Data within your environment
Designed to run admitted workloads on customer infrastructure, with access, retention and updates defined for the organisation.
TARGET INTRODUCTION · Q1–Q2 2027
THOLON
SOVEREIGN EDGE AI MICRO DATA CENTER
Compute becomes
engineering capability.
THOLON brings together Reinventy software and NVIDIA Blackwell infrastructure: Tin Man OS for system governance, Alchemi for research and AeroMag Forge for engineering design.
PRODUCT IN DEVELOPMENT · REFERENCE ARCHITECTURE · INDUSTRIAL CONCEPT
01 / REINVENTY SOFTWARE + NVIDIA BLACKWELL
Integration creates
the value.
AI models, technical knowledge, scientific tools and local computing: THOLON is designed to bring them into a governed working environment. The goal is to connect requirements, design alternatives, simulation results and the evidence needed to make engineering decisions.
An integrated platform
Compute, internal networking, control PC, cooling and protected power designed as parts of the same module.
Growth by module
A ten-node foundation, with a proposed expansion path to four modules for more services, specialists and concurrent capacity.
REINVENTY SOFTWARE / THE CORE OF THE PLATFORM
Three technologies.
One industrial capability.
THOLON’s distinctive value comes from our technologies working together. Blackwell provides the computing foundation; Reinventy software connects it to knowledge, tools and development workflows.
TIN MAN OS ↗
Governs execution.
The planned governance layer for coordinating models, tools, resources, access and work plans. A shared context for organising execution and keeping each step traceable.
ALCHEMI ↗
Explores new possibilities.
The research programme for exploring materials, candidates and technological hypotheses, connecting scientific knowledge with calculations and experiments. Proposals advance through dedicated verification.
AEROMAG FORGE ↗
Shapes the design.
The engineering environment for connecting geometry, parameters, simulation tools and comparisons between alternatives. Progress towards components and buildable systems remains subject to engineering verification.
- 01REQUIREMENTS
- 02RESEARCH
- 03DESIGN
- 04SIMULATION
- 05VERIFICATION
THOLON integration is in development. Functions and performance will be qualified for each software version, model and workload. AI capabilities do not replace validation of scientific and engineering results.
02 / SYSTEM ARCHITECTURE
A complete system.
A defined boundary.
Explore the layers of the reference architecture. The cabinet is the project’s industrial concept; the final layout and components will be defined during industrialisation.
INDUSTRIAL DESIGN / REFERENCE CONCEPT01 / COMPUTE
Ten nodes. An expandable foundation.
Ten NVIDIA Jetson T5000 nodes per module, each with local memory and storage. Nodes can host distinct specialists or participate in a distributed execution plan requiring qualification.
03 / TIN MAN OS + LLM
Multiple specialists.
A shared direction.
The same infrastructure supports two distinct design paths. Model choice, tasks and measurements on the actual configuration will determine which path fits.
DOCUMENTSCODETECHNICAL SUPPORT
TIN MAN OSPLANS & RESOURCES
N01N02N03N04N05N06N07N08N09N10
BUSINESS APPLICATIONS
ILLUSTRATIVE / LOGICAL ARCHITECTUREA team of capabilities, coordinated by Tin Man.
Complete models assigned to documents, software development or technical assistance. Orchestration allocates tasks and collects results; each specialist retains its role, permitted data and authorised tools.
Planned product architecture mode. Qualification by model and workload is required before release.
04 / MODULAR BY DESIGN
Start with a module.
Plan for growth.
Explore the proposed configurations. These quantities are nominal design capacities, not measured product performance.
10Thor nodes
20.7 PFLOPSThor theoretical AI peak20,700 TFLOPS · Sparse FP4 · Tensor Core
1,280 GBnominal distributed memory
80 TBproposed raw SSD storage
1integrated control PCs
Reference configuration: 128 GB per node and a candidate 8 TB SSD. Distributed memory and raw capacity before redundancy and reserves; the final commercial configuration remains to be confirmed.
AI / FP4 SPARSE
Compute capacity,
configuration by configuration.
From one module to an extended configuration with an external NVIDIA DGX Station GB300. Compare theoretical AI peaks on the same sparse FP4 basis.
Aggregate theoretical AI peak · up to
20,700TFLOPS
20.7 PFLOPS · Sparse FP4 · Tensor Core
Thor node peak20,700 TFLOPS
DGX Station peak0 TFLOPS
RTX PRO 6000 peak0 TFLOPS
4,000 TOPS = 4,000 equivalent TFLOPS on the same sparse FP4 basis. DGX + RTX: a nominal 24,000 TFLOPS. RTX memory remains separate from DGX coherent memory.
A proposed option with its own power, cooling and integration requirements. It is outside the module and is not required for base operation.
An NVIDIA-supported option, subject to confirmation for the OEM system: 96 GB of dedicated GDDR7, up to 4,000 additional sparse FP4 TOPS and a 600 W board. It also extends visualization and simulation; it does not double DGX AI compute.
