The case for self-hosting an open-source meeting summarization stack is straightforward. Your data never leaves your infrastructure. You control the model, the prompts, the pipeline. The components, Whisper for transcription, an open-weights LLM for summarization, a queue, a storage layer, are all freely available. On paper, it is the privacy-maximalist answer.

In practice, the question is not whether self-hosting works. It does. The question is what it costs to run.

This article is the honest cost breakdown, written by people who have actually done it.

The components you actually need

A self-hosted meeting summarization stack, end to end:

Audio ingestion and storage, chunked upload handling, format normalization, durable storage with retention policy.
Transcription, Whisper (large-v3 or distil-whisper) on GPU, with speaker diarization (pyannote or similar).
Summarization, an open-weights LLM (Llama 3.x 70B, Mixtral 8x22B, or similar) running on GPU with enough VRAM and concurrency to handle queue depth.
Orchestration, queueing, retry, observability, prompt versioning, output storage.
Web frontend, upload UI, status polling, summary download, team management, auth, billing if you charge internal teams.
Operations, monitoring, on-call, GPU capacity planning, model upgrades, security patching.

Each of these is solvable. None of them is free.

What it actually costs to run, year one

The numbers below assume a mid-market team summarizing about 200 hours of meeting audio per month, call it 30 to 50 active users. The infrastructure choices are realistic but conservative; production deployments will vary.

Cost line	Year-one estimate	Notes
GPU infrastructure (transcription + summarization)	€18,000–€36,000	Reserved instance with H100/A100 capacity for steady-state plus burst
Object storage and bandwidth	€1,200–€3,000	Audio chunks, intermediate artifacts, summary archive
Orchestration and queue infrastructure	€1,800–€3,600	Managed queue, observability stack
Engineering time (build)	€60,000–€120,000	One senior engineer, 3–6 months to ship a usable v1
Engineering time (maintain)	€30,000–€60,000 / year	0.25–0.5 FTE on-call, model upgrades, patches
Prompt engineering / quality work	€15,000–€30,000 / year	Iterating on summary quality across meeting types
Compliance and audit overhead	€5,000–€20,000 / year	If you have GDPR, SOC 2, or sectoral obligations
Total year one	~€130,000–€270,000	Excluding hardware purchase if you go fully on-prem

Year two onward drops the build cost but keeps the operating cost: roughly €50,000–€110,000 per year for the same usage profile.

For comparison, a SaaS meeting summarization tool at the same usage profile typically costs between €5,000 and €30,000 per year, depending on tier, vendor, and seat count.

Self-hosting a meeting summarization stack costs roughly five to ten times what an EU-default SaaS costs at the same usage profile. The argument for it has to come from somewhere other than cost.

When self-hosting is the right call

The cost picture changes for some teams. Self-hosting is the right call when:

You have data classification rules that prohibit any third-party processing. Some defense, intelligence, and certain healthcare contexts simply do not allow external SaaS regardless of GDPR posture. If your data classification says no third party, self-hosting is the only option.
You have an existing GPU footprint and an MLOps team that can absorb the work. If the GPUs are already there for other workloads and the on-call rotation already exists, the marginal cost of adding meeting summarization is genuinely lower than the table above suggests.
You need full control of the model behavior. Specific summarization formats, custom domain vocabulary, integration into proprietary internal pipelines. SaaS tools allow some customization; self-hosting allows all of it.
You are at the scale where SaaS per-seat pricing exceeds infrastructure cost. A 5,000-seat enterprise running heavy daily usage may flip the math. Most teams are not at that scale.

For any of these, self-hosting is reasonable. The cost is real, but the alternatives are not realistic.

When SaaS is the right call

The default for most teams. SaaS is the right call when:

You do not have an MLOps team or GPU footprint. The build cost is what it is. Adding a team to support summarization is hiring, infrastructure, and a year of work. Buying a SaaS license is a Tuesday.
GDPR or sectoral compliance is satisfied by an EU-default vendor. If a SaaS vendor stores and processes in the EU, with short retention and no training on user data, that vendor satisfies most compliance frameworks. The reason to self-host evaporates.
The use case is broad team productivity, not specialized format. Status meetings, customer calls, research recordings, generic summarization. SaaS handles these well.
You want quality improvements you do not have to ship. SaaS vendors upgrade models, add summary types, and iterate on prompts continuously. You get the upgrades without the prompt-engineering hours.

For most teams, including most EU teams under GDPR, the answer is SaaS, specifically a SaaS that ships an EU-default architecture. That is the architecture EnClair ships.

A note on retention

EnClair stores audio and summaries for 24 hours, then deletes both. Hosting and processing are in Europe. We do not train models on user inputs or outputs. The full posture is on the security page. For a structured comparison of the alternatives, see the GDPR-compliant meeting summarization article.

What to take from this

"Open source is free" is not the right framing for self-hosted meeting summarization. Open source is freely licensed. Running it costs real money, hardware, time, ops, compliance, and the running costs do not look like SaaS subscription costs at the same usage. If your data classification or scale forces self-hosting, do it; the math works at scale or in regulated contexts. For everyone else, an EU-default SaaS is the cheaper, faster, and equally private answer.