Why local AI
The math is simple.
Cloud AI looks cheap until you add up users, tokens, and compliance overhead.
Regulated industries
Built for sectors where data cannot leave.
Cloud AI is not an option when you operate under strict data sovereignty rules. Local inference is the only compliant path.
🏥
Healthcare & Pharma
GDPR · HDS · HIPAA
Clinical trial analysis, medical record summarisation, drug interaction research — on isolated infrastructure with no external data transfer.
⚖️
Legal & Compliance
Attorney-client privilege · RGPD
Contract review, due diligence, case research — confidential documents processed entirely within your firm's walls.
🏦
Finance & Asset Management
MiFID II · SOC 2 · DORA
Private dataset analysis, regulatory reporting, trading strategy research — zero data sovereignty risk.
🏭
Industry & Defence
ISO 27001 · Air-gapped
Process optimisation, maintenance prediction, sensitive R&D — full air-gap possible, no cloud dependency.
Performance
Numbers, not marketing.
NVIDIA GB10 Grace Blackwell Superchip. 128GB unified memory. 1 PETAFLOP of AI compute.
LLaMA 3 70BRecommended
128k context
~60 tokens/sec
inference speed
Mistral 7BFast
32k context
~240 tokens/sec
inference speed
LLaMA 3 405BMax quality
128k context
~14 tokens/sec
inference speed
DeepSeek-R1 70BReasoning
128k context
~55 tokens/sec
inference speed
* Indicative figures based on Grace Blackwell architecture. Exact throughput varies by quantisation and context length.
Why 128 GB unified matters
Traditional AI servers split CPU and GPU memory. The GB10 Superchip shares 128 GB between both — meaning a 70B model fits entirely in fast memory without swapping. No bottleneck. No chunking. Just full-speed inference.
Compatible model formats
GGUFGGMLHuggingFacevLLMOllamallama.cppLangChainOpenAI-compatible API
See it in action
Watch the demos.
The hardware
Acer Veriton GN100.
NVIDIA GB10 Grace Blackwell
AI Compute
1 PETAFLOP
Unified Memory
128 GB
Models supported
70B+ parameters
Connectivity
10GbE + Thunderbolt
OS
DGX OS (Linux)
Dimensions
Desktop form factor
💬 Pricing on request — enterprise quotes within 24h. Volume discounts available.
Full specifications
For the technical team.
Processor & Compute
Chip
NVIDIA GB10 Grace Blackwell Superchip
AI Performance
1 PETAFLOP (FP8)
CPU Cores
72× Arm Neoverse V2
CPU TDP
~300W total system
Memory & Storage
Unified Memory
128 GB LPDDR5x (shared CPU/GPU)
Memory Bandwidth
~273 GB/s
Storage
1 TB NVMe SSD (configurable)
Connectivity
Network
10GbE (RJ-45)
High-speed I/O
Thunderbolt 4
USB
USB-A 3.2 + USB-C
Display
HDMI 2.1
Software & OS
Operating System
DGX OS (Ubuntu-based Linux)
CUDA
Latest CUDA Toolkit included
AI Stack
Ollama, vLLM, Docker pre-configured (with service)
API
OpenAI-compatible local endpoint
Physical
Form factor
Compact desktop
Warranty
3-year on-site (official Acer)
Certifications
CE, FCC, RoHS
Power
AC adapter, ~300W peak
Security & compliance
Your data never moves. Ever.
Every query, every document, every model response stays on your hardware — physically, legally, permanently.
Zero data exfiltration
No internet connection required. Models run fully on-premise. Your prompts, your outputs — never sent anywhere.
GDPR compliant by design
No third-party processors, no data transfers outside your jurisdiction. GDPR Articles 28 & 32 satisfied without contracts.
No vendor model access
Unlike cloud APIs, your model inputs never train Anthropic, OpenAI, or any third party. Absolute confidentiality.
