Public Cloud — section overview

Q: Cloud instance or VPS — what is the difference?

Both run on the same platform and are billed in the same way: per hour, capped at the monthly price. The difference lies in the allocation of the core. As a VPS, you book the series with shared cores, NX, NP and NA, from €5.99 net per month. The series with dedicated cores, ND, NM and NS, start at €38.80. The rule of thumb from operations: if the average utilisation of a vCPU is permanently above 20% over a week, the dedicated series is cheaper per unit of work performed.

Q: How fast is an instance ready?

From the call to the first SSH connection takes an average of 35 seconds for standard images, guaranteed under 60 seconds. The state changes via BUILD and BOOT to ACTIVE, each transition is timestamped in the project's activity log. Instances that a scaling group pulls in are in the same order of magnitude.

Q: What does an instance cost if it only runs for a few hours?

Billing is per started hour at the hourly rate. The hourly rate is the monthly price divided by 730 — the average month length, not that of the respective calendar month. The count is capped at the top: if the sum of an instance's hours in the calendar month reaches the monthly price, it ends. A month with 744 hours therefore never costs more than one with 720.

Q: Do I continue to pay for a stopped instance?

Yes. A shut-down instance continues to occupy CPU quota, memory and local NVMe on the host and is therefore still billed. Only deletion ends the billing. If you want to keep the state, create a snapshot beforehand, an image of the instance at this point in time. Snapshots appear on the invoice as a separate item.

Q: What counts as egress, and what does it cost above the included volume?

Egress counts as outbound traffic to the public internet. Inbound traffic, traffic between instances via the private network — including between locations — and access to Object Storage from instances in the same region are not counted. Depending on the instance type, 20 TB to 200 TB of outbound traffic per month are included. Above that, we charge €1.09 per terabyte, without a zone tariff: A terabyte to a recipient on the other side of the world costs the same as one within Germany. The exception: NG-RTX4000, NG-RTX6000, NG-RTX6000-512 and NG-RTX6000-768 do not count outbound traffic at all — traffic is unlimited there.

Q: How do I cancel an instance?

Without a term commitment, you delete the instance in the console, via the API or the command line. There is no notice period; billing ends at the time of deletion (Section 11 (3) of the GTC). Stopping is not sufficient for this. Contracts with a 12 or 24-month term can be cancelled with a notice period of 30 calendar days to the end of the term, otherwise they are extended by twelve months (Section 11 (2) of the GTC). Cancellations require text form; the cancellation function in the customer area fulfils this.

Q: What is the difference between shared and dedicated vCPUs?

In the shared families, the hypervisor, the host's virtualisation layer, schedules multiple instances on the same physical core. If you do not use your quota, you pass the share on to your neighbours — in return, the price per vCPU is lower. Waiting time caused by neighbours is visible in the guest system as steal time — the time during which the core currently belongs to a neighbour. On average over the month, it is below 3% on the shared families, and up to 12% during peak loads on individual hosts. On dedicated families, it is zero by design.

Q: Is data retained on the local NVMe?

The specified capacity is physically located in the same server as the instance and is tied to its lifespan. It survives restarts, but not the deletion of the instance. For data that must outlive an instance, Block Storage or Object Storage belong underneath. Both are separated from the compute node and can be attached to a new instance.

Q: Can I resize an instance later?

Yes. Resizing is a logged action like start, stop and deletion and is listed with a timestamp in the project's activity log. From this timestamp, the hourly rate of the new size applies. A change of family is not covered by this, because the local NVMe is bound to the instance. The route via a snapshot and a new instance is intended for this.

Q: Is every instance type available at every location?

No. 39 instance types are distributed across 5 locations in Germany and Finland. Storage and GPU designs require special chassis and power density and are therefore only located where the corresponding racks are installed. The catalogue price applies to Frankfurt. Every other location carries a factor that reflects the electricity price, network connection and local taxes; the table in the "Billing" section calculates it using an example.

View from space of the night side of the Earth. The globe fills the lower half of the image; its landmasses can only be recognised by the city lights, which lie like golden nets of dots on a dark background — North America on the left, Europe and Africa in the middle, the edge of Asia on the right. Above the horizon is a narrow, glowing purple arc of the atmosphere, and above that a deep purple starry sky. Stylised clouds of translucent, glowing purple-white material float above the globe: two smaller ones on the left, one large one on the right. Thin, bright arcs run down from each cloud to the globe, ending there in small green glowing dots; further arcs span between the clouds. The centre of the image remains empty.

Instances, storage, network, databases, AI endpoints and developer tools on one control plane — in the same projects and in the same private network.

Configure instance Price list

Entry: €5.99$6.95; NX2, monthly net
Instance types: 39; 7 series in 6 families
Largest instance: 2 TB; Memory on NG-H200×8
Locations: 5; in Germany and Finland

Your path to the Public Cloud

Three decisions in the order they actually arise: which machine, how it scales, and where it is located.

Three identical server racks stand side by side in a dark, purple-lit room, seen straight from the front. Each is a tall floor-standing rack made of matte dark steel with a perforated front door, behind which stacked device bays are located. The three show three states of the same device: the left rack is dark, only a single dot glows deep inside. In the middle one, the lower half of the bays glows in rich purple, the upper half is still dark. The right one is fully powered — all bays glow from purple to an almost white core, the light shines through the perforation to the front and lies as an elongated reflection on the polished floor in front of it. A tiny lime-green status dot sits on each glowing bay. Further unlit racks stand on the left and right, fading into the darkness.

Deploy

You choose a load profile and a size: 39 instance types range from a shared vCPU — a virtual processor core shared by multiple instances — for a development environment to a node with eight GPUs. Entry starts at €5.99$6.95 net per month.

From the call to the first SSH connection, it takes on average 35 seconds, guaranteed under 60 seconds for standard images, the operating system images maintained by us. The state changes via BUILD and BOOT to ACTIVE. Every transition is recorded with a timestamp in the project's activity log.

