DeepKernel has completed its seed round financing, with L2F Lighthouse Founders' Fund participating in the investment.

Recently, DeepKernel, a leading company in sovereign AI infrastructure, announced the completion of a multi-million RMB seed round financing. This round of financing was led by Shunwei Capital, with participation from L2F Lighthouse Founders' Fund. The funds from this round will be mainly used for the engineering implementation of the three-layer architecture of Sovereign AI, core technology R&D breakthroughs and the building of high-level technical teams.

DeepKernel focuses on the most core yet most easily overlooked issue in the AI era —— when enterprises are leveraging the best intelligence and infrastructure, can they retain full control over their own data, agent trajectories, learning loops and accumulated intelligence. DeepKernel is committed to making intelligence operate within verifiable boundaries of enterprises, and ensuring that the intelligence precipitated from real business operations belongs exclusively to the enterprises themselves.

From Owning Data to Owning Intelligence

In 2026, "Sovereign AI" has become one of the most mainstream topics in the global AI industry. In its whitepaper *Institutional Sovereignty in the AI Era*, Palantir advocates that institutions must hold on to their own alpha; Microsoft CEO Satya Nadella proposed the "Reverse Information Paradox" —— enterprises may end up paying twice for intelligence: once for calling model APIs, and the other for the proprietary knowledge they have to surrender to make the intelligence useful; Sonya Huang, Partner at Sequoia Capital, put it more directly: who should own the intelligence at the core of your business?

Despite different positions from all parties, they all point to the same concern: when intelligence becomes a core means of production, enterprises must never lose control over "how intelligence is created, improved and used".

However, Sovereign AI is often understood too narrowly —— running an open-source model locally, purchasing GPUs on your own, and locking data into a private cloud. DeepKernel believes that these are only some of the ways enterprises use AI at present, and do not equate to sovereignty. The fact that an enterprise owns its own servers does not mean that the models and learning processes running on them are also defined and controlled by itself; conversely, when an enterprise rents cloud infrastructure, as long as the computing process is verifiable and provable, it can also retain control over its data and intelligence.

Therefore, DeepKernel defines Sovereign AI as: while using the best intelligence and infrastructure, retain control over your own data, agent trajectories, learning loops and accumulated intelligence.

DeepKernel Builds a Three-layer Architecture Around Sovereign AI Infrastructure

Based on the above judgment, DeepKernel breaks down the construction of Sovereign AI into three layers.

1. Sovereign Compute Infra, which answers the question "Where does my intelligence run?"

Enterprises today are facing two extremes: calling public model APIs delivers the strongest capabilities at the lowest cost, but the intelligence provider that enterprises rely most on always stays outside the enterprise's trust boundary; full private deployment offers the strongest control, but the scale and iteration speed of cutting-edge models are making it economically unsustainable. Neither of the two extremes is feasible, so Sovereign AI requires a third architecture.

Based on cryptography, trusted chips and distributed systems, DeepKernel establishes a clear and enforceable Trusted Computing Boundary for AI training and inference infrastructure, and breaks it down into three verifiable items —— Confidentiality, which ensures data and models are not exposed to unauthorized computing power operators and other tenants; Integrity, which ensures AI training and inference are executed in accordance with expected codes, models and policies; Verifiability, which based on cryptography and hardware root of trust, realizes the measurability, auditability and provability of end-to-end computing for AI workloads.

2. Sovereign Learning Loop, which answers the question "Who benefits from the intelligence I precipitate?"

The Execution Trajectory left by agents when performing real tasks —— the context they receive, the plans they form, the tools they select, failures and human intervention —— encodes information that no general-purpose model can pre-equip: in the real operational environment of this organization, what is "Reward".

DeepKernel is building an Agentic Learning Infrastructure covering Execution, Evaluation, Learning and Orchestration, to turn real-world AI operations into enterprise-controllable Intelligence Compounding.

3. Sovereign Agentic Interoperability, which answers the question "Can my agents interact without losing control?"

When agents start to act on behalf of organizations, an increasingly large proportion of interactions will take place between agents. Agents need to understand goals, hold granted permissions, access private context, spend money and take consequential actions.

DeepKernel is building the core protocol stack of the Internet of Agents, covering key capabilities including Identity, Capability, Delegation, Accountability and Settlement for agents.

Team and Progress

DeepKernel brings together young scientific researchers from top universities at home and abroad and core engineering talents from leading technology enterprises in the industry. Team members have previously worked at companies including Google, VMware, Broadcom, ByteDance, Alibaba, Tencent and Meituan; the core R&D team has in-depth cooperation with the InspiringGroup of Tsinghua University, has published nearly a hundred papers in international Class A academic conferences and journals, and has won the Outstanding Paper Award at top international academic conferences for many times, realizing systematic innovation covering algorithms, models, systems, chips and networks, and forming a globally leading technology stack.

In terms of engineering progress, DeepKernel has built three technical routes for trusted AI Infra based on heterogeneous computing power: based on existing trusted hardware, it has completed large-scale cross-network trusted AI cluster networking, and realized end-to-end verifiable remote attestation; for domestic computing power, it has completed the capability building of trusted AI Infra around the domestic chip system; based on self-developed confidential computing chips, it redefines the chip-level trusted AI computing paradigm.

In terms of industrial implementation, DeepKernel has launched commercial delivery for enterprise customers and platform partners: for MaaS, the company builds its own trusted AI Infra, and promotes joint solutions with leading domestic intelligent computing cloud platforms, providing verifiable model inference and training services to the outside world, allowing enterprises to independently verify the trusted status of the AI computing environment and the authenticity of the called models while using elastic computing power; for PaaS, it has reached cooperation with leading domestic world model companies to build trusted AI infrastructure, enabling intelligence to form a closed loop in the physical world.

Sovereign AI: The Default Paradigm for Future AI Computing

At present, three conditions for Sovereign AI to become the next-generation AI computing paradigm are emerging simultaneously: the model scale has crossed the economic critical point of private deployment, and full self-construction is changing from "expensive" to "impossible" for the vast majority of enterprises, and faces the dilemma of being outdated as soon as it is completed; AI has begun to make consequential decisions in real business, rather than just giving suggestions, and AI needs to complete closed loops and intelligence compounding in the real world; at the same time, the supply chain of trusted computing power with hardware root of trust is gradually improving, and the foundation of trusted AI Infra is maturing.

In the next decade, every enterprise needs to answer: Where is our intelligence precipitated, and who owns it?

This is exactly the problem DeepKernel aims to solve, and its answer to Sovereign AI —— Control the Compute. Compound the Intelligence. Connect the Agents.

Ji Xing, Managing Partner of L2F Lighthouse Founders' Fund, said: "AI is evolving from providing information to undertaking decision-making, and enterprises are beginning to provide their own core knowledge and decision-making processes to AI. We judge that the key to AI competition in the next step is no longer just who can train stronger models, but who can enable enterprises to use the most powerful intelligence while still owning their own data, learning loops and intelligence compounding —— which determines who will eventually get the dividend of intelligence in the next decade. DeepKernel does not confine sovereignty to 'moving the machines back', but builds a Sovereign AI system from the three layers of computing, learning and interaction at the same time, which is exactly the key path for Sovereign AI to move from concept to implementation. The team has long-term research accumulation in trusted and verifiable computing, as well as engineering implementation capabilities for large-scale AI Infra. We look forward to accompanying DeepKernel to define intelligence sovereignty in the AI era, so that every enterprise can truly own its own intelligence."