On July 17, 2026, Sichip Technology unveiled its AGX Agentic Compute strategy in Shanghai, built on three pillars: proprietary high-performance agentic CPUs, end-edge-cloud collaboration, and end-to-end software-hardware integration. The company introduced the AGX Station, a desktop-grade supercomputer capable of local inference for large models with 70B to 150B parameters, delivering a total computing power range of 160–320 TOPS. Founded in 2021, Sichip Technology’s P1 chip, initially researched five years ago, now aligns with agentic computing demands and is being developed in partnership with ecosystem collaborators such as Lenovo and Arm to build an open computing foundation. Amid intensifying global competition in AI chips, Chinese chip manufacturers are redefining agentic computing at the architectural level.

Article author and source: AI World

July 17, 2026, Shanghai, by the Huangpu River, the summer wind is scorching.

The World Artificial Intelligence Conference 2026 (WAIC 2026) will be simultaneously held across three venues in Shanghai—Expo, Zhangjiang, and West Bund—featuring four exhibition halls: over 100,000 square meters of exhibition space, more than 1,100 companies, over 3,000 exhibits, and more than 300 global debut products, with nine Turing Award and Nobel Prize laureates in attendance.

The scale of this event is so immense that even "unprecedented" feels understated.

On the same morning as the opening of WAIC, we noticed a domestic chip company accomplished something significant.

Cixin Technology, a company founded in 2021 and focused on developing proprietary high-performance intelligent agent CPUs, held a launch event in Shanghai themed "Unleashing Infinite Potential, Igniting a New Era."

This time, they officially launched the AGX Agentic Compute strategy—a unified approach integrating chips, hardware systems, and operating systems to build a comprehensive computing foundation spanning end devices, edge, and cloud.

This move is perfectly timed.

The wave of domestic chip adaptation sparked by DeepSeek continues to gain momentum, as OpenAI and Anthropic have both officially announced their self-developed chips, reshaping the global AI computing landscape.

On the edge side, a deeper transformation is underway—AI is no longer just about "chatting"; it is beginning to truly "get things done." AI agents are moving from concept to reality, and the focus of computing power demand is rapidly shifting from "training" to "inference and execution."

The two curves collide, revealing a problem that has been collectively ignored—on which type of chip should agents run?

How impressive is the answer? On stage at the launch event, a machine the size of an office desk was brought up—AGX Station. With the appropriate AI accelerator cards installed, it can perform local inference on large models ranging from 70B to 150B parameters, and multiple units can be linked together to form a desktop-scale computing cluster.

The new currency for computing power is called Token.

This morning, Sun Wenjian, founder and CEO of Cixin Technology, took the stage and immediately made a bold assertion:

We are at the dawn of a paradigm shift from "generating content" to "getting work done," and this chip technology will be the core driver of this transformation.

This might sound a bit bold at first, but if you follow its logic, it holds up.

Over the past few years, AI computing power has revolved around scaling laws, with all focus placed on "training"—training is an arms race, but it only happens once.

The model has finished training; the real cost comes next with the non-stop, 24/7 inference bills.

But in 2026, the tide turned. AI agents are no longer just visions on PowerPoint slides—they’ve begun actively breaking down tasks, invoking tools, and executing workflows. Gartner predicts that by 2028, about one-third of enterprise software will embed agent capabilities.

Thus, the standard for measuring computing power has shifted from peak computing power to task completion efficiency, and tokens have become the new "currency" of computing power value in the agent era.

Why 2026 specifically? Sun Wenjian provided four reasons:

• The iteration cycle for large models has been shortened to "every few weeks," and their capabilities are hitting a critical threshold;
• Toolchains and frameworks have lowered the barrier to entry, drawing a surge of developers.
• Capital recognition agents represent another core赛道 following large models;
• And the most critical business loop has been successfully validated over the past six months.

He casually cited an example from this chip itself:

The chip was released the year before last, and only last year did we, along with a group of partners, complete the underlying hardware setup—“At that time, there weren’t even agents yet.”

When agents truly take off, they look back and realize their hardware is "perfectly suited for agents." This foreshadowing has now come to fruition, and the opportunity belongs to those who were prepared.

Three Pillars: Starting with a Single CPU

Build a comprehensive computing power network

In Sun Wenjian’s view, agents are no longer passive knowledge bases that respond to queries, but rather “task executors” that can proactively break down goals, invoke tools, and carry out tasks.

For it to truly scale and be widely adopted, it must overcome three challenges: stability, affordability, and sustainability.

Based on this insight, Zhixing Technology has pioneered the development of an agent-native integrated platform for the AGI era.

At the press conference, Sun Wenjian officially launched the AGX Agentic Compute strategy.

Centered on the three pillars of "stable performance, affordability, and sustainability," this chip technology company has established its three strategic pillars.

First, focus on agent-specific CPUs.