All instance types compared

A single straight row of identical server racks runs diagonally into the depth of a dark, purple-lit room. The front one stands large and close, every detail is readable — perforated door, edge shadows, feet —, and it glows the brightest, with an almost white core behind the front. Behind it, the same rack follows again and again at equal distances, each slightly smaller and glowing less brightly, until the rearmost ones are just purple blocks of light in a thin ground fog, converging into a single glow at the vanishing point. No rack deviates in design from the front one. The polished floor beneath each bears its own, shrinking field of reflected light. A tiny lime-green status dot sits on each of the front racks.

Scale

A scaling group consists of an instance template, a capacity range and rules. The evaluation runs every 30 seconds. The shortest cooldown — the block period after an adjustment during which no further adjustment is triggered — is 60 seconds. A group comprises 1 to 200 instances.

Where a share of a server is no longer sufficient, an entire node without a hypervisor is available in the same project and in the same private network, i.e. without the virtualisation layer between hardware and operating system. The “Metal” section draws the line row by row.

Auto scaling in detail

Stay within EU jurisdiction

The cloud runs at 5 locations in Germany and Finland, 4 of them in Germany. Not every family is available at every location: storage and GPU designs require special chassis, i.e. specially designed cases.

Outbound traffic has no zone tariff, i.e. no separate rate per destination region. At least 20 TB are included per instance, above which a single rate of €1.09$1.26 per terabyte applies — regardless of where the recipient is located. Inbound traffic is always free of charge.

Availability per location

The entire section, sorted by tasks

11 groups from compute to developer tools. All are billed according to the same model, all can be controlled via the same API, and all are located in the same projects.

Configurators for this section

View across the full width along a row of black server racks. A busbar runs on galvanised brackets above the rack heads: a straight aluminium housing, from whose open underside two bare copper bars are visible, with evenly clipped-in grey tap-off boxes, from each of which a short hose leads into a rack. On the left, two racks are open and completely filled with identical 1U servers, each with a green status light. In the centre of the image, the smooth flank of a rack is turned towards the camera; on it sits a large transverse brushed aluminium plaque with four countersunk screws and the embedded inscription ENTRONYX CLOUD. On the right, the row ends at a light wall above a light raised floor. — All product groups in this section run in the same rows, on the same hardware family and on the same power rail. They are separated by allocation and guarantee, not by location — which is why traffic between them in the private network incurs no transit costs.

Price up your instance before you order it

Series, size, location, network and the traffic you expect — the amount alongside updates after every choice. What you select here is carried over to the configurator as the default.

Always included

at least 20 TB egressper instance and month
IPv4 and IPv6 /64included in the instance price
no setup feewith every instance type
billed per hourcapped at the monthly price
Deliveryon average 35 seconds, guaranteed under 60 seconds

Cost-optimised for development, small services and bursty workloads. CPU time is shared fairly.

vCPU

€5.99$6.95

NXIntel/AMD x86NPAMD EPYC

2 vCPU16 vCPU

NX24 GB RAM40 GB NVMe20 TB traffic

Location

Reference location — the catalogue price applies without a factor.

Public network

included

Primary IPv4 + IPv6-/64DefaultIPv6-/64 only− €0.50$0.58 per month

The primary IPv4 address is included in the instance price. If you deselect it, you will only be reachable inbound from IPv6 networks — and you can add it later at any time.

Expected traffic

€0.00$0.00

1 TB200 TB

20 TB included. Each additional outbound terabyte costs €1.09$1.26 net, without zone pricing. Inbound traffic is always free.

Choose the load profile first, then the size

The series identifier tells you how CPU time is allocated and what core-to-memory ratio you get. The number after it is just the size.

NXNP

Shared vCPU

Cost-optimised for development, small services and bursty workloads. CPU time is shared fairly.

NX runs on Intel/AMD x86, NP on AMD EPYC with twice as much local NVMe per vCPU.

For services that are rarely under full load for long

Web frontends and blogs with low traffic
Development and staging
Cron jobs and short build runs

Types: 10
vCPU: 1–16
RAM max.: 32 GB
from · net per month: €5.99$6.95

Configure

Dedicated vCPU

A physical core belongs exclusively to your instance. Constant performance without steal time — and from 64 vCPU, it is worth looking at TITAN bare metal.

A physical core is permanently assigned via cpuset. No neighbour, no quota, no steal time.

For consistent continuous load without neighbours on the core

Production databases
API backends with latency requirements
Kubernetes workers and CI runners

Types: 7
vCPU: 2–128
RAM max.: 512 GB
from: €38.80$45.01

Configure

Memory-optimised

8 GB RAM per vCPU for in-memory databases, caches and analytics.

8 GB memory per vCPU, on dedicated cores.

For data that must reside entirely in memory

In-memory databases
Caches with a large working set
Analyses across entire datasets

Types: 5
vCPU: 2–32
RAM max.: 256 GB
from: €48.40$56.14

Configure

Storage-optimised

Up to 7.68 TB local NVMe directly attached to the instance — without network latency.

NVMe is attached directly to the instance's PCIe bus. No network storage, but bound to the instance.

For workloads that depend on local NVMe

Search indices
Delivery nodes with local cache
Replicated databases with a high write ratio

Types: 3
vCPU: 4–16
RAM max.: 64 GB
from: €220.00$255.20

Configure

ARM / Ampere

Shared Ampere Altra cores for ARM64 workloads built and tested on ARM. Windows Server does not run on this series.

Shared Ampere-Altra cores for arm64 software, from €6.49$7.53 net per month.

For software available as arm64

Container fleets
Horizontally scaled web services
Build nodes for arm64 images

Types: 4
vCPU: 2–16
RAM max.: 32 GB
from: €6.49$7.53

Configure Book ARM as VPS

GPU

RTX PRO 4000 (24 GB GDDR7) up to H200 cluster (8 × 141 GB HBM3e) for inference, fine-tuning and distributed training — single cards plus NVLink scaling for H100/H200.

Training costs, allocation model and availability are on the GPU page.

