Software Is No Longer a Moat

Sovereign Signal

For two decades, software defended itself with effort. Years of engineering, a polished interface, and a catalogue of edge cases made a product expensive to copy. That defense is eroding in public. A single open-source repository now lets an agent inspect an application without its source code, explain how a feature works, show the evidence, and build a version for someone else's project.¹

Creative Destruction Arrives in Order

Creative destruction rarely hits everything at once. It moves through the economy in a sequence, and this one has been legible. It started with labour: tasks that once justified headcount were absorbed by models. It moved next to companies whose entire product was a thin layer over a model, because the model vendor could ship the same capability as a feature. That second step is a pattern often described as platform risk.

The third step is the one that reaches SaaS. When the thing being sold is a feature, and a feature can be reconstructed from observable behavior, the price of that feature heads toward the cost of the compute that rebuilds it.

The Suites Were Cloned

Brandon Thomas's ArtCraft project is reported to have produced open-source, clean-room-style alternatives to the Adobe creative suite, followed by WordCraft, GridCraft, and DeckCraft as alternatives to Word, Excel, and PowerPoint, written in Rust with help from Claude. They are early software, and they lack decades of accumulated bug fixes. That caveat is real. It is also the kind of gap that closes quickly once the method is public.

If a feature can be rebuilt from its observable behavior, the feature is no longer the moat. It is the invoice.

The Decision Layer Became a Commodity

TypeSafe's Jev, released in mid-September, is a decision model and not a language model. It takes application state and typed questions and returns calibrated choices, not prose. Within weeks, Cloudflare shipped Clef and Clef-flash as open-weight alternatives that it trained over a single weekend, then cut Clef-flash pricing to $0.038 per million input tokens, below Jev's.³ OpenAI, AWS, and Perplexity are reported to have released comparable offerings.

The detail that matters is not any one model. A frontier-adjacent capability went from a startup launch to an open-weight, cheaper clone in under a month. Routing, classification, moderation, and guardrail checks used to be product features. They are now a priced utility.

Executives Are Saying the Quiet Part

Palo Alto Networks CEO Nikesh Arora has been reported describing most software as a pile of data wrapped in a user interface, and conceding that Wall Street's fears are closer to right than vendors admit. In the same breath he exempts cybersecurity, where the work lives in rare edge cases. NVIDIA CEO Jensen Huang has been reported saying every software company will become an agentic company, and that licensing will give way to renting specialized tokens.

The Counter-Thesis: Taste

The strongest argument against the apocalypse comes from the owner of Grand Theft Auto. Take-Two Interactive CEO Strauss Zelnick has been reported saying that the notion AI can make GTA is laughable, that AI is backward-looking because it learns from existing data, and that hits are created by genius, which data sets, compute, and language models do not equal.

Take that seriously. It suggests the remaining moat is taste: the judgment about what to build, what to cut, and what feels right. But it also sharpens the point. Taste is a property of people and teams, not of code. A vendor whose only edge was the code has nothing left to point to, while one with a distinctive product sensibility still does.

Where the Moat Went

None of this means software has no value. It means the value moved. Code, interface, and features are reproducible. What agents cannot copy from a screenshot is the system of record that holds the data, the integrations that took years to negotiate, the audit trail a regulator will accept, and the party that is liable when something goes wrong. Those are governance assets, not engineering assets.

The same shift raises the exposure on the buyer's side. An organization that replaces a licensed suite with an agent-built clone inherits the security, licensing, and provenance questions the vendor used to absorb.

Creative destruction does not ask whether an industry is ready. It reprices whatever has become cheap to reproduce. Software spent twenty years charging for the effort of building. The next decade will charge for what cannot be rebuilt on a weekend: the data, the trust, and the accountability around it.