Skip to content
State of GPU Compute 2026

Enterprise

Production GPUs, under one contract.

Reserve 8-GPU servers in the regions you choose, burst on-demand above them, and get one post-paid invoice in euros. Start on a card; sign when the usage curve settles.
Talk to salesSee pricing
console.nexpod.ai/org/agreementActive
Enterprise agreement
your-org · EU-AMS-1 · EU-FRA-1
termmonth 4 of 12
Reserved
24servers
Burst · Oct
455server·h
Uptime · Oct
99.98%
Servers in use
reserved 2433 peak
burst +9
B300 committed
€5.00/GPU·hr
B300 burst
€6.00/GPU·hr
Commit · month
€607,360
post-paid · invoiced 1 Nov · net 30SEnamed engineer

Teams running production on Nexpod

Placeholder
  • Customer
  • CUSTOMER
  • customer
  • Customer Co
  • customer.ai
  • CUSTOMER

Why teams sign

Where GPU fleets go wrong.

Five problems that grow with the fleet, and the part of a Nexpod agreement that handles each one.

  1. 01

    The plan said 24

    October peaked at 33 servers and dipped to 18. No single server count fits a month like that.

    Hold 24 at the reserved rate. The other 9 run on the per-second meter.

    plan 24peak 33low 18
  2. 02

    The B300s sold out

    The family you benchmarked on is gone in the week you need twelve more of it.

    Reserved servers are named machines, held for your org in EU-AMS-1 or EU-FRA-1.

    • B300 · on-demand pool0 free
    • ams-r-01…16held
  3. 03

    Your ML team got a pager

    Firmware updates, NVLink errors and dead drives land on engineers you hired to train models.

    We swap the part and re-image the server. You keep root over SSH.

    • 02:14ams-r-02 · nvme1 fault
    • 02:31drive swapped by Nexpod
    • 02:52re-imaged · in service
  4. 04

    Idle hours on the invoice

    A fleet bought for launch week sits at 40% busy for the rest of the quarter and bills at 100%.

    Only the floor is fixed. Burst servers bill to the second and stop when you stop them.

    fixed fleet 40%bills 100%
    floor + burst 95%bills what ran
  5. 05

    Five cards, no total

    Each team pays for its own GPUs, and finance rebuilds the monthly figure from receipts.

    One org, a cost centre per team, one invoice in euros.

    card ··4412card ··0937team-lab acct
    1 invoiceEUR

From card to contract

From one H100 to a reserved fleet.

Nobody signs on day one. This is how a typical account grows, counted in servers and euros.
  1. 01
    Day 1 · card

    One H100 server

    Deployed from the console at €24.00 per hour, stopped when the benchmark finished.

  2. 02
    Weeks 2 to 10 · on-demand

    A fleet of 18 to 33 servers

    Usage settles around 24 servers with launch peaks above it. Every hour bills per second.

  3. 03
    Month 3 · agreement

    24 reserved, the rest on the meter

    16 B300 and 8 H200 move to €40.00 and €24.00 per server-hour. Peaks keep billing per second.

Products

Three products, one agreement.

Reserved servers, burst and Serverless land on the same rate card and the same invoice. Everything you do in the console has a nexctl command and an API call.
n0
n1
n2
n3
node fabric32 GPUs · one job
Instant Clusters

Multi-node training

8-GPU servers joined by a fast fabric, on-demand or reserved.

2 to 8 servers self-serve · larger fabrics on contract
Workers · 24 h
0 → 12 → 1
Serverless

Autoscaling inference

Endpoints that add workers under load and drop to zero after.

Billed per GPU-second
NODENVIDIA DGX H200
8/8 online
0
H200
1
H200
2
H200
3
H200
4
H200
5
H200
6
H200
7
H200
Cloud Servers

Whole 8-GPU servers

For development, training runs and long-running services.

