§ 01 Sovereign AI Stack · 1,801 planned

Planned multi-silicon AI factory

As of August 2026 the Vasilikos campus is pre-construction. Phase 1 CapEx is Tenstorrent Galaxy RISC-V inference. Later halls evaluate AMD Helios, Cerebras CS-3/CS-4, IBM z17, Trainium 4, and photonic fabrics — mixed ISA, vendor-agnostic, no exclusive lock-in.

Planned Galaxy servers · TARGET

1,801

Planned RISC-V chips · TARGET

57,632

CY meter (1 Galaxy = 1 CY) · MODELED

540K tok/s

Power per server · DESIGN

8–10 kW

Air-cooled Galaxy PUE · TARGET

1.20

Architecture · TARGET

P1 RISC-V · mixed later

Reserve Capacity →

Tenstorrent Galaxy Blackhole — planned Phase 1 RISC-V inference server

§ 02 Philosophy · Open architecture

Demand-driven, not brand-driven

The AI silicon market is in rapid design churn — new XPUs, 3D-stacked inference chips, rack-scale open standards, and lightwave fabrics ship on 12–24 month cycles. AGICY evaluates and integrates the most efficient open-weight-compatible hardware per workload class. We do not sign exclusive, decade-long single-vendor infrastructure marriages: SRAs reserve capacity; the technology committee re-scores silicon each generation.

Efficiency First

We select silicon by tokens-per-watt and workload fit, not by brand logo. Galaxy, Cerebras, Helios, Trainium, and photonic paths compete on benchmarks.

Fact-Backed

Every hardware decision is supported by published benchmarks, vendor specifications, and internal validation testing.

No Vendor Lock-In

Open fabrics (OCP ORW, Ethernet, UALoE) and modular halls let us adopt Trainium 4, MI450, or lightwave switching when diligence clears — without rewriting the campus.

§ 03 Zone Architecture · Two planned zones

Two operating zones

Token Factory is planned RISC-V (Galaxy). Training Zone is mixed ISA on evaluation — Helios GPUs, Cerebras wafers, Trainium XPUs. Not two RISC-V halls.

Training Zone

Dense model training, fine-tuning, and research workloads.

HARDWARE

AMD Helios eval (Phase 2) + Cerebras CS-3 / CS-4 (announced 18 Aug 2026 · not ordered) + future Trainium-class XPUs

WORKLOADS

Foundation model training (100B+ params)

Enterprise fine-tuning (LoRA, QLoRA)

Research experimentation

Token Factory

Real-time inference serving at scale. Planned output is tokens — a token factory, not a storage warehouse.

HARDWARE

Tenstorrent Galaxy fleet (Phase 1) + d-Matrix Corsair decode (Watchlist) + photonic fabric eval

WORKLOADS

Production API serving

Real-time chat & assistants

Batch processing & embeddings

§ 04 Heterogeneous Fleet · 8 silicon lanes

Multi-silicon factory — Galaxy + Cerebras + open eval queue

Phase 1 and Phase 2 anchors receive equal architectural weight. Additional lanes (AMD Helios, Taalas HC1, AWS Trainium 4, photonic fabrics) stay on the technology committee watchlist until independent EU workload validation completes.

Phase 1 (Day 1)

Tenstorrent Galaxy

RISC-V · Blackhole

32 Blackhole chips per server · 300W TDP each

BEST FOR Inference at Scale

Phase 1 CapEx anchor: Galaxy air-cooled inference at ~$110K per server (vendor public price). 16 Linux-capable RISC-V cores + ~752 baby RISC-V cores per chip; matrix compute is Tensix engines — not Helios, not Cerebras. Campus is pre-construction.

Phase 2 (Evaluation)

AMD Helios · MI450

OCP Open Rack Wide · CDNA

Up to 72 MI450 GPUs per liquid-cooled rack

BEST FOR Open Rack-Scale Training & Inference

Phase 2 open-rack priority: Meta ORW v3 with UALoE scale-up — volume target H2 2026. Evaluated on workload fit and EU supply chain; no exclusive AMD commitment.

Phase 2+ (Watchlist · AMD deal announced)

Taalas HC1

Model-hardwired ASIC · TSMC 6nm

Llama 3.1 8B cast into silicon · ~53B transistors

BEST FOR Fixed-model ultra-throughput inference

Technology demonstrator that hardwires Llama 3.1 8B instead of streaming weights from HBM. Vendor-claimed ~17k tok/s/user. AMD announced a definitive acquisition agreement (6 Aug 2026) — subject to closing. Primary source: https://newsroom.amd.com/news/amd-acquires-taalas-ai-inference/. Not Phase 1 CapEx; not a Galaxy substitute.

