A bank that barely existed a year ago is now chasing a valuation most European lenders can only dream of. Erebor, founded by Palmer Luckey — the creator of Oculus and now CEO of Anduril — and backed by Peter Thiel, is in talks to raise funding at a valuation of at least $8 billion. Its deposits have nearly quadrupled in just three months. These numbers tell a story of explosive growth, but they also raise an uncomfortable question: how does a newborn bank safely manage a rapidly expanding mountain of financial data without betraying the trust of customers and regulators?

For many fintechs entering sensitive markets, the answer lies in a delicate balance between scalability and infrastructure control. On one hand, the cloud offers the ability to expand services within hours, activating nearly limitless computing resources. On the other, banking data remains subject to strict regulations like GDPR, which require knowing at all times where information resides and who can access it. Demonstrating compliance is no longer enough: the architecture must make data sovereignty a foundational block, not an afterthought.

That is why more and more players, including in the financial sector, are starting to evaluate on-premise or hybrid deployments for artificial intelligence workloads. Training and inference of Large Language Models — used for risk analysis, anti-money laundering, customer support — raise even more critical privacy concerns when handling bank statements, personal details, and transactions. Bringing LLMs in-house, on proprietary servers, allows maintaining control over the data pipeline and reducing exposure to third-party risks. But it is not a cost-free choice: the Total Cost of Ownership (TCO) of local infrastructure must be weighed against cloud operating expenses, and it requires in-house expertise to manage training pipelines, quantization, and model serving.

The Erebor case, despite its technological opacity, is emblematic of a tension that runs through the entire sector. Growing at breakneck speed without compromising on data security is the real challenge for anyone aiming to become a major global fintech player. For those considering on-premise deployment, the trade-offs are clear: lower latency and guaranteed sovereignty come up against the need to invest in specialized hardware — GPUs with ample VRAM for inference, fast storage, low-latency networking — and to maintain a team capable of fine-tuning and maintenance. There is no universal recipe, but Erebor’s rise is a reminder that the market will reward architectures that can combine growth and control. Or at least it will try to.