For computing on the GPU

Inference
Fine-tuning
Distributed training

Types: 10
vCPU: 8–128
RAM max.: 2 TB
from: €214.00$248.24

Configure GPU instances in detail

Six profiles, seven series: NX and NP share the “Shared vCPU” profile because both are scheduled according to the same fair-use rule — they differ in platform and local storage, not in the allocation model.

Hardware per series

One row per series: processor, ranges and the amounts you can use to calculate without a calculator. Catalogue prices for Frankfurt, net and without a fixed term; the cost check above calculates the location factor and traffic.

Hardware per series: processor, generation, clock speed, ranges, connection and net starting prices for Frankfurt
Series	Processor	Generation	Clock speedBase / boost	vCPU	RAM	NVMelocal	Connection	fromper month	per hourEntry type	per vCPUper month, cheapest type
NXShared vCPU · 4 typesConfigure — series NX, from NX2	Intel/AMD x86vCPU shared	—	—	2–16	4–32 GB	40–320 GB	—	€5.99$6.95	€0.0082$0.0095	€1.87$2.17
NPShared vCPU · 6 typesConfigure — series NP, from NP1	AMD EPYCvCPU shared	—	—	1–16	2–32 GB	40–640 GB	—	€11.99$13.91	€0.0164$0.0191	€8.16$9.46
NDDedicated vCPU · 7 typesConfigure — series ND, from ND2	AMD EPYCvCPU dedicated	—	—	2–128	8–512 GB	80 GB – 2.4 TB	—	€38.80$45.01	€0.0532$0.0617	€16.69$19.36
NMMemory-optimised · 5 typesConfigure — series NM, from NM2	AMD EPYCvCPU dedicated	—	—	2–32	16–256 GB	80–960 GB	—	€48.40$56.14	€0.0663$0.0769	€24.16$28.02
NSStorage-optimised · 3 typesConfigure — series NS, from NS3	AMD EPYCvCPU dedicated	—	—	4–16	16–64 GB	1.92–7.68 TB	—	€220.00$255.20	€0.3014$0.3496	€55.00$63.80
NAARM / Ampere · 4 typesConfigure — series NA, from NA2	Ampere AltravCPU shared	—	—	2–16	4–32 GB	40–320 GB	—	€6.49$7.53	€0.0089$0.0103	€2.59$3.01
NGGPU · 10 typesConfigure — series NG, from NG-RTX4000	AMD EPYC · Intel Core i5-13500 (14 cores, 20 threads) · Intel Xeon Gold 5412U (24 cores, 48 threads)vCPU dedicated	—	—	8–128	48 GB – 2 TB	400 GB – 30.72 TB	—	€214.00$248.24	€0.2932$0.3401	€10.70$12.41

NXShared vCPU
from €5.99$6.95
Intel/AMD x86vCPU shared
vCPU
2–16
RAM
4–32 GB
NVMe
40–320 GB
per hour
€0.0082$0.0095
per vCPU per month
from €1.87$2.17
Generation · clock speed · connection: — · — · —
Configure — series NX, from NX2
NPShared vCPU
from €11.99$13.91
AMD EPYCvCPU shared
vCPU
1–16
RAM
2–32 GB
NVMe
40–640 GB
per hour
€0.0164$0.0191
per vCPU per month
from €8.16$9.46
Generation · clock speed · connection: — · — · —
Configure — series NP, from NP1
NDDedicated vCPU
from €38.80$45.01
AMD EPYCvCPU dedicated
vCPU
2–128
RAM
8–512 GB
NVMe
80 GB – 2.4 TB
per hour
€0.0532$0.0617
per vCPU per month
from €16.69$19.36
Generation · clock speed · connection: — · — · —
Configure — series ND, from ND2
NMMemory-optimised
from €48.40$56.14
AMD EPYCvCPU dedicated
vCPU
2–32
RAM
16–256 GB
NVMe
80–960 GB
per hour
€0.0663$0.0769
per vCPU per month
from €24.16$28.02
Generation · clock speed · connection: — · — · —
Configure — series NM, from NM2
NSStorage-optimised
from €220.00$255.20
AMD EPYCvCPU dedicated
vCPU
4–16
RAM
16–64 GB
NVMe
1.92–7.68 TB
per hour
€0.3014$0.3496
per vCPU per month
from €55.00$63.80
Generation · clock speed · connection: — · — · —
Configure — series NS, from NS3
NAARM / Ampere
from €6.49$7.53
Ampere AltravCPU shared
vCPU
2–16
RAM
4–32 GB
NVMe
40–320 GB
per hour
€0.0089$0.0103
per vCPU per month
from €2.59$3.01
Generation · clock speed · connection: — · — · —
Configure — series NA, from NA2
NGGPU
from €214.00$248.24
AMD EPYC · Intel Core i5-13500 (14 cores, 20 threads) · Intel Xeon Gold 5412U (24 cores, 48 threads)vCPU dedicated
vCPU
8–128
RAM
48 GB – 2 TB
NVMe
400 GB – 30.72 TB
per hour
€0.2932$0.3401
per vCPU per month
from €10.70$12.41
Generation · clock speed · connection: — · — · —
Configure — series NG, from NG-RTX4000

“—” means: the catalogue does not list the value per series. The vCPU series run on hosts without a fixed processor model; a model with cores and threads is only listed where the hardware of a node is fixed. We state generation, clock speed and connection as soon as they are documented per series.

Images

12 operating systems are available for cloud instances in the configurator. Without a licence specification, an image costs nothing extra.

Ubuntu24.04 LTS
Ubuntu26.04 LTS
Debian13 (Trixie)
Debian12 (Bookworm)
Rocky Linux10
AlmaLinux10
Fedora Server43
openSUSE Leap16.0
Talos Linux1.14· from 4 GB RAM
Flatcar Container Linuxstable
FreeBSD15.1
Windows Server2025 Standard· from 8 GB RAM· licence €28.00$32.48 per month

All instance types compared

Prices apply to Frankfurt. The hourly rate is the monthly price divided by 730 hours and rounded to four decimal places.

The table is wider than the screen and can be scrolled horizontally.