OPTIONAL EXTERNAL RESOURCENVIDIA
DGX Station GB300+ RTX PRO 6000 Blackwell Workstation+ 20,000 TFLOPSSparse FP4 · Tensor Core
DGX Station GB300+ RTX PRO 6000 Blackwell Workstation+ 20,000 TFLOPSSparse FP4 · Tensor Core
Arithmetic sum of NVIDIA peak ratings: up to 2070 sparse FP4 TFLOPS per Jetson T5000 and up to 20,000 TFLOPS for DGX Station GB300. This is a distributed theoretical maximum, not a system benchmark or guaranteed performance for a single model. Networking, runtime, sparsity, precision and power limits determine actual results.
Thor memory remains distributed, while DGX coherent memory is separate. Connecting the systems does not automatically create one GPU or memory pool.
| THOLON | Modules | THOLON modules onlyTFLOPS | + 1 external DGX StationTFLOPS | + DGX + RTX PRO 6000TFLOPS |
|---|---|---|---|---|
| THOLON 10 | 1 | 20,700 | 40,700 | 44,700 |
| THOLON 20 | 2 | 41,400 | 61,400 | 65,400 |
| THOLON 30 | 3 | 62,100 | 82,100 | 86,100 |
| THOLON 40 | 4 | 82,800 | 102,800 | 106,800 |
NVIDIA specifications and calculation basis
Modules × 10 × 2070 + DGX × 20,000 + RTX × 4,000 TFLOPS
NVIDIA Jetson T5000 · 2070 TFLOPS FP4 sparse ↗NVIDIA DGX Station GB300 · 20 PFLOPS FP4 sparse ↗NVIDIA DGX Station · RTX PRO GPU ↗NVIDIA RTX PRO 6000 Blackwell Workstation · 4000 FP4 sparse TOPS ↗Jetson T5000: MAXN / 130 W.
PRODUCT SCENARIOS / FROM WORK TO OUTCOMES
Inside the work
of those building what comes next.
Four scenarios illustrate how the platform could be used. These are proposed integration workflows, not services already qualified on a THOLON configuration.
01 / ENGINEERING DEPARTMENT
A new drive system, from requirements to alternatives.
- THE CHALLENGE
- A team is developing a drive system for a machine, robot or drone, balancing size, mass, thermal behaviour and manufacturing constraints.
- THE THOLON WORKFLOW
- Tin Man OS organises work and resources. Alchemi supports research into alternatives; AeroMag Forge connects parameters, geometry and the results of admitted calculations.
- THE DISTINCTIVE VALUE
- Knowledge, alternatives and verification in one workflow: a reasoned shortlist for testing, with traceable sources and assumptions.
02 / MACHINE BUILDER
The next machine starts with the company’s own knowledge.
- THE CHALLENGE
- Manuals, previous designs, field reports and specifications are distributed across departments and archives.
- THE THOLON WORKFLOW
- THOLON is designed to make authorised knowledge searchable and connect it to development workflows. Teams compare requirements, subsystems and alternatives before consolidating the design.
- THE DISTINCTIVE VALUE
- The company’s experience becomes part of engineering. Hardware, software and knowledge can evolve together, with access permissions defined by the customer.
03 / RESEARCH & NEW TECHNOLOGIES
From a materials hypothesis to the test that matters.
- THE CHALLENGE
- A new solution requires comparisons between candidates, expected properties, physical constraints and experimental results.
- THE THOLON WORKFLOW
- Alchemi supports exploration; Tin Man OS coordinates tools and workloads. Predictive models trained and validated on relevant data can help select cases for further numerical analysis and physical testing.
- THE DISTINCTIVE VALUE
- A proprietary learning cycle keeps data, hypotheses and results connected. AI helps choose what to verify; evidence determines what can advance.
04 / INFORMATION ANALYSIS
Multiple sources, an assessment that can be checked.
- THE CHALLENGE
- A technical analysis team needs to compare reports, documents and authorised sources while identifying contradictions and gaps.
- THE THOLON WORKFLOW
- Specialised services could extract information, reconstruct timelines and prepare summaries with references. Tin Man OS governs access and workflow steps; the analyst reviews the final assessment.
- THE DISTINCTIVE VALUE
- Analysis within the organisation’s chosen environment, with explicit provenance and uncertainty. A foundation for confidential work, without implying certification for classified information.
REINVENTY / ONE ECOSYSTEM
One architecture.
From requirements to design.
THOLON brings together Tin Man OS, Alchemi, AeroMag Forge and NVIDIA Blackwell computing. The Reinventy offering combines infrastructure with the software that puts it to work for research and development. Integrations with CAD/CAE tools, customer applications and processes will be defined and qualified for each configuration.
06 / COMMERCIAL ROADMAP
From architecture
to a commercial product.
Target commercial introduction between the first and second quarters of 2027. The path moves from configuration validation to integrated testing and a defined commercial offer.
Models, runtimes & workloads
Bench verification of correctness, resource use, quality and behaviour under faults.
The module as a system
Compute, networking, software, thermal management, continuity and maintenance verified together.
Target commercial introduction
Configurations, availability and support terms defined before orders open.
THOLON
Bring your AI
into your environment.
Start a conversation about workloads, data and integration. The offering is being prepared for 2027; configurations and commercial terms will be confirmed at launch.
Product preview. The cabinet is a concept; the capacities and functions described refer to the design architecture. Final specifications, performance and availability require qualification and confirmation.