By the hour or reserved for 12 months
nexctl · zsh
$ git tag v42 && git push --tags
$ nexctl endpoints release chat-api --tag v42
→ building image from tag v42
✓ v42 live on 3 workers
v41 kept · nexctl endpoints rollback
api.nexpod.ai/v1
POST /v1/endpoints/chat-api/releases
Authorization: Bearer $NEXPOD_KEY
{ "tag": "v42" }
201 { "version": "v42",
"status": "live" }
  • v42live3 workers · 14:02
  • v41standbyrollback target
  • v40archivedimage kept 30 days
nexctl + API

Release from a git tag

Tag a commit, release it with nexctl or one POST, and roll back the same way.

Any container image · volumes persist between releases

Organisation

Access and spend that follow your org chart.

Once more than a handful of people deploy, structure matters. Every resource belongs to the organisation, sits in a workspace and carries tags.
your-org
  • research14 resources
    pretrain-b300evals
  • inference22 resources
    chat-apiembeddings
  • platform6 resources
    ci-runners

Resources belong to the org

Workspaces per team. Nothing is tied to whoever clicked deploy.

login.nexpod.ai/your-org
Sign in to your-org
Continue with SSO
name@your-company.com
  • SAML 2.0connected
  • OIDCavailable
  • SCIM sync214 users
  • Enforce SSOon

Single sign-on, enforced

SAML or OIDC from your identity provider, with SCIM to add and remove people.

RoleDeployBillingMembersAudit
Owner
Admin
Billing
Developer
Viewer

Roles per group

Owner, Admin, Billing, Developer and Viewer built in. Custom roles on request.

team:researchenv:prod▍
cost:ml-410project:evalsowner:platformregion:eu-ams-1
14 of 1,206 resources match
  • pretrain-b300-01
  • pretrain-b300-02
  • evals-h200

Tags you can filter on

Label by team, project or cost centre and find any server in seconds.

14:02:11m.devriesservers.create 8× B300 · eu-ams-1
13:58:40sso-botmember.added group: inference
13:41:07j.kowalskirole.changed developer → admin
13:20:55ci-deployendpoint.release chat-api v42
12:59:02a.moreaubilling.export invoice 2026-10

An audit log for the whole org

Who created, changed or removed what. Export it to your SIEM.

ml-410 · research€233,600
ml-220 · inference€140,160
pl-100 · platform€19,680
untagged€2,304

Spend per cost centre

Tagged usage lands on the invoice grouped by cost centre, in euros.

Commercial model

24 servers held. The rest metered.

October on the example agreement: 16 B300 and 8 H200 at reserved rates all month, plus 455 server-hours above 24 on the on-demand meter. It arrives as one invoice on 1 November.

Talk to salesReserved rates
Servers in use · October baseline 24 in use idle burst
1 Oct15 Oct31 Oct
  • Reserved · B300€467,20016 servers × 730 h€40.00/hr
  • Reserved · H200€140,1608 servers × 730 h€24.00/hr
  • Burst · B300 on-demand€21,840455 server·h, per second€48.00/hr
invoiced 1 Nov · SEPA · net 30 · terms past 12 months quoted per dealOctober · excl. VAT€629,200

Security and compliance

Answers for your security review.

Single-tenant hardware, EU data residency and the paperwork reviewers ask for, ready before the first call.
your server · single tenant
G0G1G2G3G4G5G6G7
2× CPURAMNVMe
tenant Bother hardware
tenant Cother hardware

Single-tenant by design

Your workload never shares a server, a GPU or a local drive with another customer.

  • Data Processing AgreementGDPR, EU standard clauses NDA
  • Sub-processor listUpdated with 30 days notice NDA
  • Security whitepaperArchitecture, access, incident response NDA
  • Penetration test summaryMost recent external test NDA
  • ISO/IEC 27001 certificateCertification statusPlaceholder
  • SOC 2 Type II reportAudit statusPlaceholder

The documents, ready

Shared under NDA. Certification items stay placeholders until audits complete.

Request the security pack
  • At restAES-256
  • In transitTLS 1.3
  • Keysper volume

Encrypted at every layer

Volumes, snapshots and API traffic.

  • SSOenforced
  • MFArequired
  • IP liston

Locked-down access

Only your people get in.

Dedicated capacity

Hardware held for you, where you need it.