39 Instance types displayed· cheapest row €5.99$6.95 per month

All cloud instance types with vCPU, CPU type, memory, local NVMe, included traffic, and hourly and monthly price
Instance	Family	vCPU	CPU type	RAM	Local NVMe	Traffic	Price / hour	Price / month
NX2Entry	NX	2	Intel/AMD x86 (shared)	4 GB	40 GB	20 TB	€0.0082$0.0095	€5.99$6.95
NX3	NX	4	Intel/AMD x86 (shared)	8 GB	80 GB	20 TB	€0.0123$0.0143	€8.99$10.43
NX4	NX	8	Intel/AMD x86 (shared)	16 GB	160 GB	20 TB	€0.0226$0.0262	€16.49$19.13
NX5	NX	16	Intel/AMD x86 (shared)	32 GB	320 GB	20 TB	€0.0411$0.0477	€29.99$34.79
NP1	NP	1	AMD EPYC (shared)	2 GB	40 GB	20 TB	€0.0164$0.0190	€11.99$13.91
NP2	NP	2	AMD EPYC (shared)	4 GB	80 GB	20 TB	€0.0274$0.0318	€19.99$23.19
NP3Bestseller	NP	4	AMD EPYC (shared)	8 GB	160 GB	20 TB	€0.0493$0.0572	€35.99$41.75
NP4	NP	8	AMD EPYC (shared)	16 GB	320 GB	20 TB	€0.0959$0.1112	€69.99$81.19
NP5	NP	12	AMD EPYC (shared)	24 GB	480 GB	20 TB	€0.1383$0.1604	€100.99$117.15
NP6	NP	16	AMD EPYC (shared)	32 GB	640 GB	20 TB	€0.1788$0.2074	€130.49$151.37
ND2	ND	2	AMD EPYC (dedicated)	8 GB	80 GB	20 TB	€0.0532$0.0617	€38.80$45.01
ND3Production	ND	4	AMD EPYC (dedicated)	16 GB	160 GB	30 TB	€0.1063$0.1233	€77.60$90.02
ND4	ND	8	AMD EPYC (dedicated)	32 GB	240 GB	30 TB	€0.1904$0.2209	€138.99$161.23
ND5	ND	16	AMD EPYC (dedicated)	64 GB	360 GB	40 TB	€0.3788$0.4394	€276.49$320.73
ND6	ND	32	AMD EPYC (dedicated)	128 GB	600 GB	50 TB	€0.7315$0.8485	€533.99$619.43
ND7	ND	48	AMD EPYC (dedicated)	192 GB	960 GB	60 TB	€1.1698$1.3570	€853.99$990.63
ND8	ND	128	2× AMD EPYC (dedicated)	512 GB	2.4 TB	80 TB	€3.4016$3.9459	€2,483.20$2,880.51
NM2	NM	2	AMD EPYC (dedicated)	16 GB	80 GB	20 TB	€0.0663$0.0769	€48.40$56.14
NM3	NM	4	AMD EPYC (dedicated)	32 GB	160 GB	30 TB	€0.1324$0.1536	€96.65$112.11
NM4In-Memory	NM	8	AMD EPYC (dedicated)	64 GB	320 GB	30 TB	€0.2648$0.3072	€193.30$224.23
NM5	NM	16	AMD EPYC (dedicated)	128 GB	480 GB	40 TB	€0.5300$0.6148	€386.90$448.80
NM6	NM	32	AMD EPYC (dedicated)	256 GB	960 GB	50 TB	€1.0590$1.2284	€773.07$896.76
NS3	NS	4	AMD EPYC (dedicated)	16 GB	1.92 TB	40 TB	€0.3014$0.3496	€220.00$255.20
NS4	NS	8	AMD EPYC (dedicated)	32 GB	3.84 TB	50 TB	€0.6027$0.6991	€440.00$510.40
NS5	NS	16	AMD EPYC (dedicated)	64 GB	7.68 TB	60 TB	€1.2055$1.3984	€880.00$1,020.80
NA2Efficient	NA	2	Ampere Altra (shared)	4 GB	40 GB	20 TB	€0.0089$0.0103	€6.49$7.53
NA3	NA	4	Ampere Altra (shared)	8 GB	80 GB	20 TB	€0.0151$0.0175	€10.99$12.75
NA4	NA	8	Ampere Altra (shared)	16 GB	160 GB	20 TB	€0.0294$0.0341	€21.49$24.93
NA5	NA	16	Ampere Altra (shared)	32 GB	320 GB	20 TB	€0.0568$0.0659	€41.49$48.13
NG-RTX4000Entry	NG	20	Intel Core i5-13500 (14 cores, 20 threads)	64 GB DDR4	2 × 512 GB NVMe	unlimited	€0.2932$0.3401	€214.00$248.24
NG-L4Inference	NG	8	AMD EPYC (dedicated)	48 GB	400 GB	40 TB	€0.7397$0.8581	€540.00$626.40
NG-L40S	NG	16	AMD EPYC (dedicated)	96 GB	800 GB	50 TB	€1.3808$1.6017	€1,008.00$1,169.28
NG-RTX6000	NG	48	Intel Xeon Gold 5412U (24 cores, 48 threads)	256 GB DDR5 ECC	2 × 960 GB NVMe	unlimited	€1.6425$1.9053	€1,199.00$1,390.84
NG-RTX6000-512	NG	48	Intel Xeon Gold 5412U (24 cores, 48 threads)	512 GB DDR5 ECC	2 × 1.92 TB NVMe	unlimited	€2.4644$2.8587	€1,799.00$2,086.84
NG-RTX6000-768	NG	48	Intel Xeon Gold 5412U (24 cores, 48 threads)	768 GB DDR5 ECC	4 × 3.84 TB NVMe	unlimited	€3.1493$3.6532	€2,299.00$2,666.84
NG-H100Training	NG	32	AMD EPYC (dedicated)	256 GB	3.2 TB	80 TB	€2.6575$3.0827	€1,940.00$2,250.40
NG-H100×2	NG	64	2× AMD EPYC (dedicated)	512 GB	6.4 TB	100 TB	€5.3151$6.1655	€3,880.00$4,500.80
NG-H100×4	NG	128	2× AMD EPYC (dedicated)	1 TB	12.8 TB	140 TB	€10.6438$12.3468	€7,770.00$9,013.20
NG-H200×8Cluster	NG	128	2× AMD EPYC (dedicated)	2 TB	30.72 TB	200 TB	€42.0000$48.7200	€30,660.00$35,565.60

Shared or dedicated vCPU — the difference is measurable

A vCPU is a time share of a physical core. How this share is allocated determines your latency under load — not your throughput when idle.

