The Engine of
Distributed AI Factories
TonoForge™ deploys 2× racks of enterprise-grade GPU compute — choose AMD Instinct MI355X (128 GPUs) or NVIDIA GB300 NVL72 (144 GPUs) — from MiTAC, Supermicro or Dell, inside a single 300 kW containerised AI factory.
Choose Your Compute — AMD or NVIDIA
Each TonoForge™ unit houses 2× full racks of GPU compute, delivered pre-integrated and factory-tested. Choose between AMD Instinct MI355X or NVIDIA GB300 NVL72 platforms, sourced from MiTAC, Supermicro, or Dell.
Both configurations deliver enterprise-grade AI training and inference capacity with liquid cooling, 800 Gb/s networking, and full TonoFabric™ orchestration.
- AMD path: 2× 64-GPU MI355X racks — 128 GPUs, 36.8 TB HBM3E, 161 PFLOPS FP4
- NVIDIA path: 2× GB300 NVL72 racks — 144 GPUs, 42 TB HBM3e, 2.2 EXAFLOPS FP4
- Rack vendors: MiTAC MR1100L · Supermicro AS-4126GS / SRS-NVL72 · Dell XE9712
- Centralised orchestration via TonoFabric™ with RL-based optimisation
TonoForge™ deploys battery storage amplification to solve the grid bottleneck — one of the biggest blockers for AI infrastructure. By pairing compute with on-site LFP energy storage, TonoForge™ delivers peak AI performance without requiring expensive grid upgrades.
Heat waste is recovered and redistributed to adjacent buildings, creating a circular energy system with a PUE close to 1.05 — near-theoretical perfection.
- Deploy full 300 kW AI capacity without grid infrastructure upgrades
- Battery packs absorb peak loads and smooth power delivery across 128–144 GPUs
- Heat recovery converts waste into usable building heat — a revenue stream
- PUE ~1.05 — near-perfect energy efficiency at scale
Unlike traditional data centres that take 3 to 5 years to plan, permit, and build, TonoForge™ is factory pre-integrated and tested — arriving at site ready for commissioning in just 5 to 20 weeks.
Both AMD MI355X and NVIDIA GB300 NVL72 racks are pre-configured by MiTAC, Supermicro, or Dell, then integrated into the TonoForge™ container with battery, cooling, and TonoFabric™ orchestration before shipping.
- Factory pre-integrated with liquid cooling, power, and networking
- 5 to 20 weeks vs 3–5 years for a traditional data centre
- Modular expansion — 150 kW to 300 kW+ without downtime
- Re-uses existing telecom sites — no greenfield land cost
TonoForge achieves a 5x cost reduction vs traditional data centres through distributed architecture, renewable energy co-location, battery amplification, heat recovery monetisation, and re-use of existing telecom infrastructure.
Operating costs are further reduced through TonoFabric’s reinforcement learning engine, delivering 15%+ improvements in efficiency over time.
- Co-locate with renewable energy sources
- Re-use existing telecom sites — no greenfield land cost
- Monetise waste heat to adjacent buildings
- RL-based orchestration cuts operational energy costs 15%+
TonoForge Configurations at a Glance
| Specification | AMD MI355X Config | NVIDIA GB300 Config |
|---|---|---|
| GPU Model | AMD Instinct MI355X (CDNA 4) | NVIDIA B300 Blackwell Ultra |
| GPUs per Rack | 64 (8 servers x 8 GPUs) | 72 (18 trays x 4 GPUs) |
| Total GPUs (2 Racks) | 128 | 144 |
| Memory per GPU | 288 GB HBM3E | 288 GB HBM3e |
| Total HBM Memory | 36.8 TB | 42 TB (21 TB/rack) |
| Memory Bandwidth | 8 TB/s per GPU | 8 TB/s per GPU |
| FP4 Peak (per rack) | 80.5 PFLOPS (MXFP4) | 1.1 EXAFLOPS (dense) |
| FP8 Peak (per rack) | 40.3 PFLOPS | ~360 PFLOPS |
| TDP per GPU | 1,400W | 1,400W |
| GPU Interconnect | 4th Gen Infinity Fabric + PCIe 5.0 | NVLink 5 (1.8 TB/s per GPU) |
| Network per GPU | Pensando Pollara 400 + P2200G | ConnectX-8 (800 Gb/s) |
| Host CPU | 2x AMD EPYC 9755 per server | 36x Grace CPU per rack (LPDDR5X) |
| System RAM | Up to 6 TB DDR5-6400 / server | 17 TB LPDDR5X + 144x E1.S |
| Cooling | Direct Liquid Cooling (cold plate) | Full Liquid Cooling (rack-scale) |
| Server Form Factor | 4U per server (MiTAC G4826Z5) | Compute tray (4 GPUs each) |
| Rack Format | 48U EIA | 48U rack-scale |
| Rack Power | ~120 kW per rack | 132 kW (8× 33kW shelf) |
| Rack Vendors | MiTAC MR1100L-64355X-02, Supermicro AS-4126GS-TNMR | Supermicro SRS-GB300-NVL72, Dell XE9712 |
| Datatype Support | FP64/32/16 BF16 FP8 MXFP6 MXFP4 | FP64/32/16 BF16 FP8 FP4 INT8 |
| Software Stack | ROCm (PyTorch, vLLM, SGLang) | CUDA (TensorRT, Triton, NeMo) |
Both configurations are integrated into the TonoForge container with battery amplification, heat recovery, and TonoFabric orchestration. Power, cooling, and networking are pre-configured at factory.
