One of the central arguments in a recent edition of the Superhero.tech newsletter by Wojtek Strzałkowski and Piotr Kacała is that access to model intelligence is becoming increasingly commoditized. It is a short observation with long consequences, and it matches what I have been watching from the product side for a while now - so I want to unpack it, and add my own read on where the defensible value actually sits.
The edge that keeps shrinking
Three things are happening at once, and each of them erodes the same advantage.
Frontier models increasingly overlap in their capabilities. The gap between the best model and the second-best keeps narrowing, and for most practical product work the differences are no longer the deciding factor. What one lab ships this quarter, the others match within months - sometimes weeks.
Open-weight models continue to improve. The floor keeps rising, and a growing share of use cases can be served by models anyone can download, fine-tune, and run on their own terms. The capability you once had to pay a premium for becomes something you can self-host.
And the leading AI labs are moving further into the application layer themselves. The companies that make the models are now also building the assistants, the agents, and the workflows on top of them - competing directly with the products that were built on their APIs.
Put those together and the conclusion is hard to avoid: access to a powerful model - or building a polished wrapper around one - is becoming less sufficient on its own to create a durable product. If your product's main ingredient is something everyone can buy, and the supplier is moving into your market, the ingredient is not the moat.
The question worth asking instead
The more important question is: what part of the product cannot be copied or replaced overnight?
It is worth being honest when answering it. A clean interface can be rebuilt in a sprint. A clever prompt chain can be reverse-engineered from the outputs. A feature that demos well can be matched by a competitor watching the same demo. None of these are worthless - they matter for adoption and experience - but none of them survive contact with a determined competitor and a capable model.
What does survive is everything that took years to accumulate and cannot be prompted into existence.
The value sits around the model
From my perspective, whether a company develops its own models or builds on top of existing ones, the defensible value increasingly sits around the model:
- Deep industry know-how - understanding how the work actually gets done, where the exceptions live, and which errors are expensive. This is learned in the field, not from documentation.
- High-quality, domain-specific datasets - data that reflects the real distribution of cases in a given industry, collected and cleaned over years, with the labels and context a general-purpose model has never seen.
- Integrations with existing systems and workflows - the unglamorous plumbing into the tools, formats, and processes an organization already runs on. Every integration is slow to build and slower to replace.
- Compliance with established procedures - products that fit the way regulated organizations are required to work, instead of asking those organizations to change their procedures for the tool.
- Security, governance and standards such as ISO - certifications and controls that take audits, documentation, and organizational discipline to earn, and that procurement departments increasingly treat as entry requirements.
- Traceability, auditability and repeatability of results - the ability to show how an output was produced, reproduce it on demand, and stand behind it when someone questions it months later.
- Outputs that can support verification, acceptance and commercial settlement - results reliable enough to be the basis for signing something off, releasing a payment, or resolving a disagreement between parties.
These elements are much harder to reproduce than another AI interface. A competitor with the same model and a better designer can copy your screens. They cannot copy your decade of domain data, your certified processes, or your position inside a client's workflow - at least not overnight, and anything slower than overnight is what the current pace punishes.
What raw intelligence cannot provide: trust
There is a second thing this list creates, and I think it is the more important one. These elements produce something that raw model intelligence alone cannot provide: trust.
In many industries - especially those involving infrastructure, construction, finance, healthcare or regulated operations - an impressive answer is not the product. An AI output becomes valuable when the result can be verified, reproduced, audited and used as a reliable basis for a decision, an approval, a payment or a contractual discussion.
I see this daily in my own corner of the market. In construction, a progress measurement is not interesting because a model produced it quickly; it is interesting when both sides of a contract can rely on it during settlement. The intelligence generates the result. Everything around the intelligence is what makes the result count.
That distinction is easy to miss when the demos are dazzling. A model that is right 95% of the time is a remarkable achievement and, simultaneously, a liability in any process where the remaining 5% carries legal or financial weight - unless the product around it can show which outputs to trust, and why.
From useful tools to operational infrastructure
This, to me, is the real dividing line emerging in AI products. On one side: tools that make individual people faster - genuinely useful, easily adopted, and easily swapped for whichever alternative is cheaper or newer this quarter. On the other: systems that organizations run their actual operations through, where the outputs feed decisions, approvals and money.
That is where AI products can move from being useful tools to becoming trusted operational infrastructure. Infrastructure is not chosen for being impressive; it is chosen for being dependable - and once it is woven into how an organization works, it is very hard to remove.
The intelligence will keep getting cheaper. The trust will not. If I were placing product bets right now, I would place them on the side of the product that a client cannot rip out without rewriting their own procedures - because that is the part no model release can commoditize.
Sources
- Superhero.tech newsletter - Wojtek Strzałkowski and Piotr Kacała, on the commoditization of access to model intelligence.