For the shared families NX and NP, the hypervisor schedules multiple instances on the same physical core. Each vCPU gets a quota; if you do not use it, you give your share to your neighbours. This is why these families are cheap: you pay for an expected value, not a guarantee.

If a neighbour computes while you want to compute, your instance waits. This waiting time is visible in the guest system — as %st in top, vmstat or mpstat. Steal time is not a fault, but the honest display of a shared system. On NX and NP, we measure a monthly average of under 3%, and up to 12% during load peaks on individual hosts. On dedicated families, it is zero by design.

We regulate burst behaviour via a credit model instead of hard throttling: a shared vCPU can draw full core performance for up to 20 minutes per hour. After that, the base share applies — 25% per vCPU for NX, 40% for NP. Unused credit is saved for up to eight hours. A build server that compiles for ten minutes and waits for fifty minutes never sees this limit. A video transcoder sees it after twenty minutes.

The rule of thumb from our operations: if the average load of a vCPU is permanently above 20% over a week, a dedicated instance is not only faster, but also cheaper per unit of work done.

Price per vCPU and month

NX / NP — shared: €1.87$2.17 – €11.99$13.91
ND — dedicated (x86): €16.69$19.36 – €19.40$22.50
NA — shared (ARM): €2.59$3.00 – €3.25$3.77

Catalogue prices for Frankfurt without a fixed term. Which series is cheaper per unit of work done is determined by the load, not the list price per vCPU — the rule of thumb is on the left.

Allocation models compared: shared vCPU, dedicated x86 cores and shared ARM cores
Feature	NX / NPShared vCPU	ND / NM / NSDedicated x86 cores	NAShared ARM cores
CPU allocationHow the core is allocated	Quota in the scheduler, core is shared	cpuset, core permanently assigned	Core is shared
Steal time%st in the guest system	Average < 3%, peaks up to 12%	0% by design	visible, no measured value shown
Recommended continuous load	up to 20% per vCPU on weekly average	100%, permanently	no base share shown
Price per vCPU / month	€1.87$2.17 – €11.99$13.91	€16.69$19.36 – €19.40$22.50	€2.59$3.00 – €3.25$3.77
Instruction set	x86-64 (Intel or AMD)	x86-64 (AMD EPYC)	arm64 — custom images required
Typical use	Development, staging, web frontends, cronjobs	Databases, API backends, CI runners, Kubernetes workers	Cloud-native services, container fleets, horizontal scaling
Not suitable for	Continuous load, real-time, licensed core counting	Cost-optimised test environments	Software without arm64 build, Windows

If you want the machine all to yourself

Dedicated servers without a hypervisor — in the same project, in the same vRack and with the same access credentials as your instances. Four series, 17 configurations, from €65.54$76.03 per month.

COREDedicated hardware at an entry-level price4 models · from €65.54$76.03
FORGEThe server platform that grows with you4 models · from €90.75$105.27
TITANCore density6 models · from €299.99$347.99
VAULTStorage density3 models · from €165.54$192.03

The models with processor, memory configuration, drives and connection are listed in full on the dedicated servers page. This just explains when switching is actually worthwhile.

Dedicated vCPU (ND series, largest expansion stage ND8 with 128 vCPU) compared to the dedicated server
Feature	Dedicated vCPU	Metal
What you rent	a share of a server	the server
Hypervisor	present, you see a virtual machine	none, you boot on the metal
Core allocation	fixed vCPU via cpuset	all cores of the processor
Billing	hourly, capped per month	Monthly rent; setup fee only with monthly term
Delivery	on average 35 seconds, guaranteed under 60 seconds	usually 90 minutes for stock items, up to 24 hours for build to order
Hardware access	none	IPMI, RAID control, BIOS
Licences by core count	counted per vCPU, evidence complex	counted per socket and core, unambiguous
Private network	vRack, shared with the instances	vRack, shared with the instances

Scaling groups instead of gut feeling

A scaling group consists of an instance template, a capacity range and rules. The evaluation runs every 30 seconds against metrics we collect anyway.

Metrics

CPU and memory average across the group, active connections and requests per second at the Load Balancer, as well as any custom values via the Prometheus-compatible endpoint.

Response time

Evaluation every 30 seconds, shortest cooldown 60 seconds. Provisioning a new instance from a standard image takes on average 35 seconds, guaranteed under 60 seconds.

Capacity

1 to 200 instances per group, up to 25 groups per project. A warm pool keeps prepared, stopped instances ready for load spikes with short notice.

Scale down

When scaling in, instances are first deregistered, then drained for 120 seconds, then deleted. Instances with scale-in protection remain, such as the leader of a consensus algorithm.

scaling-group.yamlYAML

name:   web-tier
region: fra1
 
template:
  flavor:      np3
  image:       ubuntu-2404
  private_net: vrack-prod
  user_data:   ./cloud-init.yaml
 
capacity:
  min:     2
  max:     12
  desired: 3
 
policies:
  - name:       scale-out-cpu
    metric:     cpu_avg      # Average across all instances in the group
    comparison: ">"
    threshold:  65           # percent
    periods:    3            # 3 × 30 s must be exceeded
    change:     +2
    cooldown:   180
 
  - name:       scale-in-cpu
    metric:     cpu_avg
    comparison: "<"
    threshold:  25
    periods:    6
    change:     -1
    cooldown:   300
 
health_check:
  type:            http
  path:            /healthz
  interval:        15
  unhealthy_after: 3
 
drain_seconds:   120
scale_in_protect: ["web-tier-leader"]

Apply with entronyx compute autoscale apply -f scaling-group.yaml. Changes to the template only affect newly created instances; existing ones are replaced by a rolling replacement.