How TonoForge™ compares
Five approaches to AI infrastructure — from prefab pods to hyperscale campuses. Not all are created equal when speed, sovereignty, and sustainability converge.
| Dimension | APrefab Pod | BPFM CoolChip | CRuggedised MDC | DHyperscale Campus | TonoForge™Distributed AI Factory |
|---|---|---|---|---|---|
| Architecture | Prefab pod shell 8–42 racks per pod | Modular building units Multi-MW assembly | Shipping container MW-class relocatable | Purpose-built campus Multi-GW portfolio | Self-contained container 2-rack AI factory · 300 kW |
| Compute Integration | None — shell only BYO servers & GPUs | None — shell only BYO compute platform | GPU-ready Configurable density | GPU clusters available NVIDIA Preferred Partner | Factory-integrated MI355X / GB300 NVL72 |
| GPU Density per Unit | N/A Infrastructure only | N/A Infrastructure only | Not disclosed Chip-agnostic | 20k+ GPUs / 50 MW Air + liquid halls | 128–144 GPUs / 300 kW 480 GPUs/MW density |
| Time to Deploy | 30% faster than on-site Weeks to months | 50% faster than on-site Months | Weeks Rapid relocatable | 12–24+ months Fixed campus build | 5–20 weeks Turnkey containerised |
| Site Requirements | Prepared building shell Grid + permits + land | Prepared site or retrofit Grid + permits + land | Minimal — pad or surface Power-agnostic | 1,800+ acres owned Grid + substation + permits | Flat surface only Existing telecom sites OK |
| Cooling Technology | Air / InRow / RDHx / DLC Flexible options | Direct-to-chip liquid CoolChip CDU | Configurable Air or liquid per spec | Air + liquid (new halls) 10 MW DLC facility | Full DLC integrated 250 kW CDU · N+1 pumps |
| PUE | ~1.3–1.4 Improved vs traditional | Improved vs traditional No figure published | Not disclosed — | 1.1 Renewable grid | ~1.05 DLC + heat recovery |
| Energy Strategy | Grid-dependent No on-site generation | Grid-dependent UPS/battery backup only | Power-agnostic Stranded gas / solar / grid | 100% renewable Grid + RECs | Battery amplification 2× factor · renewable co-loc |
| Heat Recovery | Not integrated — | Not integrated — | Not integrated — | Not integrated — | Circular architecture HVAC / district heating |
| AI Orchestration Software | EcoStruxure DCIM Facility monitoring | Vertiv DCIM Facility monitoring | Edge Platform (AEP) Bridge GPU orchestration | Cloud dashboard GPU cluster mgmt | TonoFabric™ RL optimisation · workload sched. |
| Relocatability | No Fixed installation | No Assembled on-site | Yes Truck / ship / rail | No Fixed campus | Yes ISO container · any site |
| Modular Scaling | Pod-by-pod 8–12 rack increments | Building-by-building MW increments | Container-by-container Leviathan stacking | Hall-by-hall Multi-MW phases | 150 kW → 300 kW → n MW Add units to mesh |
| Data Sovereignty | Customer-managed Depends on location | Customer-managed Depends on location | Sovereign by design US / allied nations | Fixed jurisdictions US & Canada only | Deploy anywhere EU sovereign · GDPR-ready |
| CAPEX Efficiency | €10–15M / MW Shell only — no compute | €12–16M / MW Shell only — no compute | €12–15M / MW Undisclosed detail | €10–15M / MW Fixed campus scale | ~€2.5M / MW 5–6× less · compute included |
| Best Suited For | Enterprise white-space fit-out inside existing facilities | New-build or retrofit AI halls at prepared sites | Remote / defence edge AI in contested environments | Hyperscale cloud & GPU-as-a-service at fixed campuses | Distributed sovereign AI at telecom / renewable sites |
Benchmark based on publicly available product specifications and press materials as of Q3 2025. A–D represent representative industry approaches, not specific endorsements or claims about named competitors. TonoForge™ figures reflect Tonomia design specifications.
Deploy TonoForge at your core network
Connect with our engineering team to explore a deployment tailored to your GPU requirements, network topology, energy profile, and scale. Choose AMD MI355X or NVIDIA GB300 NVL72 from MiTAC, Supermicro, or Dell.