT-5.5. It was three months last October, so the gap widened slightly, and Epoch pins that on the compute spending gap rather than anything about data or talent. SemiAnalysis ran the same exercise across generations in August and found that each era, open models take about half as long to catch the first closed model of that era as they did the era before. Mozilla’s State of Open Source AI report put open models at roughly a third of all tokens served.
Then there’s the Google question, which already has a two-year-old answer. In May 2024 more than 2,500 pages of Google’s internal Content Warehouse API documentation leaked through GitHub, covering around 14,000 ranking attributes. Rand Fishkin and Mike King published it. It contradicted years of Google’s public statements about click signals and Chrome data. Nobody built a better Google with it. The documents show what Google collects, and even if they showed the weights too, you would still need twenty-five years of click data and an ad business big enough to pay for crawling the whole web. The algorithm was probably never the hard part.
What a clone cannot copy
Look at what Adobe reported the same month it got downgraded. Q3 revenue was $6.76 billion, up 13 percent, with $27.5 billion in annualized recurring revenue and a billion monthly active users. A company that just lost its product to a free clone does not post those numbers, so people are paying Adobe for something other than the code.
They pay for the stuff that doesn’t fit in a GitHub repo. A PSD file that opens the same way at the agency and at the printer. A plug-in ecosystem built over twenty years. A font library with the licensing already cleared. Training programs at every design school in the country. An enterprise contract with an indemnity clause, so when something breaks there’s a vendor with deep pockets to call. None of that is code, and Claude can’t one-shot any of it.
That same list explains why the SaaS selloff hit everyone but the actual disruption is picking targets. The analysts sorting through the wreckage keep landing on the same categories: companies that hold customer-specific operational data, run deterministic mission-critical workflows, carry regulatory or liability weight, or sit where money moves. A hospital doesn’t swap out the system that runs its pharmacy because a hobbyist posted a clone with 50 open crash bugs. A bank doesn’t move its core ledger to a vibe-coded alternative because the license is free. Switching costs in those environments run into the tens of millions and take years, and none of that has anything to do with features.
There’s a 1980s precedent for this. Compaq cloned IBM’s PC BIOS in a clean room and IBM lost the PC market, but not to Compaq. IBM lost it to the commodity hardware ecosystem the clone made possible. The clones themselves never made much money. Microsoft and Intel did, because they owned the operating system and the chip, and nobody could clean-room those.
In 2026 the thing underneath is compute. Epoch’s explanation for why the open-weight gap widened slightly this year is that the closed labs widened their compute lead, and the Chinese labs got accused of distillation because training from scratch at the frontier costs more than renting the answers. Everything sitting on top of the compute keeps getting cheaper to copy. The compute is the part that still costs real money.
Implosion or stronghold
So does this eat itself? The implosion case goes like this. If any application can be cloned in weeks and any model can be distilled in months, nobody can charge enough for either to fund the next generation, so the capex stops and the whole thing stalls. The IMF warned about an AI bubble the same week ArtCraft went viral, so the timing didn’t help.
I don’t buy the full version of that. The copying is real and it’s speeding up, but it keeps hitting the same wall, because the thing getting copied is the cheap part. Thomas copied Adobe’s features and none of the stuff a billion users actually depend on, like a PSD that opens the same way everywhere. DeepSeek got Claude’s answers and none of the data centers. What I think we’re watching is value draining out of the middle of the stack, meaning features and model weights, and pooling at the two ends, meaning the infrastructure underneath and the contracts and data on top.
The corporate stronghold version is wrong too, or at least it’s wrong for the companies that confused a feature set with a business. Adobe’s net new ARR was down 36 to 37 percent year over year last quarter, and management blamed its own freemium push, which means Adobe started pricing against free before anyone made it. That’s probably the right read on the next few years. The incumbents that survive will be the ones that gave away the copyable part early and charged for what was left. The ones still charging a monthly subscription plus a cancellation fee for features a frontier model can reproduce in a weekend are going to find out what their moat actually was, and some of them won’t like the answer.
My guess is this lands somewhere boring. The clones get good enough for students and small shops, which is a real chunk of Adobe’s billion users but a small chunk of its $27.5 billion in ARR. Adobe, Microsoft and Oracle keep the enterprise because the enterprise is buying insurance, and nobody buys insurance from a GitHub repo. The model labs keep fighting distillation with terms of service and congressional memos, lose a little more of that fight every four months, and make it up by owning more of the compute. And the Googles of the world keep winning search because the algorithm leaked two years ago and nothing changed.
The democratization of IP is real. I just think the IP was a smaller piece of the moat than any of us were sold.