Functions on demand — in preparation

This service does not exist yet. It is in progress, but we do not give a date because we cannot promise one. What works in its place today is listed below — and it already covers most cases.

In preparation · no date committed

What is possible here today

Peak loads: scaling groups

Rule-based between 1 and 200 instances, evaluation every 30 seconds, warm pool for short warning times. This covers the most common reason people ask for serverless — load that does not occur evenly.

View Auto Scaling

Tasks without continuous operation: Kubernetes

Jobs and CronJobs start a container, process it and terminate it. The control plane of the managed cluster is included in the standard plan at no extra charge; you pay for the nodes underneath.

Managed Kubernetes

Short-lived compute time: instances via the API

Provisioning takes on average 35 seconds, guaranteed under 60 seconds; the runtime is recorded to the second and billed by the hour. For tasks that run for minutes to hours, this is a cheaper calculation than any execution fee.

API and Terraform

What the service should be able to do

5 points we are working on. They are listed here as an intention, not a commitment — and without pre-orders, because a waiting list for a service without a date is just a more polite form of waiting.

Triggers via HTTP, queue and schedule — without your own listener in front
Scaling to zero, so a service without requests costs nothing
Billing by execution time instead of the runtime of a machine
Custom container image instead of predefined runtime environments
The same API, the same CLI and the same Terraform provider as for instances

Tracked by the second, billed by the hour, capped monthly

We log the start and stop of every instance to the second. The invoice still shows an hourly figure — because an item that no one can verify is not transparent.

01
Ingestion
per second
The term is logged to the second. Start, stop, resize and deletion are timestamped in the project's activity log and can be retrieved via the API.
02
Rating
per started hour
Billing is at the hourly rate. The hourly rate is the monthly price divided by 730 — the average month length, not that of the respective calendar month.
03
Cap
Upper limit
If the sum of an instance's hours in the calendar month reaches the monthly price, the counting stops. A 744-hour month never costs more than a 720-hour month.
04
Term discount
optional
A 12-month commitment reduces the monthly price by 8%, a 24-month commitment by 14%. The discount applies to the cap and therefore also to every single billed hour.

Actual billed amount per term — the penultimate column shows a month with 744 hours in which the cap applies
Instance	Hourly rate	24 h	100 h	400 h	744 h	Cap
NX22 vCPU · 4 GB	€0.0082$0.0095	€0.20$0.23	€0.82$0.95	€3.28$3.80	€5.99$6.95	€5.99$6.95
NP34 vCPU · 8 GB	€0.0493$0.0572	€1.18$1.37	€4.93$5.72	€19.72$22.88	€35.99$41.75	€35.99$41.75
ND34 vCPU · 16 GB	€0.1063$0.1233	€2.55$2.96	€10.63$12.33	€42.52$49.32	€77.60$90.02	€77.60$90.02
NM48 vCPU · 64 GB	€0.2648$0.3072	€6.36$7.38	€26.48$30.72	€105.92$122.87	€193.30$224.23	€193.30$224.23
NS34 vCPU · 16 GB	€0.3014$0.3496	€7.23$8.39	€30.14$34.96	€120.56$139.85	€220.00$255.20	€220.00$255.20

Region factor

The catalogue price applies to Frankfurt. Every other location carries a factor that reflects the electricity price, network connection and local taxes. Example NP3 (4 vCPU, 8 GB):

Frankfurt am Mainfra1
€35.99$41.75
Factor 1.00 · 3 ms
Frankfurt am Mainfra2
€35.99$41.75
Factor 1.00 · 3 ms
Berlinber1
€35.99$41.75
Factor 1.00 · 6 ms
Munichmuc1
€35.99$41.75
Factor 1.00 · 5 ms
Helsinkihel1
€33.83$39.24
Factor 0.94 · 21 ms

The same API that our own console uses

No reduced feature set for machines: everything that works in the web interface works via REST, CLI and the Terraform provider. Authentication is done with project-bound tokens that can be restricted to individual actions.

POST /v1/regions/fra1/instancesHTTP

curl -X POST https://api.entronyx.cloud/v1/regions/fra1/instances \
  -H "Authorization: Bearer $ENTRONYX_TOKEN" \
  -H "Idempotency-Key: 8f2c-web-01-2026-08" \
  -H "Content-Type: application/json" \
  -d '{
    "name":        "web-01",
    "flavor":      "np3",
    "image":       "ubuntu-2404",
    "ssh_keys":    ["ops-2026"],
    "private_net": "vrack-prod",
    "backup":      "bk-snap",
    "tags":        { "env": "prod", "tier": "web" },
    "user_data":   "#cloud-config\npackages:\n  - nginx\n"
  }'
 
# 202 Accepted
# {
#   "id":      "i-4c81f0a9",
#   "status":  "BUILD",
#   "flavor":  "np3",
#   "region":  "fra1",
#   "price":   { "hourly_net": 0.0493, "monthly_cap_net": 35.99 },
#   "job":     "/v1/jobs/j-71ad3e"
# }

Response 202 Accepted with the instance ID and a job reference. The state changes via BUILD and BOOT to ACTIVE; it takes on average 35 seconds, guaranteed under 60 seconds until the first SSH connection.

