Build AI on infrastructure you control, in the jurisdiction you choose.
AI labs, SaaS platforms, and enterprise teams run AI agents, retrieval pipelines, and vector databases on ZCP today, keeping training data and models in the region they select under a single accountable operator. GPU is available now for private cloud and bare metal; GPU on public cloud is on the roadmap.
GPU instances, coming soon
CPU and high-RAM VM instances are available today for AI agents, RAG pipelines, embeddings, and CPU-optimized inference. GPU-accelerated compute for large model training and high-throughput inference is on the roadmap. Join the waitlist.
Ship agent workloads today
High-memory instances run agent frameworks, embedding pipelines, and retrieval orchestration right now, with no GPU required for most agent workloads. You can move from idea to production without waiting on scarce accelerator capacity.
Your jurisdiction, your control
Training data and model weights stay in the region you select, under one named operator. That matters when export controls, customer contracts, or internal policy dictate exactly where sensitive AI development is allowed to run.
A foundation for retrieval and RAG
Fast persistent storage and high-memory instances run the leading vector databases at scale, with object storage for training data and model checkpoints, so your retrieval stack performs without exposing it to infrastructure you do not control.
GPU available now for serious training
GPU-optimized bare metal and private cloud are available today for large model training, fine-tuning, and high-throughput inference, with GPU on public cloud on the roadmap. Talk to us to scope a dedicated GPU build-out.
Common workloads
Where organizations like yours put ZCP to work.
AI agent hosting
Deploy multi-agent frameworks, LangChain, AutoGen, CrewAI, LlamaIndex, on high-RAM VM instances (up to 96 GB). Run agent orchestration, tool-calling loops, and memory systems on dedicated CPU compute. Available now.
Ollama on YUL Intel VMs
Run the Ollama Marketplace image on YUL-1 Intel compute. The ci2c.4xl plan provides 16 vCPU and 64 GB RAM, with a root disk sized to your model (for example 320 GB). It suits CPU inference with large quantized models such as Llama 3.3 70B. Response speed depends on model size, context, and concurrency. GPU public-cloud VM inference is not currently available.
RAG pipeline infrastructure
Vector databases, document chunking, embedding pipelines, and retrieval APIs on dedicated, region-controlled compute. No shared-tenant resource contention. NVMe block storage for index persistence.
LLM inference, CPU-optimized
Run llama.cpp, Ollama, or vLLM with CPU-optimized models (Phi-3, Mistral 7B quantized, Gemma 2B) on high-core-count VM instances. Practical for agent sub-tasks, classification, and summarization at lower cost.
Embedding generation
Bulk embedding jobs on large-vCPU instances for indexing pipelines, semantic search, and document retrieval. Predictable CAD compute cost, no per-token billing, no shared quota limits.
MLOps and experiment tracking
Host MLflow, DVC, or custom experiment tracking on dedicated VMs with persistent NVMe storage. Keep training runs, metrics, and model artifacts under your control.
GPU training and large-model inference
GPU-optimized bare metal and private cloud are available now for large model training, LoRA fine-tuning, and high-throughput inference. GPU on public cloud VM instances is on the roadmap. Contact us to scope a GPU deployment.
ZCP Object Storage
Object Storage for AI datasets and checkpoints
Machine learning teams need one place for datasets, checkpoints, evaluation output, and the documents behind a retrieval pipeline. ZCP Object Storage keeps those artifacts in the region you select, behind the industry-standard S3 API, and data moves in and out under published terms with no transfer charges.
Common S3 workflows
- Upload raw and curated training datasets
- Write checkpoints and training logs from your jobs
- Keep evaluation artifacts and experiment outputs
- Store source documents before chunking and embedding for RAG
- Move artifacts between S3-compatible tools with standard clients and SDKs
Questions to settle before you move data
- Where should the datasets and model artifacts live?
- Do any datasets carry sensitive or customer data that needs a separate review?
- Is shared, self-service storage enough, or do you need a dedicated cluster for your AI data?
Comparing with Hugging Face Storage Buckets
Hugging Face Storage Buckets also offer S3 API access, through a gateway that its documentation describes as single-region, path-style only, and without object versioning. ZCP Object Storage keeps data in the region you choose and supports versioned buckets with Object Lock. Billing units, request charges, and AI workflow features differ between the two, so compare them against your own dataset sizes and access patterns.
ZCP Object Storage is self-service and S3-compatible, in the region you choose. If you need dedicated hardware or root access, ZSoftly Cloud Storage is a dedicated, single-tenant cluster with object, block, and file storage.
Get started
Start building on
sovereign cloud today.
Start with CA$100 in credit. Earn up to CA$300 as you spend. No long-term contract.
Eligible new accounts only · CA$200 spend unlocks CA$200 more, up to CA$300 · Offer ends Dec 31, 2026 · Prepaid adds a CA$1 verification payment, credited to your account · Terms apply · ZSoftly Technologies Inc., Ottawa ON