Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.nyc-ai.app/llms.txt

Use this file to discover all available pages before exploring further.

Empire AI is being built as a phased AI supercomputing environment centered on accelerated training, inference, and data-intensive research.

​
Alpha system

Alpha is the first production-facing phase of the platform. Public Empire AI materials describe the October 2024 launch as a first-stage system powered by 96 advanced processors. CCR’s user documentation describes Alpha as the current onboarding and early-use environment while Beta procedures continue to mature. In practice, Alpha is where researchers validate software stacks, data paths, and job patterns before moving to larger allocations.
Alpha is an early-access research environment, not the final target architecture. It is useful for onboarding and experimentation, but the larger Blackwell-based Beta platform is where the consortium’s long-term AI scale begins.

​
Beta system

Beta is the major architecture jump. Empire AI announced Beta in June 2025 with a $40 million investment and described it as roughly 11 times stronger for AI training than Alpha, with about 40 times the inference throughput and 8 times the storage footprint.
Beta uses an NVIDIA DGX SuperPOD built on DGX GB200 systems. That matters because it combines Blackwell GPUs, Grace CPUs, high-bandwidth memory, and a unified NVLink fabric in a design aimed at very large distributed AI workloads.

​
Beta architecture details

ComponentSpecification
GPU architecture288 NVIDIA Blackwell GPUs
CPU nodes60 nodes with NVIDIA Grace CPU Superchips
Blackwell transistor count208 billion transistors per GPU
Process technologyTSMC 4NP
Superchip designDGX GB200 systems built around Grace Blackwell Superchips
InterconnectNVLink unified fabric
Rack-scale bandwidthUp to 130 TB/s across a single NVL72 rack
Rack-scale AI performanceUp to 1.44 exaFLOPS at FP4 precision
Storage tier 120+ PB on the VAST Data AI Operating System
Storage tier 210+ PB on DDN’s data intelligence platform
Operations softwareNVIDIA AI Enterprise and NVIDIA Mission Control

​
Beta hardware

NVIDIA GB200 NVL72 rack

NVIDIA DGX SuperPOD (DGX GB200)

NVIDIA DGX SuperPOD built on DGX GB200
Turnkey AI data-center platform built from DGX GB200 systems. This is the platform that houses Beta’s 288 Blackwell GPUs.

NVIDIA Grace CPU Superchip

NVIDIA Grace CPU Superchip
CPU platform used for HPC and data-processing nodes alongside the GPU tier. Empire AI lists 60 Grace CPU Superchip nodes for those workloads.

VAST Data AI Operating System

VAST Data storage system
Tier 1 storage, 20+ PB. All-flash platform designed for AI training pipelines and high-concurrency reads.

DDN data intelligence platform

DDN storage platform
Tier 2 storage, 10+ PB. Additional high-performance data infrastructure for datasets, checkpoints, and shared research workflows.

​
Why the Blackwell stack matters

The key design choice is not just “more GPUs.” Beta is tuned for modern large-model workflows:
  • FP4 support and the Transformer Engine make it easier to run very large model training and inference efficiently.
  • High-bandwidth NVLink reduces the communication penalty that often slows distributed GPU jobs.
  • Grace CPUs and large storage tiers support the data movement and preprocessing that AI pipelines depend on.
In practical terms, Beta is meant to support workloads that would be difficult or cost-prohibitive on standard institutional clusters: very large multimodal models, long-context training jobs, large-scale simulation-assisted AI, and high-resolution medical or climate datasets.