What the AI Server actually runs
The NZ$2,900 AI Server is not a general-purpose PC. DataForge pre-images it on Ubuntu 24.04 LTS with a purpose-built AI stack before delivery. Every service runs as a background process, starts automatically on boot, and exposes an API on your local network. No configuration is needed after installation.
The hardware is built around an AMD Ryzen 5 7600 paired with an NVIDIA RTX 5060 Ti. The 5060 Ti is a Blackwell-generation card, NVIDIA's newest GPU architecture, delivering 16 GB of VRAM and CUDA 13.0 support. Combined with 62 GB of system RAM (31 GB reserved for shared memory so models can spill beyond VRAM without thrashing), the server runs a 20-billion-parameter model at approximately 31 tokens per second. That is fast enough that the response appears to stream in real time.
Ollama and local language models
Ollama is the inference engine that runs open-weight language models across the GPU and shared system RAM. DataForge pre-loads a curated set of models suited to business use: a 20B general-purpose model for everyday tasks, a code-focused model for technical work, and a compact 7B model for fast, lightweight queries. The 62 GB RAM configuration means large models are loaded into shared memory and served without the cold-start delays typical of VRAM-constrained setups.
The server exposes an OpenAI-compatible API endpoint at the local network address. Any software that can talk to the OpenAI API (including the ForgeBox, n8n, Continue IDE extension, and custom applications) connects to the local server instead of calling the cloud. Rate limits, per-token billing, and data egress are all eliminated.
Whisper ASR for speech-to-text
The server runs a dedicated Whisper ASR service using approximately 1.9 GB of VRAM, leaving the remainder available for LLM inference. The result is accurate, multi-language transcription that processes meeting recordings, voice notes, customer call recordings, and audio files entirely on your hardware. A transcription API sits on the local network, accessible to any ForgeBox or internal application.
Use cases that become straightforward with a local Whisper server: automatic meeting minutes, searchable call recordings, voice-to-CRM note creation, and accessibility features for internal tools.
AI image generation (SDXL) and OCR
With 16 GB of VRAM on the RTX 5060 Ti, the server runs Stable Diffusion XL concurrently with other services. SDXL holds a small resident footprint in VRAM (around 130 MB at idle) and scales up when a generation request arrives. The image generation API accepts text prompts and returns images without any cloud service involved. An OCR API also runs as a resident service (approximately 284 MB VRAM), making document digitisation and structured data extraction from scanned files available across the network.
AI video generation
Lighter video generation models run on the server for short-clip generation from image or text prompts. The 16 GB VRAM headroom of the RTX 5060 Ti handles these loads more comfortably than the 12 GB cards common in previous-generation setups. This covers a genuine business need: short product clips, animated explainers, and social content that would otherwise require external services or creative agencies.
Want to see the AI Server demonstrated before you commit?
We run a live remote demo showing Ollama, Whisper, and image generation working together on a reference server. Book a session and see what is possible.
Book a Demo
What DataForge delivers
The value DataForge provides is not the hardware. It is the work between "server delivered" and "AI integrated with your business." That gap is where most self-managed implementations stall.
1
Discovery and use-case mapping
We map your business processes, identify the five to ten highest-value AI integration points, and decide which ForgeBox mode fits each one. This typically takes two to three days and produces a prioritised integration plan with a clear first-90-days roadmap.
2
Hardware procurement, pre-imaging, and delivery
We procure, configure, and image both the AI Server and ForgeBox hardware before delivery. The AI Server arrives with all services running and tested. Each ForgeBox arrives with Claude Code, n8n, and the RAG indexer pre-installed and authenticated against your Anthropic account. Plug in, connect to the network, and everything starts.
3
Documentation ingestion and RAG library build
We ingest your existing business documentation into the local vector database: SOPs, product catalogs, pricing, policy documents, historical project files. We clean, chunk, and index the content so the knowledge agent answers accurately from your actual materials, not from hallucination.
4
Systems integration
We wire the ForgeBoxes to your line-of-business systems: ERP, CRM, accounting, ticketing, or whatever your stack includes. DataForge has existing integration experience with SAP, ServiceNow, Salesforce, Shopify, and MySQL-backed custom applications. The agent reads from and writes to your systems via authenticated API connections.
5
Governance, approval workflows, and AI policy
We configure human approval gates for sensitive agent actions, establish spending limits on hybrid cloud API calls, and document the AI usage policy for your organisation. Staff get a clear, written guide on what the agents can do, what they cannot do, and how to report issues.
6
Staff enablement and 30/60/90 review
We run role-specific training sessions for the teams using each ForgeBox, then check in at 30, 60, and 90 days to measure outcomes, address friction points, and expand the integration scope where value has been proven.
