When the model is the software
For 60 years, software was prerecorded: write code, compile, ship, run. The model is a different artifact. It generates the program at the moment of use, contextually, in response to intent rather than instruction. That changes what software is, how it is built, and what business models work on top.
The prerecorded era
From the IBM System/360 through the latest Windows release, software has been a frozen artifact. A developer writes source code. A compiler translates it into a binary. The binary is shipped to the user. The user runs it. The instructions are fixed at compile time. The machine just executes them.
This is the world Microsoft was built for. Office is a binary. Windows is a binary. Every line of behaviour was decided before the user touched the product. The product roadmap was the literal roadmap of which binaries would be shipped next.
The generated era
A modern frontier model is not a fixed program. It is a function that generates a program in response to context. Ask it to write code: it generates code. Ask it to summarise a document: it generates a summary. Ask it to drive a car: it generates control signals. The "instructions" are produced at the moment of use, contextually consistent with whatever just happened.
Jensen Huang frames this as the first reinvention of computing in 64 years. The mental model has to flip. Software is no longer a bill of materials to be authored. It is the output of a learned function running on a vast accelerator.
Source: Jensen Huang, Stanford CS153 Frontier Systems lecture, April 30, 2026 (https://cs153.stanford.edu/)
What changes when the model is the product
The unit of release changes. You no longer ship a binary; you ship a model weight, a system prompt, a tool harness, an eval suite. Each can change independently. The cadence of release flips from quarterly to continuous.
The cost structure changes. A binary has near-zero marginal cost to run. A model has real marginal cost: every request consumes tokens, electricity, and inference capacity. Pricing models that worked for software (perpetual licences, seat fees) bend under workloads that are usage-priced at the chip level.
The competitive surface changes. The frozen-binary era rewarded distribution and lock-in. The generated-software era rewards model quality, eval coverage, latency, and the integration of the model into a working product. The winning stack is the one where the model, the harness, and the product evolve together.
How to read every product decision through this lens
Any AI product roadmap is implicitly a bet on which side of the prerecorded/generated boundary the product sits. A copilot that just generates text is on the generated side. A workflow that uses a model to fill in a fixed template is mostly prerecorded with a sprinkle of generation. The buyer's economics depend on which side dominates.
The same lens applies to companies. A "software business that uses AI" is closer to Microsoft. An "AI-native business" is closer to OpenAI. Their gross margins, distribution challenges, and platform risks differ in ways the lens makes obvious.