When large models move from the cloud to phones, cars, and robots, a long-overlooked bottleneck emerges: the bandwidth and memory consumed by data movement are severely limiting the release of edge computing power. The strategic signing between Shanghai Tongtu Semiconductor, a subsidiary of Toread (300005.SZ), and Horizon Robotics brings this hidden pain point to the forefront. Tongtu's full-stack large model compression technology aims to 'depressurize' edge AI at the data flow source.
Technical Deconstruction: Three-Layer Compression Targets Memory Wall
Tongtu's compression system covers three key stages of large model deployment. First, the input layer: for high-capacity data like video, it can compress data to one-fifth of its original size, significantly reducing the input bandwidth pressure on edge devices. Second, the intermediate layer: for data such as KV Cache generated during model inference, Tongtu achieves an additional 50% compression on top of Google's existing results, effectively alleviating the memory wall bottleneck. Third, the model parameter layer: for billion and trillion-parameter models, it achieves 30% parameter compression without retraining, enabling plug-and-play deployment.
This technical path is highly consistent with Google's TurboQuant memory compression algorithm, pointing to the industry consensus of 'using compression to free computing power.' Tongtu's verified tests show breakthrough results in inference performance, bandwidth usage, and business effectiveness. This means that large model applications that previously required expensive computing power can now be deployed efficiently on existing hardware platforms.
Industry Landing: Commercial Validation on Horizon's Platform
Horizon Robotics, a leading domestic provider of automotive-grade AI chips, reported full-year 2025 revenue of 3.758 billion yuan, up 57.7% year-on-year. In 2026, it released the cockpit-driving integrated chip 'Starry 6P' and the vehicle intelligent operating system 'Kakaxia.' In this collaboration, Tongtu's full-stack compression technology will deeply integrate with Horizon's intelligent driving computing platform, directly optimizing inference bandwidth and memory usage for large models on automotive-grade chips.
For buyers, this partnership means the deployment cost of edge AI may see a structural decline. For example, in intelligent driving, improved inference efficiency can reduce reliance on high-computing-power chips, thereby lowering the hardware BOM cost of the vehicle's electronic/electrical architecture. For robots, AR/VR, and other scenarios, lower bandwidth requirements translate to longer battery life and smoother interaction experiences.
Performance Confirmation: Chip Business Becomes Primary Growth Curve
The confidence in technology is ultimately validated by performance data. In Q1 2026, Toread reported operating revenue of 496 million yuan, up 39.41% year-on-year; net profit attributable to shareholders was 56.4457 million yuan, up 14.44%. The chip business stood out: quarterly revenue of 179 million yuan, up 206.92% year-on-year, accounting for 36.09% of total revenue.
The chip business has become the company's primary growth driver, while the outdoor business grew 8.02% in the same period, forming a strong pattern of 'high chip growth, stable outdoor growth.' Institutional signals are also positive. On June 17, Toread received surveys from 11 institutions. Huaxi Securities maintained a 'buy' rating, forecasting net profits of 237/369/454 million yuan for 2025-2027. CICC raised its 2026 revenue forecast by 29% to 1.952 billion yuan and its EPS forecast by 60% to 0.15 yuan.
Strategic Depth: Dual-Main Business Synergy Accelerates Full-Chain Autonomy
Since establishing the 'outdoor + chip' dual-main business strategy in 2021, Toread has acquired five chip-related companies, building a chip product matrix covering the entire chain of 'sensing, interaction, and display.' Tongtu's full-stack compression technology serves both as an independent chip business growth pole and as core technical support for Toread's second-generation intelligent motion exoskeleton and outdoor smart shoes.
During the 618 shopping festival, the company's self-developed intelligent motion exoskeleton Crest C3 started pre-sales, achieving a single-day sales volume of over 100 units on Tmall. The whole unit weighs only 1.8 kg with a thrust-to-weight ratio of 10:1. Full-chain autonomous control from underlying chips to terminal products is accelerating. Wang Hongjian, Vice President of Toread Group and CEO of Shanghai Tongtu Semiconductor, predicted at the 2026 Zhongguancun Forum that the edge AI market will form a huge blue ocean, with China's edge AI market size expected to grow from 600 billion yuan in 2026 to 2.1 trillion yuan by 2030.