Command lineSHELL

# Filter catalogue — same data as in the table above
entronyx compute flavor list --family dedicated --region fra1
 
# Create instance and wait until ACTIVE state
entronyx compute instance create nd3-prod-01 \
  --flavor nd3 --image ubuntu-2404 --region fra1 \
  --ssh-key ops-2026 --private-net vrack-prod --wait
 
# Move vertically: stop instance, swap flavor, start
entronyx compute instance resize nd3-prod-01 --flavor nd4 --confirm
 
# Running costs of the current billing period
entronyx billing usage --since month-start --group-by flavor --output table

The CLI reads the same tokens from ENTRONYX_TOKEN and can pass any output to scripts with --output json.

main.tfTERRAFORM

terraform {
  required_providers {
    entronyx = {
      source  = "entronyx/entronyx"
      version = "~> 2.4"
    }
  }
}
 
provider "entronyx" {
  region = "fra1" # Token comes from ENTRONYX_TOKEN
}
 
resource "entronyx_vrack" "prod" {
  name = "vrack-prod"
  cidr = "10.42.0.0/16"
}
 
resource "entronyx_instance" "web" {
  count       = 3
  name        = "web-${count.index + 1}"
  flavor      = "np3"
  image       = "ubuntu-2404"
  ssh_keys    = ["ops-2026"]
  private_net = entronyx_vrack.prod.id
  backup      = "bk-snap"
 
  lifecycle {
    create_before_destroy = true
  }
}

Rate limit: 600 requests/min; per token, header X-RateLimit-Remaining
Response time p95: 180 ms; Control plane, region fra1
Idempotency: Idempotency-Key; 24-hour reuse
Provider: entronyx/entronyx; Terraform ≥ 1.7, OpenTofu supported

Not every family is available at every location

Storage and GPU families require special chassis and are therefore only located where the corresponding racks are installed. The table shows how many of the 39 instance types can be ordered at each location.

Cloud locations with number of bookable instance types, price factor compared to Frankfurt, connection latency, PUE and current occupancy
Location	Description	Types	Factor	Latency	PUE	Occupancy
Frankfurt am Mainfra1	Location Rhein-Main I	39 / 39	1.00	3 ms	1.14	71%
Frankfurt am Mainfra2	Location Rhein-Main II	39 / 39	1.00	3 ms	1.09	44%
Berlinber1	Location Spree	22 / 39	1.00	6 ms	1.16	58%
Munichmuc1	Location Isar	25 / 39	1.00	5 ms	1.11	82%limited
Helsinkihel1	Location Uusimaa	39 / 39	0.94	21 ms	1.08	36%

Occupancy above 80% does not mean that no more instances will start — it means that large types are temporarily unavailable there. For planned capacity from 50 instances or for GPU nodes, we recommend a reservation via sales.

Six points that can be verified

Every point carries a value from the catalogue or from a section of this page. An advantage without a number would be just as true on any competitor's page and is therefore not listed here.

One tariff for thirty-nine sizes
The term is recorded to the second, billed at the hourly rate and capped at the monthly price: a month with 744 hours never costs more than one with 720.
Entry per month€5.99$6.95 net
Outbound traffic without zone tariff
At least 20 TB of outbound traffic is included per instance. Above that, a single rate per terabyte applies, regardless of the destination.
Excess usage€1.09$1.26 per TB
Deployment in seconds
From the call to the ACTIVE state, it takes on average 35 seconds, guaranteed under 60 seconds for standard images. The same order of magnitude applies to instances that a scaling group pulls in.
Until ACTIVE stateunder 60 s
Everything via one interface
No reduced functionality for machines: what works in the web interface works via REST, command line and the Terraform provider with the same calls. A p95 response time of the control plane — the time that 95 out of 100 calls do not exceed — is only shown here with the measurement period, endpoint and location. Without these three details, such a number is arbitrary.
InterfaceREST · CLI · Terraform
Guaranteed availability
Cloud instances are in class B of the availability commitment. It is measured at the network connection and the control plane, not by the state of your application.
Class B per month99.95%
Processing without third-country transfer
The cloud runs exclusively in Germany and Finland. No location is outside the EU; a third-country transfer within the meaning of the GDPR therefore does not take place. The data processing agreement according to Art. 28 is part of the main contract and not an addendum.
Locations outside the EUnone

Three recurring situations

No company names, no quotes, no logos. Instead, each case names the starting situation, the chosen approach and a metric that you can look up in the tables on this page.

E-commerce
Peak load on campaign day without permanent provisioning
Initial situation
Delivery ran on a fixed number of permanently running instances, designed for the busiest day of the year. On all other days, a large part of them was idle and was paid for anyway.
Decision
A scaling group of NP3 instances with evaluation every 30 seconds and 120 seconds of drainage during teardown — open connections are served to the end before an instance is deleted. The base load remains, the peak is paid by the hour and ends when the group is scaled down.
Result
€0.0493$0.0572
Hourly rate per node of the group (NP3, Frankfurt, net)
How it's built — Peak load on campaign day without permanent provisioning
Research and evaluation
A dataset that must fit entirely into memory
Initial situation
The evaluation kept the entire dataset in memory. On shared cores, the runtime fluctuated depending on the neighbourhood on the same host: waiting time for the core arrived as steal time in the guest system — the time during which the core currently belongs to a neighbour.
Decision
The memory-optimised series with dedicated cores and 8 GB memory per vCPU. Steal time is zero there by design; the large models are only available at locations with the corresponding rack density.
Result
256 GB
Memory per instance (NM6)
How it's built — A dataset that must fit entirely into memory
Media delivery
Outbound traffic as the largest item on the invoice
Initial situation
The previous provider charged outgoing traffic per target region with its own rate. The invoice could not be estimated in advance because it depended on the origin of the requests.
Decision
Delivery nodes of the storage-optimised series with local NVMe. The included volume grows with the model; above this, a single rate per terabyte applies, regardless of the destination.
Result
40 TB
Outbound traffic included (NS3, per month)
How it's built — Outbound traffic as the largest item on the invoice

What needs to be clarified before the first invoice

Questions from sales conversations and from the first month of operations — not the ones that are easy to answer.

Cloud instance or VPS — what is the difference?

Both run on the same platform and are billed the same way: per hour, capped at the monthly price. The difference lies in the allocation of the core.

As a VPS you book the series with shared cores, NX, NP and NA, from €5.99 net per month. The series with dedicated cores, ND, NM and NS, start at €38.80.

The rule of thumb from operations: if the average load of a vCPU over a week is permanently above 20%, the dedicated series is cheaper per unit of work performed.

