Cost anatomy

The server bill was about ₩400,000. The work was not.

A small radiotherapy challenge can be cheap to host. That does not make it cheap to run.

The monthly figure is my operating recollection. I still need to reconcile it against AWS Cost Explorer, so it is not a cloud-price quote.

AccurayLab distributed cases, accepted DICOM result files, ran evaluations, stored the results, and maintained a leaderboard. Seven final plans went through the full workflow. I remember the infrastructure landing around ₩400,000 a month.

That sounds like the price of the platform until I list what the bill did not contain.

There are three different costs

CostWhat I mean
Keeping the service aliveApplication server, database, object storage, queue, logs, mail, and backups.
Evaluating submissionsCPU or GPU time, failed jobs, retries, finalist reruns, and retained artifacts.
Actually running the challengeWriting rules, checking data, arguing about metrics, dry-runs, answering people, fixing incidents, verifying winners, and producing the final record.

The first cost can be surprisingly small. The third is the product.

Why this challenge did not need an H100

Participants uploaded treatment-planning result objects. The platform did not train their models or rerun a large neural network for every submission. Most scoring work was ordinary CPU computation over DICOM and dose-volume information.

That experience does not transfer directly to 3D segmentation, pathology slides, or a medical VLM. Those may spend more on storage throughput or inference than on the web application itself.

How I would quote the next one

Line itemHow I would charge
Two-week discoveryFixed fee to inspect the data, run representative models, choose metrics, and estimate the real workload.
Build and dry-runMilestones with an agreed acceptance test.
Open operationBased on duration, participant volume, support hours, and incident responsibility.
ComputeMeasured usage or a capped allowance, shown separately.
Final verificationNumber of finalist reruns and the depth of technical or clinical review.
ArchiveRetention period, access, reporting, and deletion evidence.
I do not want to sell GPU hours. I want the GPU bill to be visible, while the service fee pays for making the challenge finish with a result people can trust.

What I need before calling ₩400,000 a case study

I need the Cost Explorer export by month and service, the exact operating dates, storage and transfer, the number of submissions and retries, and any credits or tax. A credit-card statement is less useful and contains information that should never be published.

Until I have that, the honest version is narrow: I operated one real radiotherapy result-evaluation platform at approximately that monthly infrastructure level. It says nothing by itself about the price of the work.

← All field notes