Best VPS for LangGraph (2026): Self-Hosting the Graph Runtime

LangGraph took over as the default LangChain agent framework in 2025, and most teams using it run on LangSmith Cloud. That works fine until your bill hits 200 USD a month and the data residency questions start. Self-hosting LangGraph is straightforward, but the VPS sizing advice in the docs is conservative to the point of being misleading.

I host two production LangGraph deployments and have load-tested it on four different VPS configs. Here is what actually holds up.

Real LangGraph Footprint

The LangGraph runtime is small, around 200 MB of pure Python state machine code. The bulk comes from the LangChain ecosystem it depends on. Once you import the common integrations (OpenAI, Anthropic, Pinecone, your favorite tools), resident memory hits 400 to 600 MB before a single graph runs.

Each active graph session holds the current state plus checkpoint history. For typical conversational graphs, this is 50 to 150 MB per session. The state can grow surprisingly fast if you accumulate tool call history without trimming. Plan for state to be the dominant memory cost in long-running graphs.

VPS Comparison for LangGraph

Hetzner Cloud CCX23: The default for serious deployments

The CCX23 has the room for LangGraph plus a co-located Postgres instance for checkpoints. Postgres for graph state matters because the alternative (SQLite or in-memory) breaks down with concurrent sessions. The 16 GB RAM also handles spike scenarios when a graph branches into multiple parallel paths.

What makes this the sane pick: graph executions can hit short bursts of memory when multiple branches run in parallel. The CCX23’s dedicated CPU keeps these bursts from stretching into latency stalls.

Pros:

• Room for LangGraph plus Postgres plus Redis on one node
• Dedicated CPU stable under burst load
• NVMe matters because Postgres checkpoint writes are constant

The downside: 29.74 EUR is mid-tier pricing, not budget.

Get Hetzner: Hetzner Cloud.

Contabo VPS M: Budget production

For 8.49 EUR a month, the VPS M with 6 vCPU and 16 GB RAM fits LangGraph plus its supporting services comfortably. The shared CPU is fine until you hit concurrent graph execution, at which point you will see latency variance.

For batch agent workloads or sequential graph execution, Contabo is the right call. For interactive multi-user agent applications, the variance becomes user-visible.

Pros:

• 16 GB RAM at unbeatable price
• 200 GB NVMe leaves room for years of checkpoint history
• 6 vCPU absorbs spike load reasonably

Negative: provisioning slowness and occasional outbound rate limits, same as for other agent setups on Contabo.

Try Contabo: Contabo VPS.

DigitalOcean Premium AMD 8 GB: For US-east teams

56 USD a month with 8 GB RAM and managed Postgres available as an add-on is genuinely convenient for ops-light teams. The managed database removes one operational concern from your LangGraph deployment.

Honest take: 8 GB RAM is the floor here, and LangGraph plus Postgres co-located needs more for production. If you use the managed database, 8 GB on the app server is fine.

Get DigitalOcean: DigitalOcean.

Hetzner CCX13: For small deployments

If you run LangGraph for a single application with under 10 concurrent sessions and you use SQLite or hosted Postgres elsewhere, the CCX13 with 8 GB RAM handles it. Around 2 to 3 GB resident memory in steady state leaves headroom.

Pick this for prototypes and small production loads. Move up the moment you need concurrent agent sessions for multiple users.

What I Would Pick

For production LangGraph with co-located services: Hetzner CCX23. For a budget production setup where occasional latency variance is acceptable: Contabo VPS M. Skip LangSmith Cloud once your monthly usage hits 100 USD, the self-hosted cost stays flat from there.

The full self-hosting context sits at the SelfHostVPS comparison. LangGraph’s API has stabilized in 2026, so the hosting recommendations here should hold without major revision for a while.