Phase 2 (Watchlist · CS-4 announced, not ordered)

Cerebras CS-3 / CS-4

Wafer-Scale · 4T Transistors / wafer

CS-3: one WSE-3 · CS-4: three WSE-3 Turbo (announced 18 Aug 2026)

BEST FOR Training + High-Throughput Inference

Phase 2 wafer-scale lane: CS-3 WSE-3 with 44 GB on-wafer SRAM, plus CS-4 Nexus (three WSE-3 Turbo wafers, not a WSE-4 die). Not an owned Phase 1 fleet. CS-3 vendor-claimed ~1,800 tok/s on Llama 70B-class; CS-4 vendor-claimed up to 2× vs CS-3 on Cerebras’ comparisons.

Phase 2+ (Watchlist)

IBM z17 / LinuxONE 5

Telum II · on-chip AI + Spyre

Single-frame and rack-mount · GA 12 Aug 2026

BEST FOR AI next to core transactions (fraud, payments, regulated data)

IBM GA 12 Aug 2026: smaller-footprint z17 and LinuxONE 5. Different job than Galaxy — inferencing beside the ledger, not campus GPU training. US vendor; customer-owned iron ≠ CLOUD Act-proof. Quote-only; no endorsement.

Phase 2+ (When Available)

AWS Trainium 4

Custom XPU · Annapurna Labs

UltraServer clusters · FP4 native

BEST FOR Cost-Optimised Training at Hyperscale

AWS targets broad availability ~2027 with ~3× FP8 vs Trainium 3, NVLink Fusion hybrid options, and ~288 GB HBM — tracked for EU-compatible deployment paths, not day-one procurement.

Phase 2–3 (Watchlist)

d-Matrix Corsair

Digital In-Memory Compute

Dual-chiplet PCIe card · TSMC 6nm

BEST FOR Sub-2ms Latency Inference Decode

On Watchlist for decode-phase latency — digital in-memory compute with 4,800 TFLOPS MXINT8 and 300 TB/s on-chip bandwidth. Evaluated alongside wafer-scale and rack-scale paths; no exclusive commit.

Phase 2+ (Evaluation)

Photonic & Lightwave Fabric

Optical Circuit Switching

Vendor-agnostic · PCIe / fabric attach

BEST FOR Higher Effective Tokens / Watt

Industry moving beyond electrical switches: Oriole PRISM (AMD collab), Q.ANT photonic NPUs, and 3D photonic tensor research (e.g. SKYLIGHT) target idle-GPU reduction and tokens-per-watt — evaluated as fabric layer, not religion.

§ 04b Footprint · Galaxy · NVL72

Comparable floor footprint. Not the same cabinet.

A Tenstorrent Galaxy is a 32-chip Blackhole mesh in 6U — approximately 19,200 mm² of silicon at ~600 mm² per die (roadmap neighbourhood; die area, not core area), drawing 8–10 kW operating and air-cooled. Seven occupy 42U of rack space at $770,000 public list and roughly 63 kW IT, before top-of-rack switching. One NVIDIA GB200 NVL72 is a purpose-built liquid-cooled rack at ~$3,000,000 analyst mid (band $2.0–3.4M) and 120+ kW, occupying a comparable floor footprint. Campus is PRE-CONSTRUCTION. Executed offtake is 0.

Tenstorrent Galaxy6U server

$110,000

Public list · per 6U server · 8–10 kW operating · ~12 kW max

42U of rack space before ToR7 × Galaxy · $770,000~63 kW IT · air-cooled · not seven-plus-switching in one 42U drop-in

32 Blackhole · ~19,200 mm² silicon (~600 mm²/die)
16 Linux-capable RISC-V cores + ~752 baby RISC-V cores per chip; matrix compute is Tensix engines

VERIFIED

NVIDIA GB200 NVL72purpose-built rack

~$3,000,000

Analyst mid · per rack · band $2.0–3.4M

Comparable floor footprint1 × NVL72 · ~$3,000,000120+ kW liquid · purpose-built MGX rack, not a 42U Galaxy drop-in

ESTIMATE

Hardware list price for that footprint: $770,000 versus $2.0–3.4M — a 2.6× to 4.4× spread on purchase price alone. That is not a claim that either produces the same number of tokens. Cost-per-token is the thesis; same-model, same-precision benchmarks are in progress. Planned output is tokens — a token factory, not a storage warehouse. Logging and provenance are designed to support customers' EU AI Act Article 12 record-keeping obligations. Hosting compute does not make AGICY the provider of a customer's high-risk AI system. Standalone Annex III high-risk obligations apply from 2 December 2027 (Digital Omnibus).

§ 05 Performance · Labeled bases · Aug 2026

Inference throughput & economics

Four figures, four bases — do not mix them. The retired 400B dense ~2,500 tok/s (2.5× unnamed GPU) row is gone: it collided with vendor-class DeepSeek-R1 671B (~350–400 tok/s/user). Campus is pre-construction. Nothing here is a live meter.

