Recommendation for the Next Hetzner Server

Prepared for: Omar
Current production server: Intel Xeon E5-1650 v3 · 128 GB ECC · 2×480 GB datacenter SSD
Recommended next step: Use the E5 as the always-on control plane and Targon as the elastic compute plane before buying another Hetzner server.
Best local compute option if we need one: AMD Ryzen Threadripper 2950X · 128 GB ECC · 2 TB NVMe
Premium local compute option: Intel Xeon W-2295 · 128 GB ECC · 960 GB NVMe

Executive recommendation

The recently purchased Intel Xeon E5-1650 v3 with 128 GB ECC RAM and 2×480 GB datacenter SSDs is a good fit for the persistent backend and River job system. It is not the ideal machine for every CPU-heavy task, but it does not need to be: its role should be the reliable control plane.

For heavy, bursty, isolated, or memory-intensive work, use Targon’s API-controlled confidential VMs. Targon can provide Cloud Run-like elastic execution while also offering beefier VMs, including approximately 50 GB RAM configurations, root-level VM access, encrypted storage, and hardware-attested isolation depending on the selected environment.

This changes the purchase decision: we should first connect River to Targon and measure the real workload. A second Hetzner server is still useful for predictable, always-on capacity, but it is no longer automatically necessary.

Recommended architecture

Backend API
    ↓
River + Postgres on Hetzner E5
    ↓
Classify the job
    ├── Lightweight / I/O-bound → E5 worker
    ├── CPU-heavy / memory-heavy → Targon VM
    ├── Build / test / repository task → Targon VM
    ├── High-volume always-on worker → optional Hetzner compute node
    └── Persistent service → E5 or persistent Targon deployment

Intel Xeon E5-1650 v3: production and River node

Use the first server for stable, persistent services:

@ditto-assistant/backend
Postgres
River job queues
API services
Job metadata and state
Scheduling, retries, and orchestration
Long-running Docker services
Monitoring and operational tooling
Lightweight or I/O-bound workers

River jobs often wait on Postgres, external APIs, model providers, webhooks, or scheduled events. Those jobs do not require a large CPU. The E5’s 128 GB ECC RAM and SSD storage make it a sensible durable node.

Avoid allowing large builds, repository analysis, or many CPU-heavy agent jobs to compete directly with Postgres and River on this machine.

Targon: elastic execution plane

Targon is useful here as an API-driven compute backend rather than merely an inference provider. The application can provision a VM, run a workload, collect logs and artifacts, and destroy or suspend the environment through the Targon API.

Good Targon workloads include:

Ditto Code-style repository jobs
Code Mode executions
Agent jobs
Dockerized builds and tests
Large-memory transformations
Parallel benchmark workers
Temporary development environments
Experimental services
Workloads needing stronger isolation
Short-lived or bursty compute

Suggested River queues:

default — lightweight work on the E5
llm_io — API- and model-provider-heavy work
targon — jobs dispatched to Targon VMs
builds — repository builds and tests
bench — benchmark and evaluation workloads
long_running — persistent workloads requiring a dedicated environment
maintenance — operational jobs with controlled concurrency

The E5 remains the durable coordinator. Targon supplies horsepower only when the workload needs it.

Local Hetzner compute options

Option 1: AMD Threadripper 2950X — best value local worker

16 cores / 32 threads
128 GB ECC RAM
2 TB NVMe
Approximately €106.70/month net in the cited listing
Approximately €127/month including 19% VAT, before account-specific charges
Roughly 30,000 PassMark CPU Mark in the cited benchmark data

This is the best value if we want a permanently available local worker pool for:

High-volume builds and tests
DittoBench workers
Agent execution
Docker image and repository caching
Continuous background processing
Workloads where Targon startup time or network transfer is undesirable

The Threadripper is the best complement to the E5 when the priority is parallel throughput per euro.

