§ 01 Sovereign Compute · planned capacity
Sovereign Compute. Bare‑Metal RISC‑V.
EU‑sovereign infrastructure designed for AI workloads—training, inference, and fine‑tuning—without vendor lock‑in. Phase 1 facility is pre‑construction; figures below are design targets unless marked otherwise.
§ 02 Compute vs GPU Cloud · jurisdiction-first
Sovereign compute changes the risk model
The differentiator isn't just performance. It's jurisdiction, isolation, and whether your stack is structurally dependent on a proprietary accelerator ecosystem.
| Specification | AGICY Sovereign Compute | US GPU Cloud Provider |
|---|---|---|
| Hardware | Tenstorrent Galaxy — 32 Blackhole chips/server | NVIDIA HGX — 8 GPUs/node |
| Architecture | RISC‑V open ISA, no proprietary firmware lock‑in | CUDA proprietary, export‑controlled |
| Deployment Model | Bare‑metal planned: reserved, dedicated, on‑demand tiers | Virtualised instances only |
| Data Residency | Vasilikos, Cyprus — 100% EU sovereign by design | US regions, CLOUD Act applies |
| Networking | High‑bandwidth fabric + isolation targets | Shared tenancy + virtualised network |
| Cost Structure | Projected structural advantage via open ISA + no licensing | Premium + ecosystem licensing |
| Power Source | Behind-the-meter generation design target + solar (not IT load; not ordered) | Shared grid, variable availability |
§ 03 TCO Calculator · model your demand
Model your token economics
Estimate the cost difference between sovereign routing and traditional cloud GPU pricing. Use this to choose a reservation tier in SRA.
Interactive TCO Calculator
Illustrative pre-COD model — projected savings if shifting inference volume from legacy GPU clouds to planned AGICY Sovereign RISC-V Compute (not live list prices).
Baseline (illustrative / design target): Planned AGICY RISC-V ($500/B inference units) vs GPU Cloud ($2500/B inference units). Model excludes data egress fees; AGICY's design target is zero egress on sovereign routing — not a live fee schedule.
§ 04 Bare‑Metal Sovereign Compute · site today · GPUs now
Built for control, not tenancy
Dedicated Tenstorrent Galaxy servers — 32 Blackhole RISC‑V chips each — air-cooled Galaxy halls on a 25,000 m² Vasilikos campus (PUE 1.20 design TARGET; pre-construction). Facility MW and generation modeling: investor briefing. Reserve Day‑1 allocation via SRA. Academy seats ship with every tier — Academy training. Need GPUs before COD? GPU Bridge.
Rent GPUs now · EU bridge
Partner capacity before COD
EU-region partner GPUs for training or serving APIs now. Not Vasilikos campus silicon, not Tenstorrent Galaxy, not an SRA CY meter. Capacity is reserved per request and quoted in EUR.
From-price and flags are live Europe rows for that GPU class. A cheaper hour may sit outside Europe — it is not this card. Not Vasilikos.
- Europe soil
- Europe soil
- Europe soil
§ 05 Compute Access · delivery modes
Three modes. One sovereignty baseline.
Start with burst experimentation, graduate into reserved capacity, and move to dedicated racks when you need isolation guarantees.
- No commitment
- Burst workloads + prototyping
- Designed for fast provisioning after COD
- Predictable allocation
- Best economics for production inference
- Pairs with SRA capacity reservation
- Private racks + networking
- Highest isolation posture
- Enterprise controls + custom zones
- Fine‑tuning, RLHF, sovereign MLOps
- Training hours included with SRA tiers
- Certification path + partner network
§ 06 Briefing · listen + read
Hear it. Then size it.
This is Aphrodite, an AI agent.What hardware is planned?
Phase 1 is planned as Tenstorrent Galaxy servers with Blackhole RISC‑V accelerator chips (pre‑construction; not live inventory). Phase 2 design targets include wafer‑scale and additional specialised accelerators.
How does pricing work?
On‑demand, reserved, and dedicated tiers are planned for target COD H2 2027 — illustrative until commercial operation. Use the TCO calculator to model demand; reserve interest via SRA for priority allocation. Executed offtake today is zero.
Can I deploy on my own premises?
On‑premises / air‑gapped deployments are a roadmap design target, not a live product today. Installation and support options are planned for after campus COD; contact us to discuss requirements.