EN / 中文

Biren Backs Anxin Tech: Why Embodied Intelligence Needs Deterministic Edge Chips, Not GPUs

by shishuoxinyu·November 16, 2026

Shanghai Anxin Technology Co., Ltd. recently completed a change in its industrial and commercial shareholder registration, with Zhuhai Biren Integrated Circuit Co., Ltd. entering the shareholder list as a new investor. Public records indicate that Zhuhai Biren Integrated Circuit Co., Ltd. is a wholly-owned subsidiary of Shanghai Biren Technology Co., Ltd., which listed on the Hong Kong Stock Exchange in 2026.

Anxin Technology was established in December 2025, with Sun Erjun as the actual controller. Its publicly stated focus involves the integration of NPU architecture, edge-side AI inference models, and precision motion control. The company has been in existence for less than a year; to date, the author has not collected any public information regarding products, tape-outs, or customers, leaving no basis for evaluating its technical capabilities.

However, given its shareholder background (including Biren and Hillhouse) and its publicly disclosed technical direction, capital appears to be betting on the route of "dedicated edge-side chips for embodied intelligence." Since there are no products yet, rather than discussing technical strength, it is more appropriate to step back and analyze this specific technological path.

To understand this path clearly, one must first address a more fundamental question: Why can't GPUs, which dominate data centers, handle robot tasks effectively?

| GPUs Lack Timing Guarantees

The notion that GPUs are designed for throughput and are therefore "uneconomical" for robotics is easily refuted—when shipment volumes are sufficiently high, GPU manufacturers can customize and trim designs to amortize costs. The real issue lies not in economics, but in functionality.

When AMD launched the Ryzen AI X100 for robotics in 2026, Salil Raje, head of AMD's embedded business, emphasized in official statements that robot manufacturers prioritize "precise execution of computational results within required time constraints," rather than peak AI computing power. GPUs operate on a time-slicing model where resources are queued, and memory synchronization also requires waiting, leading to variable inference latency at the edge. This irregular rhythm cannot drive motion loops that require stable control cycles. This is not an issue of insufficient computing power, but rather that an architecture designed for "average throughput" was never intended to target "predictability even in worst-case scenarios."

From this, we can infer that training compute and embodied intelligence compute are diverging along different functional goals. The former continues to pursue scale and throughput, while the latter prioritizes determinism, low power consumption, and hardware-software integration.

| Both Paths Are Difficult

Computing for embodied intelligence must simultaneously serve two time scales. The motion control side requires control instructions to arrive within predictable cycles; the inference side, such as visual perception and large model inference, is essentially a throughput problem suitable for NPUs. The difficulty lies not in having both, but in ensuring they coexist without interfering with each other.

For example, SemiDrive's D9-MAX uses a hard isolation design, placing motion control and interaction functions in different regions of the same chip, operating independently without mutual interference: dedicated real-time cores run real-time control and communication tasks, while the NPU handles perceptual inference. Two generations of this series have already achieved mass production with leading humanoid robot manufacturers. The engineering threshold for this route of "integrating different functional cores into a single chip" is not low—it is difficult for new companies to catch up quickly in terms of mass production validation, functional safety, and the maturity of the real-time software stack.

If Anxin follows this path, the challenge is to prove that a more integrated chip brings system-level benefits compared to existing mass-produced solutions. If it aims to optimize the NPU architecture itself to better serve deterministic tasks, that would constitute a technical hypothesis distinct from existing chip solutions at the architectural level. It should be noted that this distinction between possibilities is used only to illustrate key dimensions for evaluating this path and does not constitute a prediction of Anxin's actual technical solution.

| Half the Cost Is Underestimated

Anxin listing "edge-side AI inference models" as a self-developed direction implies collaborative design between algorithms and chips. A common cost is that "model iterations take months, while chip tape-outs take years," posing a risk that hardware may be left behind by the next generation of versions.

Research in2026 shows clear divergence: some studies split models into asynchronous "slow thinking" and "fast action" components, while others insist on end-to-end processing in a single pass. Empirical results from the former category show that if run synchronously, large models slow down control frequency—as the model size increases from 0.8B to 8B, control frequency drops from 9.7 Hz to 4.1 Hz; after splitting the two, control frequency is determined solely by the action component, no longer dragged down by model size, stabilizing at 20 Hz. This indicates that different approaches have indeed different hardware requirements: some rely more on task scheduling and intra-chip communication, while others depend more on the computing power and bandwidth of a single large core.

If Anxin performs collaborative optimization for a specific structure, it faces not only the risk of being overtaken by "next-generation versions" but also the risk of technological path divergence. The verification cycle and uncertainty of this trade-off are far greater than those implied by "fast models, slow chips."

| The Market Looks at Replacement Costs

Replacing motion control chips is not simply a matter of re-running models. It involves re-establishing the real-time software stack, communication protocols, functional safety certifications, and field reliability records. Customers are not unwilling to give new companies a chance, but the hidden costs are too high—so high that it is only worth initiating a round of recertification when the chip itself offers significant material cost savings or performance improvements.

Having the right architectural direction is merely an entry ticket. Building trust requires answers from products, customers, and mass production records. Until then, what can be confirmed is that dedicated edge-side chips for embodied intelligence have moved from academic topics to engineering problems, but there remains a line drawn by determinism, reliability, and replacement costs between "being noticed by capital" and "being adopted by customers."

 

Editor: Shuo Xin Yu - Xiaoming Ba