This is the strongest foundation of the entire strategy.

An agent's workload is entirely different from that of a chatbot—it must think continuously, frequently invoke tools, and maintain memory across long contexts. The CPU is no longer just a "supporting player," but the central hub for task orchestration and execution.

Therefore, the choice for this chip is to redefine a CPU from the ground up for agents, rather than settling for off-the-shelf general-purpose chips.

Our self-developed Cixin P1 is this foundational cornerstone.

Sun Wenjian revealed a detail on stage: when defining P1 five years ago, they combined a high-performance CPU, a highly compatible GPU, and a "forward-looking" NPU.

Five years ago, we approached this chip with uncertainty. Five years later, this 6-nanometer chip has been adopted across industries, turning our uncertainty into confidence.

Second, achieve end-edge-cloud collaboration across the entire domain.

Agents won't live only in the cloud, nor will they live only on the device.

A real business process might, one moment, decompose a diagram locally, the next moment invoke a large cloud-based model to enhance reasoning, and then immediately return to edge execution. Once computing power is fragmented across different architectures and ecosystems, collaboration becomes an empty phrase.

The solution of this chip is its long-standing "one chip, multiple uses" approach: a high-performance intelligent agent CPU paired with different operating systems, large models, and software to serve four major computing scenarios—consumer, industrial, automotive, and edge.

The next step is the broader vision of this strategy: building a comprehensive computing infrastructure covering end, edge, and cloud scenarios, centered on our self-developed P1 chip, to truly achieve “one chip driving all AGI scenarios.”

Third, establish a full-chain hardware and software layout.

From the chip level all the way up to the entire device, operating system, and application ecosystem—leaving no blind spots.

The focus of this layer is our proprietary Agentic OS—enabling agents to run in isolated sandboxes while ensuring both efficiency and security; simultaneously, we deliver industry-standard toolchains and comprehensive solutions for end-to-end deployment.

The culmination of this pipeline is "maintaining an open ecosystem."

Sun Wenjian described this as the company’s consistent strategy: opening up the underlying software, the upper-layer operating system, and the pre-adapted large models. More importantly, he emphasized partnering with endpoint computing solution providers that focus solely on GPUs or NPUs to jointly develop open endpoint and edge AI computing solutions.

Supported by these three pillars is a more forward-looking vision: Cixin Technology will uphold the principles of "openness, empowerment, and coexistence," enabling intelligent agents to truly enter every scenario of the real world.

On-site, Sun Wenjian demonstrated the Lenovo AI Mini host built on the Cixin P1.

He referred to it as a "milestone product" and the first tangible outcome of the AGX strategy.

This chip is not fighting alone.

As a key global computing platform partner of this chip technology company, Parag Beeraka, Senior Director of Edge AI Consumer Computing at Arm, delivered a welcome address via online video link:

AI is transitioning from the era of simple question-and-answer interactions to the era of agent-based computing, a shift that relies on efficient system-level collaboration between CPUs and various accelerators.

By combining the Arm computing platform with CIX's innovation capabilities, we empower developers and ecosystem partners to unlock greater innovation potential and drive widespread adoption of applications through enhanced AI capabilities.

In the age of agents, a new computing "heart" is needed.

With the strategy covered, it's now the product's turn.

CMO Danny Zhang of Cixin Technology officially launched the AGX Agentic Compute product line and provided a definition for AGX:

The first heterogeneous computing platform designed from the ground up specifically for agent computing.

The "X" in AGX represents X Possibilities, X Accelerator, and X Future Matrix.

It offers three core advantages: unparalleled performance and energy efficiency, end-edge-cloud collaboration, and native security and trustworthiness.

These three points correspond exactly to the three propositions mentioned earlier.

What enables affordability is a bold claim made by this chip itself: the only domestic SoC+NPU integrated solution that reduces energy consumption per token by over 50%. The logic is straightforward—if tokens are a new form of currency, then every joule saved is real money.

The first physical product of the AGX platform is the desktop supercomputer—AGX Station.

A box that fits on a desk, capable of running native inference for large models with 70B to 150B parameters, enabling parallel computing and intelligent scheduling for multi-agent systems via a distributed architecture, and allowing the creation of a desktop cluster through 2×10G RDMA inter-machine cascading.

The total computing power expansion range covers 160 TOPS to 320 TOPS.

But its most critical feature is not performance—it's the interface.

A modular, scalable open hardware architecture supporting various AI computing card form factors such as M.2, MXM, and PCIe, with deep optimization for multiple domestic AI chips including TianShu, HouMo, YuanLi, ZhiChen, and GuangYu, while also compatible with half-height flagship GPUs.

In other words: Use whichever card works best, and run whichever model is most suitable.

