EUPHORIAINTELLIGENCE CITY

The computational heart of EIC

A liquid-cooled AI and scientific supercomputing infrastructure designed to turn computation into physical discovery.

Phase 1 configuration is an engineering visualization basis. Final equipment quantities are subject to procurement, power, financing and vendor design.

AI Factory — cinematicPhase 1 visualization · not a photograph of a built plant

Phase 1, as a machine

Measured planning plan — not an issued drawing. Hover a zone for the section sketch. Click to pin it. Dimensions in metres, areas in square metres.

Phase 1 AI Factory measured site planSecure entryOperations168 m²Metsamor / powerHall A8 NVL72120 m²Hall B8 NVL72120 m²Hall C8 NVL72120 m²Hall D8 NVL72120 m²E–HFuture140 m²CoolingAI fabricNetwork core / MMRStorage120 m²Control plane108 m²52.5 m · four halls12.0 m10.0 m26 m87 m envelope10 mNPLANNING PLAN · 87 × 38 m · 3,300 m² footprint · 32 RACKS · ~6 MW IT
Selected · First of four identical Phase 1 hallsCompute Hall A12.0 × 10.0 m120 m²Clear 4.5 m · NVL72 2.3 mNVIDIA’s own scale-out building block is 8 racks / 576 GPUs. Hall A repeats that block. NVLink stays inside each rack — 72 GPUs, not 576.
SECTION · PLANNING1.5 m aisle123456788 × 1.2 m NVL7212.0 × 10.0 m · 120 m²

Planning geometry for visualization. Not an issued architectural drawing. NVL72 shown at 1.2 × 1.2 m footprint, 2.3 m high, 1.5 m service aisle.

What the factory is made of

Compute

32 NVIDIA GB300 NVL72 racks. 2,304 Blackwell Ultra GPUs. 1,152 Grace CPUs. 576 compute trays. Direct liquid cooling. One NVLink domain per rack (72 GPUs).

Data

Hot NVMe for training and checkpoints, object storage for laboratory output, archive for long-term research. Capacity to be sized by workload.

Energy

~6 MW IT / ~7–7.5 MW facility from the Armenian grid. Metsamor Unit 2 (416 MWe operable, WNA July 2026) is the nuclear share of that grid now. A future imported SMR is planned as the national replacement in the 2030s — not inside the factory fence.

Thermal

GPU cold plate → rack manifold → technology loop → CDU → facility water → heat rejection. Optional heat reuse. Not a row of residential CRAC units.

This is not a data center. This is the computational engine of Euphoria Intelligence City.

National nuclear · public record

Power now: Metsamor. Later: an imported SMR.

The AI Factory draws from the Armenian grid. Metsamor is the country’s operating nuclear plant and the Phase 1 planning source for firm low-carbon megawatts. A future SMR is a national replacement for Metsamor, sized for Armenia’s electricity system (hundreds of megawatts), not for the factory’s 6–7.5 MW. Neither plant is inside the Phase 1 $140–280M factory envelope.

Aerial view of the Metsamor nuclear power plant, 2023
Wikimedia Commons — Metsamor NPP aerial, 2023

Armenian Nuclear Power Plant (Metsamor / ANPP)

Metsamor, Armavir Province — about 30–36 km west of Yerevan

Operator
Haykakan Atomayin Electrakayan CJSC
Reactor
One operating VVER-440 Model V-270 (Unit 2). Unit 1 shut in 1989 and is being decommissioned.
Operable now
416 MWe (WNA, July 2026)
Often-cited gross
~440 MW after 2021 uprate
2023 output
2.7 TWh · 30% of 8.8 TWh national generation

World Nuclear Association lists 416 MWe operable (July 2026). After the 2021 turbine modernization, Armenian and secondary sources often cite about 440 MW gross. Original design was 407.5 MWe gross / 376 MWe net per unit.

IEA via World Nuclear Association: Armenia generated 8.8 TWh in 2023, of which nuclear 2.7 TWh (30%). Gas 42%, hydro 18%, solar 8%. Older citations of ~40% nuclear refer to earlier years (e.g. 2015).

What the factory actually draws

Capacity comparison against Metsamor’s 416 MWe operable rating. The factory does not need, and will not be given, a dedicated reactor.

  • Phase 1 IT design

    ~6 MW

    ~1.4% of 416 MWe operable

  • Phase 1 facility demand

    ~7–7.5 MW

    ~1.8% of 416 MWe

  • Halls E–H (later IT)

    ~12 MW IT

    ~2.9% of 416 MWe

  • Full-campus AI potential

    100+ MW

    ~24% of 416 MWe — still a grid customer, not a dedicated reactor

When Metsamor is replaced

In October 2025 the prime minister said Armenia’s next reactor would be a small modular reactor (SMR). That is a national energy decision, not an AI Factory purchase. As of March 2026 the government was evaluating proposals from the United States, Russia, France, China and South Korea. Technology selection is expected in 2026 or 2027. No vendor is contracted.

