A Chinese processor startup just raised nearly $150 million to build chips specifically designed for AI that runs locally on devices, not in the cloud. This funding milestone signals a fundamental shift in how companies are thinking about where artificial intelligence actually happens. Instead of sending data to distant servers, more organizations want their AI to stay put, processing information right where it's needed.

What's Driving the Move Away From Cloud AI?

CIX Technology, founded in 2021 by semiconductor veterans from AMD, Microsoft, Apple, and Meta, just completed a Series B funding round worth nearly 1 billion RMB (roughly $140 to $150 million USD). The Shanghai IC Fund and Pudong Venture Capital led the round, with existing investors including Lenovo Capital increasing their stakes.

The company's timing reflects a broader industry realization: cloud AI isn't always the right answer. Organizations are choosing local, on-device inference for several practical reasons:

• Cost Savings: Running AI locally eliminates the continuous cloud computing bills that accumulate when processing happens on remote servers.
• Data Privacy: Sensitive information stays on customer-owned infrastructure rather than traveling to and residing in cloud environments outside their control.
• Latency: Local processing eliminates network delays, enabling real-time responses critical for robotics, autonomous systems, and interactive applications.
• Offline Operation: Devices can continue functioning without internet connectivity, essential for remote locations or disconnected environments.
• Model Flexibility: Organizations can deploy custom or specialized AI models tailored to their specific needs rather than relying on one-size-fits-all cloud services.

How Is CIX Building Processors for Local AI?

CIX's approach centers on what the company calls "agent CPUs," which are heterogeneous processors combining multiple types of computing components optimized for AI workloads that run locally. The company's first production chip, the P1, launched in 2024 as a 6-nanometer system-on-chip combining a 12-core Arm processor with a GPU and a specialized neural processing unit (NPU). The P1 delivers up to 45 TOPS (trillion operations per second) of aggregate AI performance and supports up to 64GB of shared memory.

In May 2026, Lenovo launched two commercial products built around CIX's silicon: the AI Host P7 and AI Host mini. The flagship P7 is a compact, portable device designed specifically for running large AI models locally. Lenovo and CIX claim the complete P7 system can deliver up to 190 TOPS of local AI performance, support up to 80GB of memory, maintain a 128,000-token context window (roughly 100,000 words), and run locally deployed models with as many as 122 billion parameters at a maximum power consumption of 30 watts. The system achieves inference speeds of up to 50 tokens per second, meaning it can generate roughly 50 words of output every second.

CIX has introduced a three-tier product roadmap under the ClawCore family name, each targeting different deployment scenarios. The ClawCore-P version builds on the company's high-performance architecture for demanding applications. ClawCore-A targets edge inference and systems requiring additional accelerator expansion. ClawCore-E is designed for always-available endpoint devices with minimal power consumption.

Why Does Lenovo's Partnership Matter?

Getting a processor into a commercially launched product from one of the world's largest PC manufacturers represents a critical validation milestone for a chip startup. Lenovo is not just a customer; the company is also an investor in CIX through Lenovo Capital. This dual relationship gives CIX something most processor startups struggle to obtain: real distribution channels and access to volume manufacturing.

CIX's founding team brings decades of experience taking complex processors from design through mass production. CEO Sun Wenjian previously headed AMD's custom silicon activities in China and worked on Xbox CPU development at Microsoft. Co-founder and Chief Technology Officer Liu Fang spent over two decades in Silicon Valley, including roles as a core architect on multiple generations of Apple's A-series and M-series processors. This combination of expertise in both chip architecture and customer deployment distinguishes CIX from the hundreds of other AI processor startups competing for attention.

What Security Challenges Come With Local AI?

Moving AI from cloud environments to customer-owned devices introduces new security responsibilities that organizations must address. When AI models, credentials, and sensitive data live on local infrastructure outside a provider's direct control, the threat landscape changes fundamentally.

Edge AI deployments face unique attack vectors that cloud-based systems don't encounter. Attackers may have physical access to devices, direct access to model files, opportunities to tamper with the data that feeds AI systems, or ways to compromise the supply chain delivering components. Unlike cloud AI, where providers handle security verification, edge AI deployments place customers in control of more of the AI stack, making them responsible for establishing trust across the entire environment.

Traditional software security controls alone are insufficient for protecting local AI systems. AI models can be influenced by untrusted content, including prompt injection attacks that alter model behavior through carefully crafted inputs. The same input may produce different outputs depending on context, making signature-based detection and other conventional security techniques less effective.

How Can Organizations Secure Local AI Systems?

Microsoft's security research identifies several architectural approaches for protecting edge AI deployments:

• Runtime Attestation: Verify that the hardware and software environment where AI runs matches an approved baseline before releasing sensitive model weights, credentials, or data to that system.
• Artifact Provenance: Establish trust in the components that shape AI behavior, including model weights, tool descriptors, agent definitions, and retrieval indexes, by verifying the systems in which they were built and delivered.
• Deterministic Mediation: Implement a logical boundary outside the AI model that enforces policy by allowlisting actions, scoping arguments, limiting frequency, and releasing credentials only when approved, rather than allowing the model to authorize actions directly.
• Confidential Computing: Use hardware-protected memory and attestation to prevent privileged processes or accelerators from reading or modifying decrypted model weights, credentials, and data.

These approaches work together to answer two critical questions before releasing sensitive assets to an edge AI system: Do I trust this runtime and the platform where the workload will execute? Do I trust the components such as model weights and tool descriptors because I trust the system in which they were built and delivered?

The shift toward local AI processing represents a fundamental rethinking of where intelligence happens in computing systems. As companies like CIX demonstrate that local inference can deliver competitive performance while offering cost, privacy, and latency advantages, the infrastructure supporting AI is becoming increasingly distributed. The challenge now is ensuring that this distributed intelligence remains secure and trustworthy as it moves beyond the controlled environments of cloud data centers into customer-owned infrastructure.