And the philosophy Danny Zhang put forward is where AGX's true ambition lies:

The vision of Cixin Technology is to build an open computing foundation. Through the Agentic Open Compute philosophy, we break down barriers between software and hardware, enabling diverse AI ecosystems, GPUs, and VPUs to collaborate efficiently on our platform and unlock their full potential.

The implication is clear: currently, domestic AI chips are each excellent but isolated. AGX aims to insert these fragments into the same slot and bind them together as one.

Openness is the core design philosophy of this strategy.

At the hardware level, this openness is carried out to an almost "defenseless" extent.

Chu Ranzhou, Co-founder and Vice President of System Engineering at Cixin Technology, said that the AGX Station is a mass-producible reference product, commonly known in the industry as a turnkey solution. But Cixin doesn’t just hand over the keys—it shares all design details, materials, and testing documentation with customers, partners, and developers.

He asked his head of hardware: “Are we worried competitors will copy us when we hand over everything so thoroughly?”

They replied with eight characters:

The design is fine, but the technology is lacking.

Over the past two years, this chip has onboarded a large number of domestic and international component and module suppliers, and he hopes more domestic manufacturers will join.

Chu Ranzhou concluded by saying that previously, when discussing smart cities, people always talked about cold machines and networks—lifeless, without the power of life. But the future will be different—there will be digital residents, or AI residents.

He may have his own ID, as well as a physical and virtual address. A smart city only truly comes to life when its residents are powered by AI.

What this chip is doing is creating a home for these future AI residents.

This is their home, which could be an apartment or a villa.

You have the house, but you still need furniture; you have the hardware, but you still need software.

Based on this, Chen Guoyin, Senior Director of Cixi Technology, provided an in-depth analysis of the core value of the AGX OS software solution.

He pointed out that the large-scale deployment of agents requires a "secure and controllable" software environment.

Guided by the philosophy of "hardware-software integration, flexible deployment, and full autonomy," AGX OS directly addresses the industry's three major pain points:

• Protect data: Local inference and storage ensure core data "stays within the domain," mitigating compliance risks at the source.
• Break down barriers: Unified model gateway to end vendor lock-in, enabling seamless switching between thousands of models and giving customers ultimate choice.
• Risk Control: Built-in independent sandbox, four-level permissions, and end-to-end audit trail ensure that every action taken by the agent is “plannable, reviewable, approvable, traceable, and reversible.”

Through this "transparency" execution mechanism, this chip technology company is making every autonomous decision secure and trustworthy, safeguarding the practical application of AI.

Ecosystem Collaboration: From Cool to UnionPay, deploying intelligent agents isn't a one-person job.

Even the strongest chip can't carry the whole performance alone.

In the second half of the press conference, ecosystem partners including Leku Technology, Houmo Intelligence, TianShu Intelligent Chip, and UnionPay Business sequentially took the stage, showcasing their collaborative implementation plans with this chip technology from perspectives such as smart home, edge-side acceleration, computing power accessibility, and payment security.

One number mentioned by Houmo Intelligence is noteworthy: the computational power required for continuous reasoning in complex tasks during the Agent era will increase by 1,000 times—this also explains why this chip company is building an open platform to enable more AI accelerator manufacturers to join and share the burden.

At the conclusion of the press event, thisX Technology signed collective agreements with over ten partners, including China Telecom Digital Intelligence, Guanghong Communications, and Tiantong Jingdian, covering multiple areas such as AGX solutions, expanded computing power, and scientific intelligent agents.

From chips to complete devices, from operating systems to application scenarios, a complete industrial chain is taking shape.

The Age of Intelligent Agents with China's Core Technology

Looking back, the timing of thischip technology’s release of the AGX Agentic Compute at the opening of WAIC was exceptionally precise.

On one hand, the theme of WAIC 2026 itself is “Intelligent Partners, Co-Creating the Future”—AI evolving from a tool into a partner, which perfectly aligns with the story thischip tells: agents evolving from passive knowledge repositories into proactive “digital employees” that break down goals and execute tasks.

On the other hand, the wave of domestic chip substitution entered a new phase of acceleration in 2026. Just after news emerged that DeepSeek developed its own inference chip, OpenAI released its custom chip Jalapeño, and Anthropic is in discussions with Samsung to explore in-house solutions—global AI computing power is undergoing a profound reshaping.

In this tide of change, a Chinese chip company dares to redefine "agent computing" at the architectural level—extending from SoC to full systems, from hardware to operating systems, and from single chips to holistic collaboration—making this achievement worthy of serious attention.

Of course, strategy is one thing, but execution is another. Can AGX Station truly operate in enterprise scenarios? Can Agentic OS withstand real-world business challenges? Can the open ecosystem move beyond signing agreements to achieve deep collaboration? These questions all require time to verify.

But at the very least, this company has already set up the board.

This AGX city has just broken ground. Perhaps China’s indigenous chip-powered agent era is setting sail today from the shores of Shanghai.