Dedicated AI inference capacity for regulated industries: banks, insurers and the software vendors that serve them. Reserved by the month, with a ready DORA and MNB contract pack and audit rights. The alternative to buying your own GPU servers for sensitive workloads.
Quick to start, but the data leaves the building and the DORA vendor obligations are hard to cover.
A US parent company, standard contracts and costly on-site audits. GPU-hours here are the most expensive on the market.
An 8-GPU server costs over $150K. Procurement takes four to six months, and you need a team to run it.
Dedicated capacity in Budapest for a monthly fee. We own the hardware, the compute time is yours alone, and the contract pack comes ready.
For document work, classification and drafting, open-weight models now handle work that needed the most expensive US APIs a year ago. And they run on hardware your bank can inspect.
Since January 2025, DORA makes financial entities answer for every ICT vendor. The MNB's Recommendation 2/2025 expects cloud contracts to record where data is processed, at least down to the data center.
The list price of an RTX PRO 6000 went from $8,565 in March 2025 to $16,000 in August 2026. Buy your own server today and you carry that price and the obsolescence risk alone.
We bill per GPU-month on fixed-term reservations. Choose raw hardware, or a ready endpoint that runs only for you. The contract pack comes with both.
A dedicated server or GPU, with no hypervisor and nothing shared. Bring your own software stack.
A single-tenant, OpenAI-compatible inference endpoint for open-weight models. We run the hardware and the model; you call the API.
What your compliance and legal teams will ask for, ready at signing.
This is the exact card your workload runs on: NVIDIA RTX PRO 6000 Blackwell Server Edition. It is built for enterprise inference and fine-tuning, and it has more memory per GPU than an H100. Go-live is in the first quarter of 2027.
| Spec | Per GPU | Per server (8 GPUs) |
|---|---|---|
| VRAM | 96 GB GDDR7 with ECC | 768 GB |
| Memory bandwidth | 1,597 GB/s | 12.8 TB/s aggregate |
| Peak FP4 compute | 4 PFLOPS | 32 PFLOPS |
| FP32 compute | 120 TFLOPS | 960 TFLOPS |
| Hardware GPU slices (MIG) | up to 4 × 24 GB | 32 |
| Host interface | PCIe 5.0 ×16 | Dedicated server, no hypervisor |
Built for: production inference, fine-tuning and agent workloads, as a full server, a single GPU or a 24 GB hardware slice. Not built for: training large frontier models. We would rather tell you that before you sign. Card figures are from NVIDIA's datasheet.
Any leading open-weight model, from the US, Europe or China. A few examples, and how much hardware each one needs:
OpenAI's open-weight model. Runs on a single GPU.
Understands more than 140 languages. Good for internal assistants and document work.
A European model under Apache 2.0, with a 256K-token context and image understanding.
Strong at coding and reasoning, MIT-licensed, with a 1M-token context.
In 2026 our team installed and tested open models on RTX PRO 6000 servers in a client project.
Credit application triage, missing-document requests, document checks, searchable internal policies.
Claims intake, comparing contract wording, summaries of customer complaints.
Contract analysis and risk signals alongside your existing portfolio systems.
The AI features in the software you ship to banks run on capacity the bank's auditors accept.
One 8-GPU server, three ways.
| Your own server | 1337 Compute | US hyperscaler | |
|---|---|---|---|
| Upfront investment | Over $150K per server | None | None |
| Time to start | 4–6 months of procurement | A date fixed in the contract, from Q1 2027 | Immediate, if you have quota |
| Operations | 1–2 people on your side | We run it | You configure it |
| Obsolescence risk | Yours | Ours | Theirs |
| Data location | Your premises | Budapest, a named data center | A region, US parent (CLOUD Act) |
| DORA contract, audit rights | Not needed | Ready pack with every reservation | Standard terms, costly on-site audits |
| Contract | – | Hungarian law, in Hungarian or English | Standard terms in English |
| Support | Your own team | A named engineer in Budapest | Ticket tiers |
| What you pay | Hardware, people, power | A monthly fee on a fixed term | Hourly rates plus data transfer |
The cost of owning: an 8-GPU RTX PRO 6000 server costs over $150K in autumn 2026, and the card's list price has almost doubled since March 2025. We can send a detailed comparison for your own workload on request.
Tell us what you want to run, on what data, and what your compliance team expects.
We size and price it, and send the DORA annex with it, so your lawyers see it from day one.
Once signed, the contract sets the go-live date and the SLA. We go live together and stay close through the first weeks.
15% off the launch rate card for 36 months, not indexed. A reservation needs an agreement on GPU count, term and go-live date.
Request an early reservation offerThat takes the hardware here at home, and the know-how that turns it into working AI. We build it step by step, at the pace of signed reservations.
Dedicated bare metal and managed endpoints on RTX PRO 6000 Blackwell GPUs, with the DORA and MNB contract pack.
We bring new servers online at the pace of reservations. When customers need larger models or training, we add further NVIDIA configurations.
A model catalog, fine-tuning on your own data, and benchmarks on Hungarian-language tasks. These are plans: we set the order together with our customers.
1337 Compute is the sister company of 1337 Partners. The core team digitalized a bank from the inside and builds production AI in regulated environments.
NVIDIA RTX PRO 6000 Blackwell Server Edition, 96 GB of VRAM per GPU, 8 GPUs per server. Go-live is in the first quarter of 2027, and your reservation contract fixes the exact date. If you want to test earlier, we run a pilot on partner EU capacity, on terms agreed case by case.
In an enterprise-grade data center in Budapest. Partner selection is in its final stage. The contract names the facility, and we hand over the documentation on physical security, redundancy and certifications.
Yes. The smallest unit is a 24 GB hardware-isolated GPU slice (MIG), the largest a full 8-GPU server. You can also reserve several servers.
A model whose weights you can download. That means it runs on your own hardware, and your data never goes to an outside provider. Leading open models come from US, European and Chinese labs: OpenAI's gpt-oss, Google's Gemma, France's Mistral, and China's Qwen, DeepSeek and GLM. By their makers' own benchmarks, the strongest match the most expensive closed models at coding. For document work and internal assistants they are more than enough.
Any open-weight model: gpt-oss, Gemma, Mistral, Qwen, DeepSeek, GLM, Kimi and others, including versions fine-tuned on your own data. On the managed endpoint we choose and upgrade models together; on bare metal you run whatever you want.
Per GPU-month on 1–3 year reservations, with lower rates for longer terms. The price depends on the configuration and service level, and we send a quote within 48 hours. We do not charge for data transfer (egress).
Probably yes. In MNB practice, hosting hardware can count as outsourcing, and your legal team makes that call. We prepare the data and contract annexes the notification needs.
Not at all. The API is OpenAI-compatible, so your code runs elsewhere too. You use open-weight models and can take the weights with you. Data export is a contractual right, backed by an exit plan.
A fair question, and the answer is architectural. You run open-weight models through an OpenAI-compatible API on standard container images. There is no proprietary format and nothing to rewrite: moving to another provider is a configuration change. Data export is a contractual right with a fixed deadline. Ask us for the exit clause before you sign; we are happy to show it.
The 1337 Compute team, in Budapest. Monitoring and on-call are in place from go-live, and the contract sets the SLA and response times. You reach our engineers directly.
The button opens your email client with the details filled in. You can also write to us directly: koppany@1337.partners
15% off the launch rate card, fixed for 36 months, with priority on capacity.
Get a quote