Figure	Value	Basis
CY meter / server	300 tok/s	MODELED commercial meter: 1 Galaxy treated as 1 CY. Same spine as 17T/yr.
CY meter / Phase-1 fleet	540K tok/s	1,801 × 300 = 540,300. MODELED — not a live meter.
SC-36 blended / Galaxy	~6,649 tok/s	MODELED theoretical blended. Fleet ~378T/yr on the homepage. Not 100% utilization. Not the CY meter.
DeepSeek-R1 671B / user	~350–400 tok/s	Vendor-class / research citation — not AGICY lab. Not a 400B dense class row.

ECONOMICS — STRUCTURE, NOT THIRD-PARTY MARGIN PERCENTS

Operator class	How compute is paid for
AGICY	Plans to own the campus iron (pre-construction). Executed offtake is zero. Return math stays under NDA.
Frontier labs	Typically lease hyperscaler GPU. Public margin figures are not AGICY-measured and are omitted here.
GPU clouds / colo	Rent or colo GPU systems. Different stack than planned Galaxy air-cooled halls.

§ 06 Sovereign Infrastructure · Pre-construction

Planned self-powered campus. Air-cooled Galaxy racks. EU-sovereign.

Renders and energy cards are the planned Vasilikos campus as of August 2026 — not a live hall. Phase 1 Galaxy racks are air-cooled (zero water at the rack). Campus PUE 1.20 TARGET is the facility envelope — a seawater-plant design target on the homepage, not a claim that Galaxy servers are water-cooled. Behind-the-meter generation is a design target — not ordered. Phase 2 Helios / Cerebras halls would be liquid-cooled if procured.

AGICY planned Vasilikos campus — aerial concept, pre-construction

Server Hall (concept)

Planned 1,801 Galaxy racks · 57,632 RISC-V chips · campus is pre-construction

NOC / SOC (concept)

Planned 24/7 monitoring of compute, power, security and network

Visitor Lobby (concept)

Sea-facing entrance — design target, not a live hall

Rooftop (concept)

Rooftop solar + amenity courts are a campus envelope TARGET on the 25,000 m² Site A parcel — not a claim that PV covers ~80% of the site, and not 42 MW generation (that is the Boom ISO TARGET, separate from roof PV)

Power Generation (planned)

Behind-the-meter generation is a design target — not ordered · investor briefing for MW model

Campus amenities (concept)

Ground-level cafeteria and gardens under the elevated facility — design target

🏗️Pre-construction

Campus status

Vasilikos energy zone, Limassol — planned sovereign AI campus. No live halls, turbines, or meters.

🖥️1,801 Galaxy

Phase 1 fleet

Air-cooled RISC-V inference — Phase 1 CapEx anchor. Helios, Cerebras, and Taalas remain Phase 2+ evaluation.

❄️PUE 1.20 TARGET

Cooling design

Phase 1 Galaxy racks: air-cooled, zero water at the rack. Campus PUE 1.20 TARGET is the facility envelope (seawater plant design target) — not rack water cooling. Phase 2 liquid halls only if procured.

⚡Design target

Power posture

Behind-the-meter generation is modeled for investors — not ordered and not operational. Facility MW, fuel, and turbine diligence: investor briefing / data room.

Facility MW, fuel €/MWh, turbine slot, and HRSG diligence are in the investor energy model — not repeated on this public page. Investor relations →

§ 07 Phase 2 Planning · Vendor-agnostic roadmap

Phase 2 — Helios, Taalas, Cerebras, IBM z17, Trainium 4 & lightwave

Phase 1 deploys 1,801 Tenstorrent Galaxy servers for air-cooled sovereign inference. Phase 2 opens liquid-cooled training halls with AMD Helios / MI450 open racks first on the evaluation queue, then Taalas HC1 (model-hardwired inference · AMD acquisition agreement 6 Aug 2026 · not ordered · AMD newsroom), Cerebras CS-3 / CS-4 (CS-4 announced 18 Aug 2026 · not ordered), IBM z17 / LinuxONE 5 (on Watchlist), plus AWS Trainium 4 (~2027), d-Matrix Corsair (on Watchlist), and photonic fabrics. Nothing is procured on faith or exclusive vendor contracts.

Phase 1 · Tenstorrent Galaxy briefing

Planned Phase 1 CapEx — 1,801 Galaxy RISC-V servers. Campus is pre-construction. Cerebras CS-4 is announced, not ordered. We do not OCR or invent financials. This is Aphrodite, an AI agent — not a human spokesperson.

Cerebras CS-3 · Aphrodite Video Briefing (CS-3 footage · CS-4 on the hardware page)

Phase 2+ · Taalas HC1 product brief (vendor media · watchlist)

Model-hardwired Llama 3.1 8B demonstrator — vendor-claimed ~17k tok/s/user. AMD definitive acquisition agreement announced 6 Aug 2026 (subject to closing). AMD press release ↗. Not Phase 1 CapEx. Not AGICY lab measurement.

Taalas HC1 technology demonstrator board

Taalas vendor chart of tokens per second per user versus GPU baselines

Hardwired model silicon for fixed Llama 3.1 8B paths — complementary to Galaxy, not a substitute.

Taalas hardware page →

Phase 2 · Evaluation