AGGX
STRATEGIC FINANCING · CLOSED

AI Infrastructure for Distributed GPU Workloads

AGGX develops tools for workload routing, execution verification and capacity pricing across distributed GPU clusters.

Mispa Holdings, operator of AGGX, has closed a $5,000,000 Strategic Infrastructure Financing Round to accelerate this work and deploy it on NVIDIA Blackwell B200 hardware.

Contact SalesExplore the technology
NVIDIA Blackwell B200 GPU cluster
NVL72 · 1.8 TB/s NVLink · 180GB HBM3e
$5M
Strategic Round
1.5T+
Telemetry Points
200B+
MoE Gating Network
B200
NVIDIA Blackwell
Use of Compute

Core AI Engineering Modules

To process terabytes of real-time telemetry, continuous workload parallelization, and execution verification across thousands of distributed GPUs, AGGX is training three core proprietary models.

01Workload Placement

GPU Workload Router

AI-driven workload placement across heterogeneous clusters

Mechanism

Ingests continuous telemetry streams — VRAM latency, CUDA core thermal metrics, network jitter, and matrix-multiplication throughput — from globally aggregated nodes.

Training Data Scale

Trains on over 1.5 trillion time-series telemetry points to continuously improve a reinforcement-learned Mixture-of-Experts model capable of dynamic, microsecond-level AI task partitioning across heterogeneous hardware.

  • Topology-aware placement for tensor, pipeline, and data parallelism
  • Pre-emptive migration away from thermally throttled or degraded nodes
  • Continuous retraining as fleet composition changes
02Correctness

Execution Verification

Confirming distributed work ran as scheduled

Mechanism

Uses high-throughput lightweight models to audit distributed worker nodes and confirm they executed full-precision matrix operations without silent precision downgrade or result manipulation.

  • Trusted execution environments (TEEs) for attested compute
  • Signed usage records per slice and per billing period
  • Precision-downgrade and partial-execution detection
03Cost Modelling

GPU Capacity Pricing

Estimating compute cost across regions in real time

Mechanism

Analyzes GPU rental rates, electricity costs and available capacity to estimate compute pricing across regions, so capacity is costed against live market conditions rather than a fixed rate card.

  • Regional spot availability and power-cost curve modelling
  • Per-second cost attribution down to the individual slice
  • Capacity forecasting for reservation planning
Deployment Roadmap

Capital Allocation & NVIDIA B200

Capital Escrow
70%
$3.5M
of the round

Strictly placed in an escrow account dedicated to high-density data centre co-location and NVIDIA Blackwell HGX B200 / NVL72 node procurement.

Why NVIDIA B200 is imperative

The AGGX routing model relies on a 200B+ MoE gating network. The 180GB HBM3e memory bandwidth and 1.8 TB/s NVLink bi-directional topology of Blackwell B200 nodes are technically required to prevent inter-node communication bottlenecks during multi-modal time-series matrix pre-training.

HBM3e Memory
180 GB
NVLink Topology
1.8 TB/s
Gating Network
200B+ MoE
Node Format
HGX / NVL72
High-density data centre with NVIDIA Blackwell B200 nodes
High-Density Co-location
Blackwell HGX B200 / NVL72 Deployment

Mispa Holdings, operator of AGGX, is a Korea-based technology company building enterprise GPU virtualization and AI compute infrastructure. The Strategic Infrastructure Financing Round was led by an Asian commercial investment group.

Talk to our teamReturn to AGGX