AI infrastructure
Three routes carry the same workload: a shared model API, managed inference on endpoints held for you, or a private environment of your own in the cloud or on premises. Which one is right comes out of your data, your volume, your obligations and who is available to run it. We make that decision on evidence, record the reasoning, and build on whichever route wins.
What this is fixing
It gets made early, by one person, on the provider they already had an account with, and it becomes the standard by staying. All six below follow from that one moment.
It was built against whichever endpoint was quickest to sign up for, and the question of what leaves the building was left until the thing was working.
A running system that cannot be approved for the data it was built to handle.Running it for one team cost a few dollars a day. What it costs across the business at full volume has never been calculated, because nothing has measured the volume.
Usage sits on a card until somebody senior notices the bill.One engineer picked it during a proof of concept on the strength of the documentation, and it became the standard by staying.
A decision worth six figures over three years, made in a signup form.A capacity commitment was signed to secure a rate, ahead of anyone establishing what would run on it or when.
A reserved cluster running at a fraction of what is being paid for.Three teams solved three problems in three quarters, each on the platform that suited that quarter, and none of them told procurement.
Several bills, several security reviews, and no way to compare output quality.Where the data has to sit was treated as a deployment detail rather than as the constraint that decides the provider.
The build is done twice, and the second one is the expensive one.The routes
All three run the model. What differs is how much of the environment is yours, what that control costs, and how long it takes to stand up. Start at the top and move down when the workload gives you a reason to.
For Work on low-sensitivity data where a running workflow matters more than owning the environment it runs in. Most builds start here, and a good number stay.
You receive A workflow in production in weeks, on metered usage you can stop.
For Open-weight models on endpoints held for you, where model choice, predictable access, or keeping your data off a shared surface is the constraint. No cluster to operate on your side.
You receive Endpoints and a model version you control, without GPU operations.
For Sensitive, high-volume or business-critical work where the environment itself has to be yours, with named access and evidence you can put in front of a regulator. It runs in a dedicated cloud tenancy or on hardware you own, whichever your obligations require.
You receive A dedicated environment, documented from compute through to control.
Moving down the list buys isolation and costs time. A build that starts at the bottom without the case for it spends its first month in procurement instead of on the workflow, and arrives at the same output the top row would have produced in weeks.
What the route changes
A production deployment is five layers deep. All five are there on every route, and four of them are built the same way whichever one you land on. The bottom row is the one the route settles.
A build that treats the bottom row as the whole project meets the other four during the compliance review, which is the expensive place to meet them.
The decision
The answer comes out of the workload and the conditions it has to run under. Where these five are written down, the route is usually obvious, and it is often a different one from what was assumed at the start. Four of the five score a provider. The fourth removes it.
We weigh the five against each other, because they disagree. The fastest route is rarely the one residency allows, and the cheapest at pilot volume is rarely the cheapest at rollout. Settling that argument is the work, and it happens inside FusionMap, at the Build-Ready tier priced below.
Which of the three it belongs on, the provider and the region, and every option ruled out with the reason it was ruled out.
Modelled at the volume you expect once it is live rather than at pilot volume, against what providers charge today.
What gets built in what order, what each part waits on, who owns it, and where the decision gates sit.
Where this sits
The route is chosen inside the engagement that needs it and carries no separate price. Five moments, and each one belongs to an offer that already owns the method and the figure.
Price
The route changes the compute bill and the configuration work. It leaves the bands alone. These are restated from the pages that own them, and they move when those pages move.
The decision
$5,500 to $7,500
About 6 weeks
FusionMap Build-Ready. The route chosen against the five inputs, the cost modelled at your expected volume, and written requirements a builder can quote from.
The build
$4,000 to $40,000+
10 to 40 weeks
FusionBuild Tier 2 or Tier 3, set by what the work touches. The route changes what gets configured. The tier is decided by sensitivity and by whether the system acts on its own conclusions.
The dedicated environment
$12,000 to $40,000+
20 to 40 weeks
The AI Operating System. One business function rebuilt inside an environment of its own, with the agents, the approvals, the controls and the dashboards inside the scope.
Figures are Canadian dollars. Where you are invoiced in another currency, the agreement carries the figure in that currency. Compute, model and storage charges are billed to you by the provider, normally in US dollars, and sit outside these figures. The full method and the three delivery tiers are on FusionBuild. The decision engagement and its two lighter tiers are on FusionMap. A whole function inside an environment of its own is the AI Operating System.
Fit
Four things, and the second is the one that gets skipped. A route chosen against a workload nobody has written down is a guess with a cost model attached to it.
Questions
They cover whether we resell compute, which provider you end up on, where your data can sit, and what the compute itself is going to cost.
No. You buy compute directly, in your own account, and you keep it if you stop working with us. We hold no reseller or partner agreement with any cloud or model provider, which means there is nothing on our side that makes one recommendation pay better than another.
The one the evidence supports, decided during the design and recorded in writing with the reasoning and with what was ruled out. Providers change their regions, their model catalogue and their pricing often enough that naming one on a web page would be out of date before it helped you.
Residency is treated as a constraint that removes providers from the list, and it is settled before the architecture is drawn rather than after. Coverage varies: some providers have no Canadian region at all, which rules them out for work that has to stay here. Where no provider meets the requirement, the private route can run on hardware you own instead. The design records which providers met your requirement and which did not.
Most workloads do not, and the ones that do usually know why already. A private environment buys isolation and costs time, so it earns its place when sensitivity, volume, reliability or model control makes the shared routes unworkable. Where that case cannot be made from the workload, we say so.
It is modelled during the decision at your expected rollout volume, against current provider pricing. Cloud, model and usage charges are billed to you by the provider, normally in US dollars, and sit outside our professional fees. Provider pricing moves, so the model is built to be rerun rather than quoted once.
Yes, and it is a shorter piece of work. The question becomes which workloads belong on what you have already bought, what the commitment actually covers at rollout volume, and which parts of the architecture have to sit somewhere else. Clients have finished this holding the same contract they arrived with.
No. It is a decision inside FusionBuild, and inside the AI Operating System where the environment is dedicated. It carries no price of its own and it is not sold on its own. This page exists because the decision deserves explaining, and the offer it belongs to is FusionBuild.
After the decision
A route with the reasoning behind it makes the build shorter, because the questions that normally surface halfway through have already been answered on paper.
The prototype that works, the quote you have been sent, the commitment somebody wants signed this quarter. We will tell you which route it belongs on, what it costs at rollout volume rather than at pilot volume, and what has to be settled before anything is built.
