When large models move from the cloud to edge devices such as phones, cars, and robots, a hidden but fatal bottleneck emerges: data movement consumes most of the computing power, not the actual inference computation. The latest strategic cooperation between Toread's subsidiary Shanghai Tongtu Semiconductor and Horizon Robotics precisely addresses this industry pain point—large model bandwidth compression technology has officially moved from the research phase to industrial implementation.
Technology Breakthrough: Three-Layer Compression System Targets Memory Wall
Tongtu's full-stack compression system covers the complete data chain of model inference. At the input layer, for high-bandwidth data like video, the compression ratio can reach 5x, significantly reducing the data intake pressure on edge devices. At the intermediate layer, for data such as KV Cache generated during large model inference, Tongtu's technology can achieve an additional 50% compression on top of Google's similar achievements, effectively alleviating the memory wall problem during inference. At the parameter layer, models with hundreds of billions or even trillions of parameters can be compressed by 30%, without the need for retraining—plug and play.
This technical path is not an isolated case. At the 2026 Zhongguancun Forum, Toread Group Vice President Wang Hongjian pointed out that Tongtu's approach is highly consistent with Google's TurboQuant algorithm in terms of core pain points and industrial direction, jointly verifying that "using compression to liberate computing power" is a deterministic direction for the industry. This means bandwidth compression is becoming a consensus technology at the edge AI infrastructure level, not just a proprietary trick of one company.
Industrial Landing: Horizon Ecosystem Provides Testing Ground
Choosing Horizon as a joint development partner has clear industrial logic. As a leading domestic supplier of automotive-grade AI chips, Horizon achieved full-year revenue of 3.758 billion yuan in 2025, up 57.7% year-on-year, and released the cockpit-drive fusion chip "Starry 6P" and the vehicle intelligent operating system "Kaka Xia" in 2026. Its chips are already deployed in a large number of mass-produced vehicles, providing real application scenarios for edge AI.
Tongtu's compression technology, deeply integrated with Horizon's computing platform, will directly optimize the inference bandwidth and memory usage of large models on automotive-grade chips. For intelligent driving, this means lower latency, higher frame rates, and lower hardware costs—under the backdrop of slowing performance improvements in automotive-grade chips, software-level compression optimization becomes a more cost-effective choice.
Performance Verification: Chip Business as the First Growth Curve at High Speed
The confidence in technology layout is ultimately supported by performance data. Toread's 2026 Q1 report shows that the chip business achieved revenue of 179 million yuan, a staggering 206.92% year-on-year increase, accounting for 36.09% of total revenue. This growth rate far exceeds the outdoor business's 8.02%, making chips the official first engine driving the company's growth.
Positive signals also come from institutional investors. Huaxi Securities maintains a "buy" rating, forecasting net profits of 237 million, 369 million, and 454 million yuan for 2025-2027 respectively. CICC raised its 2026 revenue forecast by 29% to 1.952 billion yuan and its EPS forecast by 60% to 0.15 yuan. For a company transitioning from an outdoor brand, such a change in performance structure deserves attention from the entire industry chain.
Strategic Depth: Dual-Business Synergy Accelerates Full-Chain Self-Sufficiency
Since establishing the "outdoor + chip" dual-business strategy in 2021, Toread has acquired five chip-related companies, building a full-chain chip product matrix covering perception, interaction, and display. Tongtu's full-stack compression technology is not only an independent business growth pole but also provides core technical support for the company's smart equipment business. During the 618 shopping festival, its self-developed smart exoskeleton Crest C3 sold over 100 units in a single day on Tmall, weighing only 1.8 kg with a power-to-weight ratio of 10:1.
Full-chain self-sufficiency from underlying chips to end products is accelerating. Wang Hongjian predicted at the Zhongguancun Forum that China's edge AI market size is expected to grow from 600 billion yuan in 2026 to 2.1 trillion yuan by 2030. Tongtu's full-stack compression capability can be applied to eight major scenarios including personal devices, smart spaces, smart vehicles, industrial internet, embodied robots, and AR/VR wearables.