The board asked build or buy. The CIO answered with a table: two columns, twelve rows, a total at the bottom. Both columns end the same way. In three years the intelligence is rented on a meter, the plumbing is either yours to maintain or theirs to reprice, and the one thing that mattered, the model of your business, sits in a format someone else controls.
The wrong question
Build or buy is a procurement question, and it has a procurement answer: whichever is cheaper this budget cycle. The question that decides the decade is different. What do you own at the end?
Models converge, and they converge quickly. Compute is a commodity with a price list. Talent moves, toward whoever is building something. What does not travel is the declared model of your own operations: your entities, your rules, the seventeen exceptions that exist because of an incident in 2011, the vocabulary your people use on the floor.[1] It is the last defensible advantage, and it compounds, which almost nothing else in an enterprise does. Every exception resolved joins the asset.
So put one question to each option. When this is done, where does the model of my business live, who can read it, and who can take it away?
Build: own the plumbing, rent the intelligence
The internal platform team builds the orchestration, the connectors, the guardrails. Two years in, you own a great deal of plumbing. The intelligence inside it is rented, on a meter, from a supplier who can alter your model, your pricing, your rate limits or your terms overnight and without consultation.[2] The inference meter is the rent of the century, and a home-built pipeline pays it exactly as a bought one does.[3]
Open weights do not settle this. Weights trained elsewhere, on data selected elsewhere, according to priorities decided elsewhere, are a superior rental agreement with an attractive licence attached.[4]
The deeper cost is where the model of the firm ends up. The business writes tickets. The platform team encodes them. Fidelity is lost at every conversion, and the loss is invisible, since nobody holds both ends of the chain.[5] After two years the declared model of your operations exists as code that the people who do the work cannot read and the people who wrote it are leaving.
Buy a US platform: rent both
You rent the intelligence and the plumbing, and you configure the model of your business inside someone else’s system. Thirty years of enterprise software have required the organisation to fit the product. One does not configure the platform; the platform configures the firm, through the accumulation of small surrenders.[6]
The configuring is done by an implementation partner billing by the day, whose incentive is the persistence of the gap that employs him, and understanding the firm is outsourced with it.[7] The model is public and it is plain: pilots and bootcamps run at the vendor’s expense, at no or low cost to the customer, to be converted into longer-term contracts and grown account by account. It is a sound business. It is a business in which the compounding asset accrues to the platform.
At the end, your ontology exists. It exists in the vendor’s format, under the vendor’s licence, readable by the vendor’s tooling. The thing that does not travel has been made to travel, for him.
The third option
Own the model of your business. Run it inside your own infrastructure. Pay on the value it produces.
The ontology is held as something you own and can inspect, in the words of your operators, rather than as weights inside a rented model.[8] It installs under your accounts, your keys and your jurisdiction; we operate no hosting for your data, and it can run fully closed, on a model you host yourself, a setting you verify from your own network.[8:1] And the invoice arrives after the result: a share of the value created, payable on production, because hours, tokens and seats are three ways to bill at the door and none of them commits the vendor.[9]
Three things Galahad refuses, and why. To host your data, because a trail you must request is not yours. To hold your model in a format you cannot read, because what compounds must compound for you. To bill before production, because a vendor paid at the door has already been paid for your pilot.
What it costs
Build costs a team, forever, plus the meter. Buy costs a licence, an integrator, plus the meter. Both are paid whether the thing works or not.
The third option costs a share of what it is worth, and nothing if it is worth nothing. That is not cheaper. It is the only price that puts the vendor’s outcome on the same side of the table as yours.
Questions to put to a vendor
Put these in writing, before the demonstration.
- The day the contract ends, in what format does the model of my business leave, and who can read it?
- Where does inference run, and who can reprice it, throttle it or cut it overnight?
- Who writes the rules: my operators, or your integrator?
- Can this run fully closed, on my network, on a model I host?
- What do I pay before anything is in production?
- What happens to what you learned inside my walls?
A vendor who answers all six leaves you owning something. A vendor who answers with a roadmap leaves you renting it.
Monarch is the third option: the model of your business as an asset you own, installed inside your infrastructure, paid on the value it produces. See it on your data.