AI that lands inside your bank, not across a border.
Banks want generative AI in operations, customer service and engineering — without sending customer data to a third-country API. ARK serves open-weight models inside your own estate and gives your ICT risk framework what it asks for: where inference runs, which model answered, what was logged, and how you would exit.
Europe's central bank warns that relying on foreign AI providers, who can withdraw access or change their terms, hands them “leverage of a kind no trade partner has ever held over Europe”. The US hosts about 75% of the world's AI computing capacity; Europe, 5%.
ECB President Christine Lagarde, speech in Vienna, 14 September 2026. US share from Epoch AI (2025), as cited by the ECB; Europe's 5% from the ECB.
With a second copy of each part of the model, a failed server doesn't take the service down. Requests in flight are replayed on another path, and servers join or leave without a restart.
Read moreTo the first token of a new 64,000-token document. Asked a second question, it answers 9× sooner.
Read moreWhat a bank cannot hand to an outside API.
Customer data stays in-house
Account data, transaction histories and complaints cannot be sent to an outside model provider.
Concentration and exit risk
DORA expects you to manage concentration risk in critical ICT providers and keep a workable exit plan. Almost half of executives say a key business function would stop working if they lost their main AI vendor, and only 6% could stop using it without disruption. Access can also disappear overnight. In June 2026 a US export-control directive forced a US provider to suspend two frontier models for all customers. One returned after 19 days; the other came back only for selected US organisations.
Zapier survey of US executives, April 2026 (n=542). Anthropic statements, 12 and 30 June 2026.
Costs you can't forecast
Banks pay for completed workflows, not tokens. A loan or claims process makes several model calls per case, and a provider's price change can land mid-year. On your own GPUs, cost is set per GPU, not per call.
Many languages, one bank
Branches, call centres and back offices work in Polish, German, English and more.
What banks build on ARK.
These workloads pay back quickly. Each runs on open-weight models inside the bank, or on ARK Cloud in the EU while you prototype with synthetic or public data.
Document extraction
KYC files, loan applications and statements are read by the Qwen vision models and returned as structured JSON your systems can check.
Complaint and email triage
Classify, summarise and route incoming messages with a small, fast model such as Qwen3.5-4B. On your own GPUs there is no per-message cost. On ARK Cloud, 1,000 summaries of 1,100-token documents cost $0.062.
Call summaries
Transcribe calls with Whisper in 99 languages, then summarise them with a chat model, all on your hardware, with subtitles and timestamps if you need them.
Search over policies and products
bge-m3 embeddings in more than 100 languages for retrieval over policies, procedures and product terms, with long-context answers from the Qwen models.
Polish-language work
Bielik, built in Poland by SpeakLeash and ACK Cyfronet AGH, handles Polish text, alongside Qwen models trained on 201 languages.
Engineering assistants
Qwen3-Coder and Devstral serve as code and documentation assistants for core-banking teams, without code leaving the bank.
What your ICT risk team will see.
Running inference inside the bank changes the question from “what does the vendor do with our data?” to “what do our own logs show?”
Data flow inside your network
Gateway, compute servers and models run on your infrastructure. Large models can span several machines over standard Ethernet, all inside your network.
Logs without content
Every request carries a correlation id that follows it across servers. Prompt fields are redacted by default, and generated text is not logged.
Your identity provider
API keys through Keycloak, or your own identity and billing systems through HTTP hooks. Per-request token usage supports internal charge-back.
A real exit plan
Open-weight models with published licences, behind OpenAI- and Anthropic-compatible APIs. The same models and code run elsewhere if you ever need to leave.
Evidence before go-live
ARK's API test suite runs on your deployment and writes standard JUnit results for your change board.
Operational resilience
With a second copy of each part of the model, a failed server doesn't take the service down: a request in flight is replayed on another path. Servers can join or leave while ARK runs, so maintenance and growth need no downtime, and your resilience tests run on infrastructure you control.
The rules a bank's AI works under.
Running inference on your own infrastructure keeps the data flow, the logs and the exit plan under your control. Your compliance assessment stays yours.
Under the EU AI Act, creditworthiness assessment and credit scoring of natural persons are Annex III high-risk uses (fraud detection is excluded). Those obligations apply from 2 December 2027 under Regulation (EU) 2026/1744. Transparency obligations have applied since 2 August 2026; content marking for generative systems already on the market applies from 2 December 2026.
Two ways to run ARK. One API.
ARK Core For banks
The platform and its models on your NVIDIA GPUs, licensed per GPU and run by your team. Our engineers can size it with you, prove it on your hardware and support the rollout.
ARK CoreOur EU infrastructureARK Cloud
Prototype with synthetic or public data on the hosted API, then move in-house without code changes.
ARK CloudStart with one workload.
Bring one use case and a sample of documents or calls. We run it on ARK, on your GPUs or in our EU cloud, and show you the numbers before you commit.