Thought piece

The second shift

Neil Middleton · 12 October 2026

In 2007 a small airline was offered a cloud reservation system for $500 a month and said no, because they were proud of the server under their desk. A few years later the same airline paid ten times that without blinking. The product had barely changed. What changed was that running servers in the back office had visibly stopped being the safe choice.

The same kind of shift is building again, one layer up. Enterprise systems were built on the assumption that the ground stays still, and it no longer does.

The gunpowder is a problem enterprise software never solved at a price anyone would pay: transactions that stay fast and affordable under any load, in a system safe enough to hold everything. On a single relational database, honest resilience was uneconomic. Organisations called nightly backups a recovery plan, or bought an active-active build that cost several times more and made complexity the new failure mode. Vendors knew the limit and sold past it, because admitting it would have unravelled the sale. Buyers mostly went along, because those who chose the system rarely lived with its failure. In 2016 the head of a large aviation operation declined to automate further with his supplier because it would put too many eggs in one basket, and he was right.

The accelerant is modelling. Enterprise suites promised to fit the business and then asked the business to fit them. One operator's travel stack went through a custom model invalidated by its vendor's upgrade, a booking tool that was shut down, a replacement that could not do the core job, and an expense system whose owner vanished mid-implementation.

Agentic AI is the match. Human clicking has been a rate limiter that hid how badly the old architecture copes with correlated demand, and agents remove it. The same technology turns specialist knowledge into a commodity, which is hard on firms that sell hours and kind to those that sell outcomes.

What works on unpredictable ground is affordable flexibility: never finding yourself down a road where the money is spent and nothing works. Cloud-native engineering provides it twice.

In the architecture, all state lives in one distributed store and everything around it is rebuilt from a script, so resilience becomes a dial. A live replica in a second region costs about an eighth of the hosting bill, and a lost site means a few hours of rebuilding and a second or two of data. Running the application in both regions cuts that to a brief brown-out. Accepting writes in both barely interrupts service, in exchange for looser consistency. A profitable operation can lock in the top setting, and one under pressure can step down and still be protected. On a relational stack each step is a project, and the upper ones are uneconomic.

The line that matters is absolute. Lose a second or two of transactions and you repair a few records. Lose twenty-four hours, against finance and operations systems that kept running, and nothing agrees with anything. The dial still has to be set. In September 2026 AWS confirmed that data held only in its Bahrain region could not be restored, while customers with a copy elsewhere carried on. What separated them was whether anyone had made that decision.

In the model, customisation is held as configuration and not written as code. Code breaks every time a piece moves. Configuration survives the move, and a lesson that took two years to learn at one client becomes a setting for the next. Every awkward requirement absorbed this way leaves the system more capable, which is what Taleb means by antifragile.

None of this is free, and none of it kills the incumbents overnight. It is built deep into one cloud, which is how a lean team outperforms a much larger one, and the dependence is chosen knowingly. The large systems of record keep their place at the centre. What fails is applying their rules at the edge, where the unpredictability now lives.

The opening is with sponsors who want more and have been told the safe version does not exist. A burnt market has heard every promise. What has not been shown is a recovery that has been rehearsed and dated, a cost you can see before you commit, and a supplier whose system comes out of each shock more capable.

Also: UnityTrip and SAP: a single source of truth. The investor thesis sets out what this means for the company.