Own Your Intelligence: Epic Semi on Sovereign Enterprise AI

Why institutions that keep data on premises need to own the data, models, hardware and software

Sovereign AI is no longer only a national program. Banks, hospitals, law firms, universities and research labs are asking the same question nations ask: Who owns the data, the model, the hardware and the software? In a 2026 BARC survey of 320 companies, 62% named the use of data and AI in core processes as a leading reason to pursue data sovereignty.¹

Helping customers answer that question is the reason Epic Semi exists. We build for organizations that want to own their intelligence, whether a nation, a company or a single team. For an enterprise, that conversation usually begins with the data.

Why the data stays home

In regulated sectors, keeping data on site is often required by law or contract. Hospitals must protect patient records under privacy laws that apply wherever those records are processed. A law firm under a legal hold has to preserve the relevant records and account for where they are kept and who has accessed them. Data residency laws in many countries require certain records to stay within national borders. In Germany and several other European countries, works councils have co-determination rights over systems that handle employee data, and moving that data to an outside service can add another round of negotiation before a rollout.

Most organizations also see local storage as the safer choice. In Cisco’s 2025 Data Privacy Benchmark Study, which drew on 2,600 privacy and security professionals in 12 countries, 90% of organizations said local storage is inherently safer, even though data localization adds operating costs.²

Partners are asking, too. More vendors want to know where their data will be stored before they share it. A supplier sending design files, a pharmaceutical partner sharing trial results or a client handing over contracts for review may ask for written agreements on where that material is stored and who can access it. An institution that keeps its AI in its own data center, on hardware it controls, can answer them directly.

Keeping data close to where it is processed also means fewer transfers. This reduces chances of accidental exposure, avoids cloud egress fees and cuts round trips for AI agents that query the same records repeatedly.

Built around the customer’s needs

Each institution sets its boundaries differently. A hospital network may want agents that summarize clinical notes without any record leaving the hospital’s own systems. A bank may need separate environments for trading, compliance and customer service, each with its own audit trail.

At Epic Semi, we start with the customer’s requirements and design the stack to meet them. Customers come to us for cooling, interconnect, CPUs, accelerators and storage delivered as a full rack, and increasingly for the orchestration, applications and operations that run on top of it.

An architecture that evolves with the work

No two enterprises run the same agent workflows. A bank, a hospital and a manufacturer each split the work differently between the model and everything around it, including retrieval, database queries, tool calls and security checks that run on general-purpose cores. Research on agentic workloads shows the slowest stage changes from one job to the next.³ No one can specify today the exact mix of general-purpose cores, AI cores and memory every enterprise will need, and that mix will shift as we learn from deployments.

We now use AI in our own chip design work, and it is shortening the time between generations. Faster cycles mean each new chip can rebalance cores and memory around the workloads customers run, instead of locking in one design for years.

Intelligence inside your walls

Epic Semi recently announced the launch of Contrail AIX, our Artificial Intelligence eXecution Platform that puts these ideas into hardware that runs inside an enterprise’s own data center. At its center is Contrail Compute 5050, a RISC-V server CPU and AI superchip that places general-purpose and AI cores on one die. Both share one memory and storage space, which reduces data movement as an agent switches between retrieval, tool calls and model execution.

For institutions focused on protecting their data, security must start in the silicon. Contrail Compute includes trusted secure boot and full-chip security management. Full hardware virtualization lets separate teams, such as a bank’s trading and compliance groups, run isolated environments on shared hardware.

Ubuntu, llama.cpp and open-weight models run on Contrail AIX today. The platform ships under a 5A992C export classification that reaches substantially all markets, so later phases of a deployment fall under the same export rules as the first.

Enterprises have spent years deciding where their data should live. The next decision is who owns the AI systems that use it. If you run AI on premises and want that control, bring us your workflows, and we will help you measure them end to end on Contrail AIX.

Endnotes

1. BARC, “Data Sovereignty 2026: Reality, Relevance, Roadmap,” online survey of 320 companies, conducted February and March 2026, published May 6, 2026, https://barc.com/news/data-sovereignty-2026-survey/.

2. Cisco, “Cisco’s 2025 Data Privacy Benchmark Study: Privacy landscape grows increasingly complex in the age of AI,” April 2, 2025, https://newsroom.cisco.com/c/r/newsroom/en/us/a/y2025/m04/cisco-2025-data-privacy-benchmark-study-privacy-landscape-grows-increasingly-complex-in-the-age-of-ai.html.

3. Ritik Raj, Souvik Kundu, Ishita Vohra, Hong Wang and Tushar Krishna, “Towards Understanding, Analyzing, and Optimizing Agentic AI Execution: A CPU-Centric Perspective,” arXiv:2511.00739, submitted Nov. 1, 2025, revised April 16, 2026, https://arxiv.org/abs/2511.00739.

‍