In 2018 I helped put real-time machine learning into production inside a bank. This was before ChatGPT, before anyone in the building used the word inference in a meeting. The model was the easy part. It was ready in weeks. Getting it to serve a decision that a regulator, a risk officer and a product owner would all sign was the rest of the year.
I have thought about that ratio ever since, because it has not changed. It has just moved industries.
Installation, then deployment
Carlota Perez wrote the useful version of this in Technological Revolutions and Financial Capital in 2002. Her argument, briefly: a technological revolution arrives in two halves. First an installation period, where the new infrastructure gets built and capital rushes in and the returns are wild and uneven. Then a turning point. Then a deployment period, where the technology stops being the story and starts being the substrate everything else runs on.
Railways, electricity, and mass production all went through it. The fortunes in the installation phase were made by the people laying track. The fortunes in the deployment phase were made by everyone who could now move goods.
Applying that to AI is my analysis, not hers, so take it as analysis. But the shape fits. We have spent roughly a decade installing: the chips, the data centres, the foundation models, the capital. That work is not finished and it is still where the headlines are. It is no longer where most of the value is going to be created.
What actually changes
The uncomfortable part of a deployment period is that the skills that won the installation phase are not the skills that win the next one. Being early stops being an advantage, because everyone has the same models. What is left is unglamorous and specific:
- Knowing which decision in your business is actually worth automating
- Owning the data contract that decision depends on
- Having somewhere for the output to land that a human will actually trust
- Being able to change it next quarter without a rewrite
None of that is a model problem. All of it is an organisation problem wearing a model costume.
The tell
Here is how I know a company is still stuck in installation thinking. Ask what their AI strategy is, and they name a technology. "We are moving to agents." "We are standing up a private LLM." Those are perfectly good sentences and they answer a different question than the one asked.
Deployment thinking sounds boring by comparison. "We are cutting the time to approve a supplier from nine days to one, and the model is one of four things we changed." Nobody is going to put that on a conference slide. It is also the one that survives the budget review.
Where I am pointing this
I spent eight years doing this inside a bank, which is roughly the most deployment-constrained environment there is. Now I lead AI and robotics at a venture in deep geothermal energy, and the constraint has changed shape rather than gone away. A bad deployment used to mean a rollback. Now it means a technician driving two hours in the dark.
The reason I find the physical side interesting is that it makes the deployment problem impossible to hide. Software lets you ship something mediocre and call it a beta. Machines do not offer that. Either the thing moved or it did not.
My working bet, and it is a bet: the next ten years reward the people who are good at the last mile, not the people who were early to the model. I could be wrong about the timing. I do not think I am wrong about the direction.