In February 2026 Armenia and the United States signed a 123 Agreement on peaceful nuclear cooperation, which allows US companies to export nuclear technology, fuel and services if a US design is later chosen. In May 2026 a US State Department FIRST pre-feasibility study (Sargent & Lundy) found four US SMR designs suitable for consideration in Armenia: NuScale Power Module, GE Hitachi BWRX-300, Westinghouse AP300, and X-energy Xe-100. A final position has not been formed. World Nuclear Association currently lists a proposed ‘Armenia 3’ unit of up to 300 MWe. Armenian officials have said the replacement should provide equivalent national capacity — i.e. replace Metsamor’s hundreds of megawatts for the country, not the AI Factory’s few megawatts.

  1. 1976 / 1980

    Two VVER-440 units enter service

    Unit 1 (1976) and Unit 2 (1980). Design life 30 years. Built for a seismic region (V-270).

  2. 1989

    Both units shut after the Spitak earthquake

    The plant itself continued through the 1988 quake without damage, then closed in 1989 on seismic-safety grounds.

  3. 1995

    Unit 2 restarted

    Restarted after the post-Soviet energy crisis. Unit 1 remains shut and is being decommissioned.

  4. 2021

    Annealing, uprate, licence toward 2026

    Reactor-vessel annealing and turbine work. Power increased (sources cite ~15–18% / about 440 MW gross). ANRA licence to September 2026, with a path to 2036.

  5. December 2024

    Second life-extension programme to 2036

    Armenian–Russian commission and Rusatom Service modernization contract to support operation through 2036. Metsamor shut in April 2026 for a scheduled five-month outage for that work.

  6. January 2024

    New plant ‘within 8–10 years’

    Officials said a replacement plant was planned on an 8–10 year horizon — implying first power in the early 2030s if that schedule holds, overlapping Metsamor’s 2036 horizon.

  7. October 2025

    SMR chosen as the next reactor type

    Prime minister: the next unit will be a small modular reactor. Not yet a signed EPC contract.

  8. February 2026

    US–Armenia 123 Agreement

    Peaceful nuclear cooperation agreement. Enables US export of technology and fuel if a US design is selected. Does not by itself build a plant.

  9. 2026–2027

    Technology selection window

    Minister Khudatyan: model selection in 2026 or 2027. Ruling-party programme: clarify the new unit’s technology by end of 2027. Metsamor Unit 2 remains the operating plant until a replacement is commissioned, with 2036 cited as the extended operating horizon.

  10. 2036

    Planned end of Metsamor Unit 2’s current extension

    ARMENPRESS (2026): 2036 is described as the maximum physical operating period now established. World Nuclear Association notes Armenia has said Metsamor stays until a replacement is commissioned. EIC planning assumes grid nuclear from Metsamor through the 2030s, then from the SMR that replaces it — not a private reactor on the campus.

US SMR designs under study — not awarded

FIRST shortlist, May 2026. Russia, France, China and South Korea remain in the national competition. Images are vendor photography.

NuScale Power Module — NuScale Power — vendor photographyUnited States · 77 MWe per moduleNuScale Power ModuleA six-module plant is about 462 MWeGE Hitachi BWRX-300 — GE Vernova — vendor photographyUnited States / Japan · 300 MWe per unitGE Hitachi BWRX-300Single-unit SMR in the same class as WNA’s ‘up to 300 MWe’ placeholderUnited States · ~300 MWeWestinghouse AP300Pressurized-water SMR; on the FIRST shortlist, not selectedUnited States · 80 MWe per unitX-energy Xe-100A four-unit plant is about 320 MWe

Hardware models

Hover a card for the reading. Open the vendor page for the public specification. Phase 1 visualization basis is NVIDIA GB300 NVL72. AMD Helios is a future hall — never mixed inside the same rack.

Product photography from NVIDIA and AMD public pages. These are vendor references, not photographs of a built EIC plant.