How fast is an instance ready?

From the call to the first SSH connection, standard images take on average 35 seconds, guaranteed under 60 seconds. The state changes via BUILD and BOOT to ACTIVE, with each transition timestamped in the project's activity log.

Instances pulled by a scaling group are ready in the same timeframe.

What does an instance cost if it only runs for a few hours?

Billing is per started hour at the hourly rate. The hourly rate is the monthly price divided by 730 — the average month length, not that of the respective calendar month.

The count is capped at the top: if the sum of an instance's hours in a calendar month reaches the monthly price, it stops. A month with 744 hours therefore never costs more than one with 720.

Do I continue to pay for a stopped instance?

Yes. A shut-down instance continues to occupy CPU quota, memory and local NVMe on the host and is therefore still billed.

Only deleting ends the billing. If you want to keep the state, create a snapshot beforehand, an image of the instance at that point in time. Snapshots appear on the invoice as a separate item.

What counts as egress, and what does it cost above the included volume?

Outbound traffic to the public internet counts as Egress. Inbound traffic, traffic between instances via the private network — even between locations — and access to Object Storage from instances in the same region are not counted.

Depending on the instance type, 20 TB to 200 TB of outbound traffic per month are included. Above this, we charge €1.09 per terabyte, without a zone tariff: a terabyte to a recipient on the other side of the world costs the same as one within Germany. The exception: NG-RTX4000, NG-RTX6000, NG-RTX6000-512 and NG-RTX6000-768 do not count outbound traffic at all — traffic is unlimited there.

How do I cancel an instance?

Without a term commitment, you delete the instance in the console, via the API or the command line. There is no notice period; billing ends at the time of deletion (Section 11 (3) of the Terms and Conditions). Stopping is not enough for this.

Contracts with a 12 or 24-month term can be cancelled with a notice period of 30 calendar days to the end of the term, otherwise they are extended by twelve months (Section 11 (2) of the Terms and Conditions). Cancellations require text form; the cancellation function in the customer area fulfils this.

What is the difference between shared and dedicated vCPUs?

In the shared families, the hypervisor, the virtualisation layer of the host, schedules multiple instances on the same physical core. If you do not use your quota, you pass the share on to the neighbours — in return, the price per vCPU is lower.

Waiting time caused by neighbours is visible in the guest system as steal time — the time during which the core currently belongs to a neighbour. The monthly average on shared families is below 3%, and up to 12% during peak loads on individual hosts. On dedicated families, it is zero by design.

Is data retained on the local NVMe?

The specified capacity is physically located in the same server as the instance and is tied to its lifespan. It survives reboots, but not the deletion of the instance.

For data that must outlast an instance, Block Storage or Object Storage belong underneath. Both are separated from the compute node and can be attached to a new instance.

Can I resize an instance later?

Yes. Resizing is a logged action like start, stop and deletion and is recorded with a timestamp in the project's activity log. From this timestamp, the hourly rate of the new size applies.

A change of family is not covered by this, because the local NVMe is tied to the instance. The intended way for this is via a snapshot and a new instance.

Is every instance type available at every location?

No. 39 instance types are distributed across 5 locations in Germany and Finland. Storage and GPU designs require special chassis and power density and are therefore only located where the corresponding racks are installed.

The catalogue price applies to Frankfurt. Every other location carries a factor that reflects the electricity price, network connection and local taxes; the table in the "Billing" section calculates it using an example.

Do I have to commit to get the entry price?

No. €5.99 net is the monthly price without any commitment, and hourly billing applies unchanged.

A voluntary commitment lowers the price: 8% for 12 months, 14% for 24 months. The discount applies to the monthly cap and thus also to every single billed hour.

Can everything be controlled without the web interface?

Yes. There is no reduced feature set for machines: what works in the console works via REST, via the command line and via the Terraform provider.

Authentication is done with project-bound tokens that can be restricted to individual actions. The rate limit is 600 requests per minute and token. Write requests accept an Idempotency-Key — a key with which a repeatedly sent request only takes effect once.

from · per month, net€5.99$6.95

Configure instance

Configure instance or query catalogue via API

The configurator guides you in three steps through instance, location including image and block storage, as well as the extras — and updates the price after every decision. If you prefer automation, you can use the Terraform provider directly.

Configure instance API documentation

Delivery: under 60 s
Traffic included: 20 TB
Excess usage: €1.09$1.26 / TB
SLA: 99.95%

Eleven product groups,one billing model

Your path to the Public Cloud

Deploy

Scale

Stay within EU jurisdiction

The entire section, sorted by tasks

Compute

Containers and orchestration

Storage

Network

Databases

Analytics

AI and machine learning

Quantum Computing

Security and identity

Operations

Tools for developers

Price up your instance before you order it

Choose the load profile first, then the size

Shared vCPU

Dedicated vCPU

Memory-optimised

Storage-optimised

ARM / Ampere

GPU

Hardware per series

NXShared vCPU

NPShared vCPU

NDDedicated vCPU

NMMemory-optimised

NSStorage-optimised

NAARM / Ampere

NGGPU

Images

All instance types compared

Shared or dedicated vCPU — the difference is measurable

If you want the machine all to yourself

Scaling groups instead of gut feeling

Metrics

Response time

Capacity

Scale down

Functions on demand — in preparation

What is possible here today

Peak loads: scaling groups

Tasks without continuous operation: Kubernetes

Short-lived compute time: instances via the API

What the service should be able to do

Tracked by the second, billed by the hour, capped monthly

Ingestion

Rating

Cap

Term discount

Region factor

The same API that our own console uses

Not every family is available at every location

Six points that can be verified

One tariff for thirty-nine sizes

Outbound traffic without zone tariff

Deployment in seconds

Everything via one interface

Guaranteed availability

Processing without third-country transfer

Three recurring situations

Peak load on campaign day without permanent provisioning

A dataset that must fit entirely into memory

Outbound traffic as the largest item on the invoice

What needs to be clarified before the first invoice

Configure instance or query catalogue via API