Startups are context arbitrages

LLMs are the ultimate semantics to syntax (opens in a new tab) machines, and they have been rapidly commoditizing one of the most expensive white collar tasks: writing code. Coding agents are not perfect, but they have raised the skill floor (opens in a new tab) of software engineering dramatically. You need to be at least as good as Codex and GPT-5.4 xhigh.

The same thing is happening to buying software. The bar for a customer to buy your product is much higher, because the floor of what they can create on their own has also been dramatically raised.

We have not yet seen many large scale coding agents deployments to operationalize this, but that is coming soon. If most (all?) code will be written by LLMs, it means all software already exists within their latent space and just needs to be elicited (or brute forced, if you go down the infinite monkey theory path). Build vs buy in the future will simply be a context arbitrage between startups and their customers.

Dan Luu's "Cocktail party ideas" (opens in a new tab) essay often comes to mind:

the more common pattern is that someone with cocktail-party level knowledge of a field will give their ideas on how the field can be fixed [...] many people who have a superficial understanding of topic assume that the topic is as complex as their understanding of the topic

Salesforce is way more than a data table and an analytics page. But what more is there to it exactly? Most people don't quite know, and would be doomed to start a new CRM company or vibe code their own. The complexity is invisible until you try to rebuild it.

Luckily for startups, the value of context can also decay very quickly. Cybersecurity is a good example of this. The most elite companies know way more than their customers about what vulnerabilities are being exploited in the wild, which allows them to build better ways to protect against them, but you are only as good as the next threat you prevent. The time between gaining new context and implementing those learnings in your product needs to be almost real-time for you to stay ahead.

Another example of context arbitrage is running the product after the tokens are generated. We are seeing this today with the sandboxes market. You can vibe code your own Firecracker wrapper, but the context arbitrage that startups have is how to run these at scale with extremely high performance across clouds. Most companies don't have a lot of expertise in distributed systems on staff. The more workloads a vendor runs, the more they can learn and codify within their platform, which increases the context gap.

Startups have always been about the founders and the talent they could attract, they were never about writing better code. Your goal as a founder is to build a team that gives you a compounding context advantage over your customers and competitors. The cracked engineer of the future will be the one that can imbue the most context into the coding harness that their employer provides. And they might not even be an engineer at all.