Option 2: Intel Xeon W-2295 — premium fast compute node

18 cores / 36 threads
128 GB ECC RAM
Current cited listing: 960 GB NVMe
Approximately €152.70/month net for the NVMe configuration
Approximately €132.70/month net for a cheaper SATA configuration
Approximately 11,099 Geekbench 5 multicore on Server Radar’s CPU page

The W-2295 is the more interesting option if we want a serious, always-on local compute node with strong per-job responsiveness and many concurrent workers. I would choose the NVMe version, not the SATA version, for agent jobs, builds, Docker layers, repositories, and test artifacts.

It costs more than the Threadripper, but it is the better choice when we value:

Lower latency for individual jobs
Strong sustained concurrency
Faster local build/test cycles
A premium, always-on execution host
Independence from Targon availability

Comparison

Option	Role	Cores / threads	RAM	Storage	Approx. monthly net
Existing E5-1650 v3	Production, Postgres, River	6 / 12	128 GB ECC	2×480 GB SSD	Already purchased
Targon API VM	Elastic, isolated burst compute	Depends on VM	Up to roughly 50 GB+ configurations	VM-dependent	Usage-based
Threadripper 2950X	Best-value local worker	16 / 32	128 GB ECC	2 TB NVMe	€106.70
Xeon W-2295	Premium local compute node	18 / 36	128 GB ECC	960 GB NVMe	€152.70

Decision framework

Use Targon first when

Workloads are bursty or unpredictable.
We need temporary VMs rather than another permanently running server.
Jobs need around 50 GB of RAM or isolated environments.
We want to scale workers up and down through an API.
We are comfortable paying for compute based on usage.
We want the E5 protected from heavy builds and agent execution.

Buy the Threadripper when

We expect frequent workloads throughout the day.
Local NVMe caching is valuable.
We want the lowest-cost always-on parallel worker.
Targon startup time, availability, or network transfer becomes a bottleneck.
We want a dedicated DittoBench/build/agent worker pool.

Buy the W-2295 when

We want a premium, always-on compute node.
Individual job latency matters as much as total throughput.
We expect many concurrent Code Mode or Ditto Code-style jobs.
We want a local fallback independent of Targon.
The higher monthly cost is justified by utilization.

Recommended rollout

Keep the E5 dedicated to backend services, Postgres, River, and orchestration.
Add a River-to-Targon execution path through the Targon API.
Start with controlled concurrency and explicit queues for builds, benchmarks, and agent jobs.
Record provisioning time, startup latency, runtime, failure rate, network transfer, and cost per job.
Reuse Targon VMs when appropriate; destroy short-lived environments after collecting artifacts.
Buy a second Hetzner server only if the measured workload shows a need for always-on local capacity.
If that need appears, choose the Threadripper for value or the W-2295 NVMe for premium responsiveness.

Bottom line

The E5-1650 v3 is not too light for River. It is appropriately sized as the durable control-plane server. Targon can provide the extra compute on demand through an API, including beefy confidential VMs that are useful for Code Mode, Ditto Code-style jobs, agents, builds, and benchmarks.

Therefore, the most rational next move is:

E5 for backend + Postgres + River, Targon for elastic heavy execution, and no immediate second Hetzner purchase until usage justifies it.

If we decide we need a permanent local compute pool, the Threadripper 2950X / 128 GB ECC / 2 TB NVMe is the value pick. If we want a premium fast compute node, the Xeon W-2295 / 128 GB ECC / 960 GB NVMe is the more powerful and responsive choice.

Checks before ordering a local Hetzner server

Hetzner’s Server Auction changes continuously. Verify the exact listing immediately before purchase:

The listing explicitly states the expected ECC memory.
The storage is NVMe if the workload involves builds, Docker, repositories, or databases.
The current price and VAT treatment are acceptable.
The location works for the intended deployment.
IPv4 requirements and any IPv4 charge are understood.
Drive health and RAID layout are checked after provisioning.
The server will be sufficiently utilized to justify a fixed monthly cost.

Next Hetzner Server Recommendation