Voicy delivers the complete sovereign AI stack: GPU and HPC clusters, the platform that runs on them, models you can licence, fine-tune or have pre-trained from scratch, and custom software shipped with its full source code. It is installed inside your data centre. Nothing is left for someone else to supply.
Most vendors hand you one layer and leave you to find the rest. The integration risk between those layers is where sovereign AI projects fail — and it is exactly the risk no one wants to own. We take the whole stack, and every layer transfers to you at handover.
The software your people actually use: assistants, document processing, call handling, search over internal knowledge, whatever your specification calls for. Built to your requirement, not adapted from a product we already sell.
Large language models, speech recognition, speech synthesis, speaker identification, embeddings and document understanding. Licence one we have already trained, have it fine-tuned to your domain, or have a foundation model pre-trained from scratch when nothing suitable exists.
Data collection and cleaning, tokeniser design, pre-training runs, fine-tuning, alignment and evaluation. Your corpus never leaves your infrastructure, and the pipeline that produced the model stays behind so you can train the next one yourself.
Inference serving, orchestration, retrieval, monitoring and access control — assembled from open components so nothing in the path requires a licence renewal to keep running.
Operating system, GPU drivers, CUDA stack, scheduler, quota and multi-tenant isolation, parallel storage and the network fabric configuration that keeps large jobs from stalling on each other.
GPU and HPC systems specified, procured through allocation channels, racked, cabled, cooled and load-tested in your building until the acceptance report is signed.
Six practices that ship together. You can buy any one of them on its own, but the reason to buy them from one supplier is that no one gets to point at the layer below when something does not work.
Hardware specified, sourced, installed and accepted. We work the allocation channels so you are not waiting behind a queue you cannot see.
Three routes to a model on your own hardware, from a ready Voicy model under licence in a few weeks, through domain fine-tuning, to a foundation model pre-trained on your corpus.
Whichever route produced the model, it has to run fast enough to be used. We tune the serving stack to the throughput your hardware budget allows and put retrieval over documents that are never indexed anywhere else.
The practice Voicy was founded on. Transcription and voice generation built for the language you actually operate in, including ones with no public training data at all.
You write the specification. We build to it and hand over the repository. There is no black box, no runtime licence and no dependency on us to keep it running.
Run it yourself with your team trained, or hand operation to us under a support agreement. Either way the cluster stays yours and the access list stays short.
Three routes to a model that runs on your own hardware. They differ in what you start from, what data you need, what it costs and what you end up owning — and most organisations use more than one of them.
Our production speech and language models, delivered as files and installed on your hardware under a perpetual on-premise licence. Nothing connects back to us, nothing is billed per call, nothing expires.
A licensed Voicy model or an open-weight model, fine-tuned on your own material until it handles your terminology, your accents and your document conventions. The training runs on your cluster and the corpus never leaves it.
A new foundation model pre-trained on your corpus, with a tokeniser designed for your language. The most expensive route, and the only one that works when nothing suitable exists — which is the problem Voicy was founded to solve.
Most projects mix all three: a licensed speech model running in week four, a fine-tuned language model by month five, and a pre-trained foundation model only where nothing else fits. Pre-training is asked for more often than it is the right answer — where an open or licensed model already covers your language and domain, adapting it costs a fraction of the compute and usually performs better. We will tell you which tier each workload needs before you commit the budget.
Every model listed here is delivered as files on your storage: weights, tokeniser, configuration and the code that runs them. There is no hosted endpoint, no API key and no version you did not approve. Each is available at any of the three tiers above.
| Model | Routes available | What you receive |
|---|---|---|
| Foundation LLM | Pre-training on your corpus (03), or continued pre-training of an existing base (02) | Weights, all checkpoints, tokeniser, training code, run configs, evaluation suite |
| Adapted LLM | Fine-tuning of an open-weight or licensed base on your material (02) | Adapted weights, adapters, data pipeline, before-and-after evaluation |
| Speech recognition (ASR) | Licence a Voicy model for a supported language (01), tune it on your audio (02), or train one for a language with no data at all (03) | Model, decoder, lexicon, domain vocabulary, streaming inference server |
| Speech synthesis (TTS) | Licence an existing Voicy voice (01), or build a new one from studio recordings or your audio archive (02–03) | Voice model, vocoder, pronunciation dictionary, streaming server |
| Speaker identification | Licence our embedding model (01), or enrol and tune on your own speakers (02) | Embedding model, enrolment tooling, scoring thresholds |
| Retrieval and embeddings | Open or licensed embedding model (01), tuned on your documents with reranking fitted to your queries (02) | Embedding model, reranker, index build pipeline, retrieval benchmarks |
| Document and vision | Licensed OCR base (01), tuned on your forms, layouts and handwriting (02) | Model, pre- and post-processing code, extraction schema, accuracy report |
Numbers in brackets refer to the three routes above. Where a licensable Voicy model already exists for your language, route 01 puts a working system in front of your users in weeks rather than months, and routes 02 and 03 remain open afterwards on the same hardware. Ask us which languages are covered today.
Software built around your AI stack should not become the thing that locks you in. Every line we write for you is handed over with the repository, the history and the right to change it without asking us. If you replace us tomorrow, your next supplier starts from working code rather than a specification and an apology.
This is the list, not a summary of one. If something on it is missing at handover, the project is not finished. Everything is delivered in German and English.