Air-gap capable
Remove the ethernet cable. The GN100 runs completely offline — no updates, no telemetry, no external calls.
Full audit trail
All inference logs stay on your server. Query history, access logs, model versions — auditable, exportable, yours.
Role-based access control
Restrict which teams and users access which models. API key management, network-level isolation, SSO-ready.
Relevant frameworks:GDPRHIPAAHDSISO 27001SOC 2MiFID IIDORANIS2
What our clients say
In production.
“We needed AI for our operations but couldn't send client data to any cloud service. The GN100 gave us the performance of GPT-4 with the security of an on-premise server. Setup was done in two days — our team was running workflows by end of week.”
I
Iris Gallery
Enterprise client · Paris
✏️ Quote to be confirmed
Integrations
Plugs into your existing stack.
The GN100 exposes an OpenAI-compatible API locally. Any tool that works with ChatGPT works with GN100 — just point the endpoint to your server.
OpenAI-compatible API
ProtocolDrop-in replacement for OpenAI endpoints. Change one URL — all existing integrations work.
Ollama
RuntimeLocal model management: pull, run, and switch models with one command.
LangChain / LlamaIndex
FrameworkFull agent and RAG pipelines against your private documents and databases.
Slack & Teams
MessagingInternal AI bots querying your knowledge base — self-hosted, no data leaves your Slack workspace.
n8n / Make / Zapier
AutomationWorkflow automation nodes that call local models. Build document pipelines, summaries, classifiers.
Custom apps
DevREST API, Python SDK, or Node.js — standard HTTP calls to localhost. Any developer can integrate in hours.
# Same code. Just change the base_url.
client = OpenAI(base_url="http://your-gn100-ip:11434/v1", api_key="local")
client = OpenAI(base_url="http://your-gn100-ip:11434/v1", api_key="local")
Deployment
Production-ready in one week.
From order to running models. We handle the full stack — you don't need an AI engineer on-site.
Day 0
Order confirmed
Hardware allocated from EU stock
Setup questionnaire sent (network config, model preferences)
Engineer assigned to your account
Day 1–3
Delivery & hardware setup
GN100 delivered to your site
Physical installation and network integration
OS, CUDA drivers, and AI stack deployed (Ollama, Docker, vLLM)
Day 3–4
Model deployment
LLaMA 3 70B (or your preferred model) deployed and tested
OpenAI-compatible API endpoint live on your network
Security hardening and access control configured
Day 5–7
Team onboarding
Half-day workshop for your technical team
Use-case specific configuration (chatbot, code assistant, doc analysis)
First internal workflows running in production
Month 1
30-day support
Dedicated support channel
Model updates and performance tuning
Additional team training sessions if needed
FAQ
Questions we get from IT directors.
Can we run LLMs fully offline?
Yes. The GN100 runs entirely on-premise. No internet required after initial setup. Models are stored and served locally — zero data leaves your network.
What models are supported?
LLaMA 3 (8B, 70B, 405B), Mistral, Mixtral, DeepSeek, Qwen, Gemma, and any GGUF or HuggingFace-compatible model. We handle the full deployment stack via Ollama and vLLM.
Do we need a GPU?
The GN100 uses an NVIDIA GB10 Grace Blackwell Superchip with 128GB unified memory shared between CPU and GPU. No separate GPU purchase needed — it's all integrated.
How does it integrate with our existing tools?
The GN100 exposes an OpenAI-compatible API locally. Any tool that works with ChatGPT (Slack bots, custom apps, n8n, LangChain) works with GN100 — just point the endpoint to your server.
What's the ROI vs cloud AI?
At 50 active users on GPT-4, typical monthly spend exceeds €5,000. GN100 is a one-time hardware cost with zero per-query fees. Most teams break even in under 12 months.
What's included in the setup service?
Hardware delivery, OS and AI stack installation (Ollama, Docker, drivers), model deployment, network security configuration, team onboarding, and 30 days post-installation support.