On-premise versus cloud AI: the honest comparison
Cloud AI services are excellent for many purposes. This is not an argument against them. It is an argument for understanding when on-premise hardware serves your business better, and making that decision deliberately rather than by default.
| Factor |
DataForge AI Server |
Cloud AI APIs |
| Upfront cost |
NZ$2,900 (one-time) |
$0 upfront |
| Ongoing cost |
~$40/year electricity |
$100–$500+/month at business volume |
| Data sovereignty |
100% on-premise |
Data transmitted to and processed externally |
| Rate limits |
None |
Tier-based, can throttle under load |
| Offline operation |
Full capability |
None |
| Model quality ceiling |
Strong 7B–34B models (not frontier) |
Frontier models (GPT-4o, Claude) |
| Break-even vs $150/mo cloud |
Under 20 months |
Ongoing indefinitely |
| Compliance / data residency |
Full control |
Depends on provider region and T&Cs |
| Multimodal: speech, image, video |
All included, unlimited use |
Additional cost per service |
How the ForgeBox and AI Server work together
The ForgeBox runs Claude Code as its primary agent. Claude Code uses Anthropic's Claude models via API for all reasoning, writing, and decision-making tasks. This requires a Claude Code subscription, and heavier agentic sessions (long multi-step tasks, large codebase analysis, extended autonomous workflows) will consume additional Anthropic API credits beyond the subscription allowance. DataForge helps you understand and manage that usage.
The AI Server's role is to provide local compute for tasks where a cloud API is unnecessary or undesirable. The ForgeBox calls the AI Server for bulk document processing, speech transcription via Whisper, image generation, and high-volume repetitive inference where sending data to Anthropic is either too costly or not appropriate. Sensitive internal data stays on the local network; the Anthropic API receives only the reasoning tasks that need frontier-model quality.
The result is a layered cost structure: one Claude Code subscription, occasional API credit top-ups for heavier sessions, and one-time hardware. Staff experience a single agent that is always available and always understands their business.
Curious what your specific workload would cost on local versus cloud?
Share your current AI usage with us and we will model the break-even point and projected savings for your business specifically.
Run the Numbers
Is this right for your business?
This infrastructure makes the most sense for businesses that share some combination of the following characteristics:
- You handle commercially sensitive data that you are not comfortable processing through third-party cloud services
- You have a team generating consistent AI usage that would exceed NZ$150 per month in cloud API costs within 12 months
- You want AI that understands your specific products, processes, and terminology rather than a generic model
- You have existing business systems (ERP, CRM, job management) that you want AI integrated with, not alongside
- You need speech-to-text, image generation, or video generation capabilities and want to avoid per-use cloud charges
- You want to solve the shadow AI problem: give staff a sanctioned, governed AI tool so they stop using unapproved cloud services with company data
It is not the right fit for businesses with no IT presence, a single person operation with minimal AI volume, or organisations that specifically need frontier-model quality for all tasks. In those cases, a well-configured cloud API approach is the better starting point, and DataForge can help with that too.
50%+
of NZ organisations are experiencing shadow AI use, where staff use unapproved cloud tools with company data. On-premise infrastructure eliminates the root cause.
Frequently asked questions
Does the NZ$2,900 include DataForge's setup and integration work?
The NZ$2,900 covers the hardware, pre-imaging, and delivery of the AI Server. DataForge's integration and setup work is scoped and quoted separately based on the complexity of your business systems and the number of ForgeBoxes. A typical starter engagement (one AI Server, one ForgeBox, documentation ingestion, one system integration) is quoted as a fixed-price project.
What happens when the hardware gets old?
The AI Server is a standard mini PC. It can be upgraded (RAM, storage, eventually GPU) or replaced in three to four years as hardware costs continue to fall. The software stack (Ollama, Whisper, the agent layer) is open-source and hardware-agnostic. You are not locked into DataForge hardware for future upgrades.
Can the AI Server handle multiple ForgeBoxes at once?
Yes. The 62 GB RAM configuration with 16 GB VRAM means the server holds a 20B model in shared memory and serves concurrent requests without reloading. In practice, three to five simultaneous ForgeBox or user requests run without noticeable degradation. For teams with heavier concurrent demand, the architecture supports adding a second AI Server and load-balancing across both, which DataForge can configure as part of an expanded engagement.
Is internet connectivity required?
The AI Server's local capabilities (Ollama, Whisper, image generation) run entirely offline. The ForgeBox requires internet access because Claude Code uses the Anthropic API for its reasoning. Web browsing and live data lookup also require internet. Everything that stays on-premise (local inference, document search, file operations) works without it.
What does the Claude Code subscription cost, and how much are extra API credits?
Claude Code is available via Anthropic's Claude Max subscription plans. For most ForgeBox deployments running standard business automation and knowledge tasks, the included subscription usage is sufficient day-to-day. Extended agentic sessions, large codebase analysis, or high-volume autonomous workflows consume additional API credits beyond the plan allowance. DataForge monitors usage during the 90-day review cycle and advises on credit allocation based on your actual workload patterns. In practice, the combined subscription and top-up cost for a typical SME ForgeBox deployment is well below equivalent cloud API spend for the same output.
What support does DataForge provide after installation?
DataForge includes a 90-day check-in and tuning cycle with every installation. After that, ongoing support is available as a monthly retainer or on a time-and-materials basis. Model updates, new skill deployment, and additional system integrations are common ongoing requests.
