# Sudu Tech: The Third Position of a Horizontal Cut

Sudu Tech: The third position of a horizontal cut

Full article: https://haiguangboy.com/posts/sudu-third-position

## Core Approach

Sudu is the third kind, and the difference isn't about "general vs. vertical."

## Key Results

It makes a **horizontal cut** through the tech stack: the upper layer (LLM/Agent that understands instructions, breaks down steps, calls external knowledge) is explicitly not pursued, acknowledging "that's a capability familiar to big model companies"; the lower layer (grasping, placing, precise assembly, plugging, screwing) gets all resources, but this layer is required to be universal across all industries.

Han Zheng's analogy for himself is "an SDK or API on iOS or Android."

This sentence deserves a pause. Qianxun also used a phone analogy when discussing general use, but there it was about "hardware and operating systems being universal, with various industries installing apps on top"—**doing the OS**. Sudu is about interfaces **above the OS** that others call. So Sudu hopes that those building upper-layer models and vertical industrial applications become its customers, not competitors. The same analogy lands on two different layers.

This also explains Sudu's entry into CATL's production line—resembling Lingchu's vertical play, but with a completely different purpose.

Lingchu treats vertical scenarios as **commercial endpoints**: making the economics of one industry work. Sudu treats them as **testbeds**: battery module models change, cells differ, processes are rearranged, but the underlying actions remain grasping, precise placement, insertion, and assembly. What needs validation is whether the same skill set can transfer from one battery model to another, then to automotive and 3C. So Han Zheng says "not just doing single-scenario delivery," even saying "we're not inclined to choose short-term revenue"—most companies in our database are telling stories with order numbers, but this one is actively downplaying them.

## Related Route Comparisons

- Opposite route · [wx_界面新闻_20260605](https://haiguangboy.com/posts/wx_界面新闻_20260605) `wx_界面新闻_20260605_2026_06`: ★Core positioning: Splitting general physical intelligence into upper and lower layers, Sudu only does lower-layer execution, explicitly not doing upper-layer LLM/Agent contradicts Route bet: Doing the brain, not the body—betting on the foundation model company identity, rejecting the motion capability track
- Opposite route · [Qianxun Intelligent Han Fengtao: Embodied Intelligence Enters the 'Speed Competition' Phase](https://haiguangboy.com/posts/qianxun-general-brain-data-loop) `wx_中国财富_20260719_2026_07`: ★★Route positioning: Sudu is the 'horizontal tech stack cut' third position, neither general brain nor vertical industry brain contradicts ★Route bet: Betting on the general brain over the vertical industry brain—'the only opportunity in the big model era is the general model'
- Opposite route · [Robots Begin to 'Stand in the Light': Lingchu Intelligent Enters Optical Module Production Lines](https://haiguangboy.com/posts/lingchu-optical) `wx_星河频率_20260718_2026_07`: Precise articulation of simulation-first training route: almost no real robot data in pretraining, but post-deployment failure retries plus a small amount of real robot RL contradicts Route bet: Inverted data pyramid—human data as the base rather than internet video as the base
- Same route · [An_Open_Foundation_Model_Towards](https://haiguangboy.com/posts/an_open_foundation_model_towards) `an_open_foundation_model_towards_2026_07`: ★Core positioning: Splitting general physical intelligence into upper and lower layers, Sudu only does lower-layer execution, explicitly not doing upper-layer LLM/Agent validates Joint co-training of heterogeneous human-robot data is a structurally suboptimal solution
- Same route · [LAWM: Why Action Labels Become a Burden](https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling) `latent_action_pretraining_through_world_modeling_2026_07`: LAWM: Why action labels become a burden
- Same route · [T-Rex: Why Tactile Sensing Needs Separate Modeling](https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation) `t_rex_tactile_reactive_dexterous_manipulation_2026_07`: T-Rex: Why tactile sensing needs separate modeling

## Boundaries

Valued at 20 billion yuan in one year, but **zero rounds, zero amounts, zero investor lists**—this transparency is among the lowest tier of Chinese companies we track. The first CATL workstation POC relies solely on the company's unilateral claims, with no confirmation from the other party. "99%+ reliability" doesn't specify which tasks, who tested, or how it was calculated—and since this standard is also used as an argument for "we're not a bubble," if the standard itself isn't measurable, the claim can't be verified either.

## Author's Judgment (Not Paper Content, Cross-Paper Synthesis)

💡 Where the real divergence lies

Sudu's simulation route collides hard with several companies in our database: Lingchu's "human data as the base" and Genesis's "simulation only for evaluation, not training data" are both 3/3 strong clashes.

But what's more worth noting is its relationship with Xu Huazhe—not opposition, but **limitation**. Xu Huazhe argues that "the action part must be a unified model, opposing modular splicing," and Sudu doesn't deny this, just locks it into the lower layer: the action layer must be internally unified, but the action layer and the planning layer can—and should—be separated.

This actually answers a question we've tracked for a long time: does the value of long-horizon tasks lie in orchestration or in atomic capabilities? Sudu doesn't try to make the scheduling layer smarter—acknowledging the upper layer can be slow, can call external knowledge, or even have factory-defined processes manually—and bets everything on the reliability and transferability of atomic skills.

Betting on atomic capabilities. This is a clear stance, and one that can be falsified.