A provider incorporated in the United States can be compelled to produce customer data under the CLOUD Act, whichever country the servers stand in. An EU region label does not change who holds the keys. Sovereign delivery removes the question by construction.
| Dimension | Public AI API | EU region of a US cloud | Voicy full-stack delivery |
|---|---|---|---|
| Model weights | Never leave the provider | Never leave the provider | Delivered as files — licensed or owned |
| Application source code | Not applicable | Vendor-retained | Transferred with full history |
| Can you train your own model | No | Only on rented capacity | Yes — pipeline stays behind |
| Runs without internet | No | No | Yes — full air-gap supported |
| Cost behaviour | Per token, grows with usage | Per GPU-hour, grows with usage | Capital expenditure, flat after commissioning |
| If the supplier disappears | Service stops | Service stops | Everything keeps running |
The comparison describes the structural position of each model, not the merits of any individual provider. Public APIs remain the right answer for workloads with no confidentiality, residency or continuity requirement.
Three generations of NVIDIA data centre silicon are shipping at once. The right choice is rarely the newest one — it is the one your power envelope, your cooling and your delivery window can actually accommodate.
| Platform | Configuration | GPU memory | Cooling / rack power | Best suited to |
|---|---|---|---|---|
| NVIDIA H200 SXM | 8-GPU HGX node | 141 GB HBM3e per GPU | Air — conventional hall | Inference and fine-tuning; widest availability today |
| NVIDIA HGX B200 | 8-GPU Blackwell baseboard | 180 GB HBM3e per GPU | Air or liquid — roughly 60 kW per rack | The general-purpose choice for training and serving |
| NVIDIA GB200 NVL72 | 72 Blackwell GPUs, 36 Grace CPUs, one rack | 13.4 TB per rack | Direct-to-chip liquid — around 120 kW per rack | Trillion-parameter training and large MoE inference |
| NVIDIA GB300 NVL72 | 72 Blackwell Ultra GPUs, ConnectX-8 800 Gb/s | 20 TB per rack (288 GB per GPU) | Direct-to-chip liquid — new-build halls only | Reasoning models and test-time scaling workloads |
| NVIDIA DGX Station (GB300) | Desk-side Grace Blackwell workstation | 784 GB unified system memory | Air — standard office power | Secure development before cluster time is committed |
Blackwell-class systems are allocated rather than sold from stock, and quoted lead times track allocation tier more than manufacturing capacity. We plan procurement windows of three to nine months into every schedule and tell you early when a specification will not arrive in time. Vera Rubin was announced at CES in January 2026 with broad availability expected from late 2026 into 2027; we build roadmaps around allocation, not announcements.
Thirty-two HGX B200 nodes is a practical size for an organisation running its own large language model, a speech stack and periodic fine-tuning at the same time. It fits a conventional hall with liquid-assisted cooling and leaves headroom for a pre-training run without pausing production inference.
You own the hardware, it stands in your building, your access control governs who touches it.
No network path out. Models and updates arrive on removable media through a documented chain of custody.
Your hardware in a German-operated facility when you have the budget but not the floor space.
Confidential inference stays in-house; non-sensitive pre-training bursts to EU-sovereign capacity.
Eight phases with honest durations rather than optimistic ones. Hardware procurement dominates the critical path, so model and software work begin during it rather than after — the two highlighted phases run in parallel with phase 03.
What you run, what data exists, what may never leave the building, and whether your corpus can support training at all.
Cluster sizing, model strategy, software specification and a five-year cost model against the equivalent cloud spend.
Orders placed through OEM channels, allocation tracked weekly, a written position whenever a date moves.
Corpus collection, cleaning, deduplication and tokeniser design start immediately, on interim hardware where needed.
Application development against your specification, in your repository from the first commit so you can read the work as it happens.
Racking, cabling, coolant commissioning, then a sustained load test with results you can hold the acceptance against.
Licence installation, domain fine-tuning or a full pre-training run, depending on the route chosen — evaluated against the tasks your people do every day rather than public benchmarks.
The manifest is walked item by item, your engineers rebuild the software from source, and you sign it off — or we operate it under support.
Sovereignty is a compliance argument before it is a technical one. We design deployments so the evidence your auditor asks for already exists, in German, on day one.
Obligations for general-purpose and high-risk systems applying from 2 August 2026. Because we train the model, the provenance, data lineage and evaluation records exist as a by-product of the work rather than a reconstruction afterwards.
No third-country transfer, because there is no third country in the path. Processing records, deletion routines and access logging are configured to your retention policy.
Supply chain security obligations for essential and important entities. Hardware provenance, firmware integrity and software bill of materials are recorded through commissioning.
Operating procedures, change control and incident handling written to slot into an existing information security management system rather than sit beside it.
Where a C5 attestation is in scope, the control mapping is prepared with your auditor in mind and the physical controls are yours to demonstrate directly.
Not applicable by construction. There is no US-incorporated provider anywhere in the processing path, and that is a statement you can verify by walking into the room.
Germany is coordinating IPCEI-AI, the first Important Project of Common European Interest aimed directly at artificial intelligence, with over one billion euros committed nationally and nineteen member states in the pre-notification phase.
The German national application deadlines are 31 October 2026 and 30 April 2027. If your project could qualify, the architecture work has to start well before the deadline — talk to us early rather than late.
Voicy supports your compliance programme with documentation, control mapping and technical evidence. Certification and attestation remain matters for you and your auditor.
Voicy began by solving speech recognition for Kazakh, Uzbek and Kyrgyz — languages with almost no public training data and no commercial interest from the large AI labs. We passed 95 percent accuracy in production. That required building the corpus, the tokeniser, the training pipeline and the serving stack from nothing.
That is exactly what full-stack sovereign delivery demands. An organisation that wants to own its models needs a supplier who can train them from scratch, not one who can only call someone else's API.
Send the workload and the constraint. You will get a written architecture sketch, a model strategy and an indicative cost for the layers you need — not a sales call.