When up to 70% of computing power in large model inference is consumed by data movement rather than computation, the industry finally acknowledges: the bottleneck is not peak chip performance, but bandwidth and memory. Toread's subsidiary Shanghai Tongtu has signed a strategic agreement with Horizon Robotics to integrate full-stack data compression technology into Horizon's automotive-grade chip platforms, directly targeting the most stubborn 'memory wall' in edge AI deployment.
Technology Pathway: Three-Layer Compression System Unblocks Edge Bottlenecks
Tongtu's compression solution covers the entire large model inference pipeline. The first layer targets input data such as video, compressing it to one-fifth of its original size, directly reducing bandwidth pressure on edge devices. The second layer handles intermediate state data (e.g., KV Cache), achieving an additional 50% compression beyond Google's TurboQuant algorithm, alleviating memory expansion during inference. The third layer compresses model weight parameters by 30% losslessly, with no retraining required and plug-and-play deployment.
This technical approach aligns closely with Google's contemporary TurboQuant algorithm in core objectives, both pointing to the deterministic direction of 'liberating computing power through compression.' However, Tongtu's advantage lies in systemic integration—embedding three compression layers into a single inference framework rather than optimizing each in isolation. For downstream manufacturers, this means gaining over 30% effective computing power release without modifying existing chip architectures.
Performance Validation: Chip Business Growth Far Outpaces Outdoor Core Business
The confidence in technology is ultimately supported by financial data. In Q1 2026, Toread's chip business achieved revenue of 179 million yuan, up 206.92% year-on-year, accounting for 36.09% of total revenue. Meanwhile, the outdoor business grew only 8.02%, making chips the company's primary growth engine.
Overall, the company reported Q1 revenue of 496 million yuan, up 39.41% year-on-year, and non-GAAP net profit of 63.94 million yuan, up 32.36%. Institutional signals are also positive: CICC raised its 2026 revenue forecast by 29% to 1.952 billion yuan and EPS forecast by 60% to 0.15 yuan. Huaxi Securities maintained a 'buy' rating, forecasting net profits of 237 million, 369 million, and 454 million yuan for 2025–2027 respectively.
Notably, chip business margins are significantly higher than outdoor apparel, so its rising revenue share will directly improve the company's overall profitability structure. For investors, Toread has shifted from an 'outdoor brand' to a 'chip + outdoor' dual valuation logic.
Industrial Deployment: From Autonomous Driving to Exoskeletons, Dual-Business Synergy Takes Shape
The partnership with Horizon is not an isolated event. Since establishing its dual-business strategy in 2021, Toread has acquired five chip-related companies, building a full-chain chip matrix covering 'perception-interaction-display.' Tongtu's full-stack compression technology serves both as an independent chip business growth pole and as underlying computing power support for second-generation intelligent sports exoskeletons and outdoor hiking smart shoes.
During the 618 shopping festival, Toread's self-developed intelligent sports exoskeleton Crest C3 opened for pre-sale, achieving daily sales of over 100 units on Tmall, with a total weight of just 1.8 kg and a thrust-to-weight ratio of 10:1. The full-chain self-reliance from underlying chips to end products is accelerating. Tongtu CEO Wang Hongjian predicted at the 2026 Zhongguancun Forum that China's edge AI market will grow from 600 billion yuan in 2026 to 2.1 trillion yuan by 2030, with Tongtu's technology applicable to eight major scenarios including personal devices, smart spaces, smart vehicles, industrial internet, embodied robots, and AR/VR wearables.
For Horizon, integrating Tongtu's compression technology means its automotive-grade chip 'Starry 6P' and vehicle OS 'Kakaxia' will achieve more efficient inference, helping solidify its leading position in intelligent driving computing. The collaboration essentially aims to shorten the technology-to-product cycle through vertical integration in the face of computing bottlenecks.