Companies That Build Their Own AI Own Its Bills

A custom AI application carries no license fee, and AI coding tools have cut the cost of producing a first working version. That makes building look cheap at a time when SAPinsider’s 2026 benchmark finds cost pressure shaping how technology leaders judge every initiative. The development quote, though, is the one cost the company can fix in advance. What follows it is a set of running expenses that scale with use, which is why a build proposal should be judged on what it will cost to run over its life rather than on its development budget.

The License Is Replaced by a Bill That Doesn’t Stop

A subscription bundles the model, hosting, monitoring, security, integration, and updates into one price the vendor manages. Building unbundles them into line items the company pays and staffs itself, and most recur for as long as the application runs.

Deloitte’s analysis of AI token economics describes three ways enterprises consume AI, each with its own cost pattern:

• Packaged software hides token use behind a subscription fee.
• API access meters every query, which brings transparency and volatility together.
• Owned infrastructure internalizes the whole cost in GPUs, storage, networking, and energy.

A build decision is also a decision about which pattern the company takes on.

Maintenance adds a further layer. An SAPinsider’s guide to SAP extensibility notes that extensions built with SAP Build, ABAP Cloud, and key-user tools survive upgrades, while modifications to the core create ongoing technical debt. Whether a custom AI application that reaches into SAP data and processes carries that exposure depends on how it is built.

Usage Decides Which Option Is Cheaper

The comparison changes as the workload grows. Deloitte observes that the unit price of tokens is falling while total enterprise spend rises, because more users, larger models, and heavier workloads push consumption up faster than prices come down. Model choice, prompt length, and experimentation all move the bill.

At low utilization, consumption-based access is likely to be cheaper because the company pays only for what it uses. Deloitte’s simulation found the balance tips once token production passes a threshold: over three years, an on-premise “AI factory” can cost more than 50% less than API-based or neocloud options. Around half of that factory’s cost sits outside the GPUs, in networking, power, facilities, and software, which suggests the saving depends on keeping the capacity busy.

SAP’s own finance team has seen what happens when usage outruns the forecast. As SAPinsider reported, rising token consumption among employees and automated agents created what SAP executives called a triple-digit-million-dollar financial risk. SAP contained it with token caps, model routing, and tool rationalization, and it now revises its consumption forecasts as usage patterns change. A workload’s cost at launch is a poor guide to its cost at scale.

Model the Workload, Not the Project

A credible build case forecasts what the workload will cost to run over several years and compares that total with the subscription for the same period. The forecast has four inputs: expected usage volume in each year, the model tier each task needs, the hosting option that fits each volume level, and the integration, monitoring, and security work the company will have to staff.

Cost is only half the comparison. Lukas Deutsch, SAP’s chief controlling officer, and David Imbert, chief marketing officer for SAP financial management, have written that optimizing purely on cost leads to the wrong decisions and that finance needs value metrics alongside cost metrics. A workload that consumes more tokens is still the better investment if the output it produces is worth more than the tokens cost.

Buying moves the same arithmetic inside the vendor’s price. SAP’s chief technology officer and chief AI officer, Philipp Herzig, has said SAP sells about 200 of its roughly 400 AI capabilities as premium features priced in AI Units, a commercial unit SAP designed to reflect business outcomes rather than tokens. That makes the buy option easier to budget and harder to inspect. Building can still be the right choice, but the case has to rest on the running-cost forecast, with the development quote as one line in it.

What This Means for SAPinsiders

Treat the build budget as a down payment. Model years two and three of usage, infrastructure, monitoring, and maintenance before approving a build, and compare the total with the subscription over the same period. The gain is a decision made on lifetime cost.

Define the utilization point at which hosting is reviewed. Set a consumption threshold at which the company re-evaluates API access against dedicated infrastructure, using its own volumes. The gain is a cost curve the company manages before it arrives.

Build cost visibility and value metrics before the workload scales. Assign token or AI Unit budgets to the business area that owns the workload, and measure its output alongside its consumption. The gain is spending that can be defended on return.