Insights
The thinking behind the practice, in public.
A small body of written work on software economics, technology moats and what AI does to enterprise value.
No news, no summaries of other people’s research. Each piece answers a question that comes up in a real decision, and each one is meant to be argued with.
The Excel Problem Is Back — This Time It Builds Software
Classify what gets built by the consequence of it being wrong, not by what it cost to build.
Cheap creation is an opportunity. Unmanaged dependency is the risk. The difficulty is that nothing announces the moment a convenience became critical infrastructure.
Cheap to Build. Expensive to Fail.
The cheapest system to build can become the most expensive system to fail — because failure cost is set by the process, not by the build.
AI compresses build cost, compresses ownership cost more slowly, and does nothing at all to failure cost. Most technology decisions are still argued on the first number.
AI May Make Mission-Critical Software More Valuable, Not Less
Code becoming abundant does not make reliable systems abundant. That gap is where enterprise value moves.
The common reading of cheap software is that serious software companies are finished. The more defensible reading is that the scarce part was never the code.
Software Is Becoming Cheap. What Remains Scarce?
Six durable sources of advantage, each with a test that a company can fail.
Falling production cost does not remove competitive advantage. It relocates it. The useful work is knowing where it goes — and being able to tell the difference between owning one of those assets and merely describing one.
Technology Due Diligence Beyond Software
Ask what runs the company, not what it sells. In most non-software targets, that is where both the risk and the upside sit.
The technology question in a non-software business is rarely about the product. It is about the processes that run the company, and how much of them a competitor is about to automate.
Stop Asking Whether the Target Uses AI
Adoption is an operational fact. Exposure is a valuation fact. Only one of them belongs in the investment case.
The AI question in most diligence processes is an adoption question. Adoption is easy to observe, easy to improve after closing, and almost unrelated to whether the investment thesis survives.
The 80% Rebuild Test
Ask what a well-funded team could rebuild. Whatever survives that exercise is the actual asset.
A single counterfactual, applied honestly, separates what a company owns from what it merely built. Most of the value is in the twenty percent — if it is anywhere.
Buy the Workflow, Not the Software
Replacing a tool is a purchase decision. Replacing a workflow is an operational programme. Only the second is a moat.
The most reliable protection a software business has is not in the product. It is in the cost the customer would incur to change how they work.
The Hyperscaler Test
A category that fits inside a platform’s roadmap competes with a bundled feature, not with a peer.
Platform absorption is the risk least likely to appear on a risk register, because nobody in the process is responsible for it and it cannot be evidenced in advance.
AI EBITDA Is Not Free Money
A cost reduction available to everyone is a change in the market’s cost structure, not an increase in anyone’s margin.
Productivity gains become profit only when a company can prevent them from being competed away. Most cannot, and the value case rarely says so.
What I Would Be Careful Buying in 2026
Caution is not a view on a sector. It is a view on which questions have to be answered before the price is agreed.
Not a list of doomed categories. A list of patterns where the usual diligence answers stop being sufficient, together with the conditions under which each is perfectly investable.