Reserved servers are allocated to your organisation in the region you pick. When demand jumps, the machines are already racked, imaged and yours.
EU-AMS-1Amsterdam
EU-FRA-1Frankfurt
EU-HEL-1Helsinkisoon
Live region Planned

Choose where data lives

EU-AMS-1 and EU-FRA-1 for EU residency, US-NYC-1 for US traffic. EU-HEL-1 is planned.

forecastreserved

Capacity planned with you

Quarterly reviews turn your forecast into reserved servers ahead of time.

  • ams-r-01NVIDIA DGX B300
  • ams-r-02Supermicro SYS-822GS-NB3RT HGX
  • ams-r-03GIGABYTE G894-ZD3-AAX7
  • fra-r-07NVIDIA DGX H200
  • fra-r-08ASUS ESC NB8-E11
  • fra-r-09NVIDIA DGX B200

Named machines, not a pool

Your reserved servers have IDs, regions and a spec sheet.

Customers

What changed after the move.

Teams on reserved Nexpod servers, in their own words: the workload, the fleet they reserved and the numbers that moved.
Placeholder

Placeholder quote. A customer describes, in one or two plain sentences, what changed after moving a real workload to Nexpod.

Customer ARole, Company
Customer namePlaceholder
metric · %
Cost change after reserving
Case studies
Customer namePlaceholder
metric · ×
Throughput or latency change
Case studies
Customer namePlaceholder
metric · count
Fleet size, such as reserved servers
Case studies

Support

A team that already knows your setup.

Support scales with the account. Under an agreement you get a named solutions engineer and response times written into the contract.
Standard

Every account, from the first server

  • Docs, guides and email support
  • Reply within one business day
  • Status page and incident updates
Included
Business

Teams running production on-demand

  • Priority queue, 4 h response in business hours
  • Shared chat channel with our engineers
  • Quarterly usage and cost review
Included from 4 reserved servers
EnterpriseWith an agreement

Fleets under an agreement

  • 1 h response for P1, around the clock
  • Named solutions engineer who knows your stack
  • Migration plan and hands-on onboarding
  • 99.9% uptime SLA with service credits
Part of every enterprise agreement

FAQ

Enterprise questions.

The things procurement, security and engineering usually ask first.
Which servers can we reserve?
Any of the 20 models on the server list, as whole 8-GPU machines: B300, B200, H200 or H100. You reserve by family and region; we assign named servers such as ams-r-01 and keep them out of the on-demand pool for the full term.
How is burst billed?
Every server above your reserved count bills at the on-demand rate of its family, metered per second and listed on the same monthly invoice. For a B300 that is €48.00 per server-hour on-demand, against €40.00 reserved.
Can reserved servers stay in EU-AMS-1 only?
Yes. Each reservation is pinned to one region: EU-AMS-1 or EU-FRA-1 for EU residency, US-NYC-1 for US traffic. EU-HEL-1 is planned. Volumes and snapshots stay in that region unless you copy them out.
What does the reserved rate save?
One euro per GPU-hour on every family, from €2.00 for H100 to €5.00 for B300. That is 33% off on H100 and 17% on B300. Terms past 12 months and fleets above 64 servers are priced per deal.
Do we share a machine with another customer?
Never. Each Cloud Server and cluster node is one physical machine for one organisation: all eight GPUs, both CPUs, the RAM and the local NVMe.
How large can a reserved cluster get?
Self-serve Instant Clusters run from 2 to 8 servers, so up to 64 GPUs. Larger fabrics are reserved under the agreement and planned with your solutions engineer before the term starts.
What does the security team receive?
Under NDA: the Data Processing Agreement, the sub-processor list, the security whitepaper and the latest external penetration test summary. Certification reports are marked as placeholders on this page until the audits are complete.
What happens at month 12?
In month 10 your engineer reviews usage with you. You renew at the same count, change the mix of families, or let the term end and keep running on-demand at €3.00 to €6.00 per GPU-hour.

Talk to sales

Describe the workload. Get a quote.

A solutions engineer replies within one business day with a server plan, regions and a reserved rate.

  • Salessales@nexpod.ai
  • Security questionnairessecurity@nexpod.ai
  • Reply time1 business day
Tell us what you are planning
Demo form

We use these details only to answer you. See the privacy policy.