DeepSeek V4 ships under an MIT licence, which makes self-hosting a real option for Australian teams that want direct control over their model and their data. Running it well takes more than a download. A production deployment behaves like a small platform, with compute, security, and on-call duties attached, and it pays to plan it that way from the start.
What you are actually taking on
Before any hardware is ordered, it helps to be honest about the shape of the commitment. Self-hosting moves a set of responsibilities from a provider onto your own team, and those responsibilities do not pause on weekends or during a busy reporting period.
You own uptime, patching, and incident response for the model server
You own the security of an extra system that holds sensitive data
You own capacity planning, including the cost of idle hardware
You own testing every model and dependency upgrade before it ships
None of this is a reason to avoid self-hosting. It is a reason to size the decision properly, because the work is ongoing rather than a single project with an end date.
The core components
A production deployment needs several moving parts working together reliably. Each is straightforward on its own, and the real effort is keeping them healthy as a single system.
GPU compute sized to the active parameter count, not the total parameter count
An inference server such as vLLM or a comparable engine
Monitoring, structured logging, and autoscaling that matches demand
A staging environment for testing model and code changes safely
DeepSeek V4 uses a mixture-of-experts design, so only a fraction of its parameters are active on any given request. Sizing to the active count rather than the headline number is the single biggest lever on your hardware bill.
Sizing the hardware
The temptation is to buy for the largest workload you can imagine. A better approach is to measure real demand first, then size to it with a little headroom. A Sydney team running business-hours traffic has a very different profile from one serving customers around the clock.
Profile expected requests per hour across a normal week
Match GPU memory to the active experts plus the context window you need
Plan for peak load, but do not pay for peak capacity around the clock
Keep a clear upgrade path for when demand grows
Most Australian SMBs find their traffic is uneven and concentrated in working hours, which means a single well-chosen node often does the job at launch and leaves room to grow later.
Australian-specific requirements
A global setup guide will skip the local context that matters most here. Data residency and privacy obligations shape the architecture, not just the paperwork that sits around it.
Compute hosted in an Australian region for genuine data residency
Access controls aligned with the Privacy Act and your own policy
A backup and recovery plan you can actually audit
Clear records of where logs, caches, and backups physically live
These are not optional extras for a business handling customer or staff data. Building them in from day one is far cheaper than retrofitting them after an incident or a client request lands on your desk.
What it costs to run
The model weights are free. The platform around them, and the person who keeps it running, are not. A capable single-node setup in Australia starts near $40,000 a year before staffing, with GPU compute alone often running $3 to $12 per hour for each instance you keep live.
Compute, storage, and networking as recurring monthly costs
At least part of an engineer's salary for maintenance and on-call cover
Security and compliance work treated as ongoing, not a one-off
Testing time for every model and library upgrade
Add a full engineer and the real figure climbs past $120,000 a year. For a business whose usage would never keep that node busy, a managed Claude build often meets the same goal for less, because the provider carries the infrastructure and the on-call load. Where the volume genuinely justifies self-hosting, the maths works the other way, and we will say so plainly rather than sell you a server you do not need.
A go-live checklist
Before a self-hosted DeepSeek V4 node serves real traffic, confirm the essentials are in place. This short list separates a weekend experiment from a system the business can depend on during working hours.
Residency, access control, and audit logging confirmed and tested
Monitoring and alerting that catch failures before users do
A tested rollback for the next model update
A named owner for the system, with cover for leave
We can stand this up for Australian teams, size it to real demand, and cost it honestly against a managed alternative. If you want a clear read on which path fits your volume, book a brainstorm.
