The Numbers That Never Make It Onto the Approval Sheet
In the last installment, we covered the break-even point between an API subscription and building in-house. When we mentioned 200–300 million tokens a month — roughly 8 million KRW on the invoice — one manager in the meeting room said: "Fine, I get that. But how much more is it actually going to cost to run this for a year?"
That question is the real gate to approval. The upfront cost sits on the first page of the quote. The operational cost sits nowhere. And it usually bleeds longer — and more quietly — than the upfront figure ever did.
Not the Cost of Adopting AI — the Cost of Running It
Several 2026 benchmarks converge on roughly the same figure: 15–25% of the initial build cost recurs every year as maintenance. Korean SI quotes aren''t far off. Zoom out to a 3–5 year horizon and post-deployment lifecycle work eats 40–60% of total TCO. Some research puts it as high as two-thirds.
The problem is that this number appears nowhere on the quote. Why? Part of it is habit. Korean SI vendors have sold "the build" for decades and pushed "the operation" into separate contracts or annual renewals. Part of it is honestly hard to predict — how do you quote maintenance on a feature that doesn''t exist yet?
So the document that lands in front of the approver has the development cost in big type, with a single footnote — "* Operations billed separately" — beside it. Three years in, that footnote is half the total spend.
Four Streams of Ongoing Cost
Break down the operational bill and it splits into roughly four streams. Almost no project trips only one of them.
(a) Infrastructure and API bills that keep expanding. Public cloud looks right at first. Traffic is unpredictable in the early days, and pay-as-you-go is the safe choice. Once usage stabilizes and volume crosses a threshold, the math flips. As ZDNet Korea noted this summer, GPU hourly rates, data transfer, and storage stack up until public cloud becomes the more expensive option. Rising prices for servers, memory, and semiconductors are steepening that curve. The result, at some enterprises, is a "private / on-prem retreat." The migration cost to move back is, in practice, another hidden line item.
(b) Observability and evaluation tooling. The LLM observability market is projected to grow from $2.7B in 2026 to $9.3B by 2030 — a 36% CAGR. If those numbers feel abstract, translate them into your own line item. Langfuse starts at a Free tier, paid plans run $29/mo, enterprise at $2,499/year. Arize AX offers Free (Phoenix) and $50/mo. A small setup can start under $50/mo with Langfuse Self-Host + Helicone + OpenAI Usage. The real cost, though, isn''t the tool — it''s the human hours spent configuring it and reading the dashboards every day. That''s where maintenance actually lives.
(c) Retraining and fine-tuning. Retraining runs on signals, not schedules — "when drift is detected," not "once a quarter." In practice, it still ends up being roughly quarterly. A single small fine-tune can land in the $20–$50 GPU range. Honestly, I''ve never seen a project where that number was the problem. The real cost is the human time on data preparation, labeling, and validation. Labeler and reviewer wages typically dwarf the GPU invoice several times over. Autonomous retraining pipelines (detect → retrain → validate → deploy) are being pitched as a 2026 trend, but even the bare-minimum setup — where a human only reviews the policy — takes a couple of months to stand up.
(d) Additional development and feature expansion. This is probably the line most reliably missing from every quote. Within three months of launch, at least five features that weren''t in the original scope will get requested. "Just add this one thing." "The customer asked about it." Five of them. Each is a small piece of work; five together consume the first half of one engineer''s year.
Why "Small and Done" Ends Up Expensive
The intuition is: start small, small ops bill. Reality is otherwise. All four streams above apply regardless of scale. Observability tools still need to be configured at low traffic. Retraining still requires the same validation procedures with less data. Infrastructure still carries minimum-instance charges.
The result: even small deployments carry a "scale-independent fixed maintenance cost" every year. It''s part of why the break-even arrives later than expected. Small doesn''t mean cheap — if anything, the smaller the deployment, the larger the maintenance share of the total bill. Leave this off the approval sheet, and in year two the business unit head writes an email that reads, "is this really what it was supposed to cost?"
A New Footnote Called Compliance: the EU AI Act
In August 2026, the high-risk-system provisions of the EU AI Act came into force. Post-market monitoring and documentation of corrective action are mandatory. It looks unrelated to Korean SMBs — until you handle data belonging to an EU customer, or ship a service into the European market. That''s the moment the rule bites. The provision moves drift monitoring from "nice to have" to "no monitoring, no contract." It also converts the observability tooling cost from stream (b) from "optional" into "fixed overhead."
Only the Development Cost Reaches the Approval Sheet
40–60% of three-year TCO lives in operations, but only the development cost reaches the approval sheet. Until that habit changes, projects get reported in the black in year one and re-reported in year three as an "unexpected cost." If you''re the approver receiving a quote, don''t skim past that "* Operations billed separately" footnote on the last page — ask "so what does that actually cost?" If there''s no answer, the quote itself is incomplete.
The break-even discussion between subscription and building in-house was covered in the last installment. Today''s piece was about the items that should have been in that equation but usually aren''t.
Next installment: on top of this total cost, how do you add ROI as a number that actually clears the approval sheet — the metrics set. That''s where today''s operational-cost items will naturally reappear as the starting point for metric design.
If you''re framing this total-cost view internally, Free consultation from 5years+ is one way to start. Before the quote hits the approval sheet, we consolidate development and operating costs onto a single page.