NVIDIA GB300 NVL72 — NVIDIA photographyThe Phase 1 compute brick. One NVLink domain of 72 GPUs. 8 racks make a hall; 32 racks make the factory. Liquid-cooled; not an air-cooled server cabinet.Rack-scale computeNVIDIA GB300 NVL7272 Blackwell Ultra GPUs · 36 Grace CPUs · 18 trays · 9 NVLink switch trays · up to ~142 kW · fully liquid cooledNVIDIA · Reference hardware specificationNVIDIA Blackwell Ultra — NVIDIA photographyThe accelerator in every compute tray. Architecture, HBM and NVLink figures are NVIDIA’s. EIC has not measured a production chip.GPUNVIDIA Blackwell UltraTensor computation · HBM · NVLink 5 up to 1,800 GB/s bidirectional per GPU (vendor documentation)NVIDIA · Reference hardware specificationNVIDIA Grace CPU — NVIDIA photographyGrace sits beside Blackwell on the tray so CPU and GPU share coherent paths. 1,152 Grace CPUs in the Phase 1 design basis.CPU on each compute trayNVIDIA Grace CPU2 Grace CPUs per tray · 36 per rack · Arm Neoverse platformNVIDIA · Reference hardware specificationNVIDIA NVLink — NVIDIA photographyCopper and switch trays that make 72 GPUs one machine. Hall-to-hall traffic uses Spectrum-X, not NVLink.Scale-up inside the rackNVIDIA NVLink9 NVLink switch trays per NVL72 · 72-GPU domain · does not span hallsNVIDIA · Reference hardware specificationConnectX-8 SuperNIC — NVIDIA photographyHow a rack leaves its NVLink island. ConnectX-8 carries training traffic onto Spectrum-X Ethernet.Scale-out NICConnectX-8 SuperNICUp to 800 Gb/s · GPU compute fabric · RDMA / RoCE · PCIe Gen6NVIDIA · Reference hardware specificationBlueField-3 DPU — NVIDIA photographyOffloads networking, storage and isolation from Grace. Separate from the GPU compute SuperNIC path.North/south and storageBlueField-3 DPUOne DPU per compute tray in NVIDIA’s GB300 reference architectureNVIDIA · Reference hardware specificationSpectrum-X — NVIDIA photographyCongestion-controlled Ethernet for RDMA AI clusters. This is the scale-out fabric for 32 racks and four halls.Ethernet AI fabricSpectrum-XLeaf-spine GPU compute network, physically separate from managementNVIDIA · Reference hardware specificationNVIDIA Mission Control — NVIDIA photographyThe intended control plane for rack health, firmware and job orchestration. Not a live NOC of a built Armenian plant.Operations softwareNVIDIA Mission ControlRecommended for GB300 NVL72 · Kubernetes and Slurm workloadsNVIDIA · Reference hardware specificationAMD Helios (future hall) — AMD photographyA compatible future hall, not a mixed NVIDIA/AMD rack. Shells E–H can take Helios without touching Halls A–D.Open alternate podAMD Helios (future hall)72 × MI455X · EPYC Venice · Pensando · UALink · Ultra Ethernet. Separate hall only.AMD · Reference hardware specification
  • KubernetesReference hardware specification
  • SlurmReference hardware specification
  • CDU (OCP class)EIC design basis

One NVL72, then the factory

8 racks make a hall. 4 halls make Phase 1. NVLink never leaves the rack.

NVIDIA GB300 NVL72 rack

NVIDIA GB300 NVL72 · vendor photography

NVIDIA GB300 NVL72
ItemPer rackPhase 1
Blackwell Ultra GPUs722,304
Grace CPUs361,152
Compute trays18576
NVLink switch trays9288
Power shelves (33 kW)8256
Out-of-band switches264

NVLink domain remains 72 GPUs per NVL72 rack. Scale-out networking connects racks. 2,304 GPUs work as a distributed compute system — not one NVLink domain.

Phase 1 planning envelope

$140–280M

Illustrative EIC planning band. Not a bid, not a contracted price, not a completed plant. Final figures follow procurement, power, financing and vendor design.

  • Compute — 32 × GB300 NVL72

    Industry reporting on GB200/GB300-class rack-scale systems is often cited around $2–4M per rack. Not a vendor quotation.

    $65–130M

  • Network — Spectrum-X, optics, MMR

    Typically a substantial fraction of compute at this density.

    $12–25M

  • Storage — hot / object / archive

    Capacity not fixed. Workload-sized.

    $8–20M

  • Cooling plant — CDUs, HX, rejection

    High-density liquid cooling, N+1 sketch.

    $12–22M

  • Power campus — substation to UPS/BESS

    On-site substation to UPS/BESS for ~7–7.5 MW. Metsamor and a future SMR are national plants, not this line.

    $15–28M

  • Building, halls, operations, site

    Four halls, ops glass, MMR, landscaping, logistics.

    $25–45M

  • Software, integration, commissioning

    Mission Control, orchestration, acceptance.

    $3–8M

Compute discovers possibilities.

Laboratories discover reality.

Health, bio, materials, robotics, earth and nutrition send models in and measurements back. The factory exists so laboratories can run physical experiments against computed candidates — not so Armenia can host another colocation hall.

Informed by NVIDIA GB300 NVL72 Enterprise Reference Architecture, Spectrum-X, Mission Control, Open Compute Project cooling environments, ASHRAE liquid-cooling practice, and ANSI/TIA-942-C. Vendor specification is labeled separately from EIC planning assumption.