In 2026, the global AI industry reached a critical inflection point: capital expenditure by hyperscale cloud providers on inference surpassed that on training for the first time. This shift means the industry's focus has moved from 'building large models' to 'using large models,' with the core bottleneck evolving from compute scale to memory bandwidth and communication latency. As models like Kimi K3, DeepSeek, and GPT series continue to push parameter limits, the exponential growth in storage costs has become the biggest barrier to AI deployment. Deploying a trillion-parameter model in a single cluster now costs over one billion yuan in hardware, and the marginal returns of simply stacking GPUs are diminishing.
Structural Contradictions from Parameter Explosion
This parameter explosion has created three structural contradictions: soaring storage costs, a prominent memory wall bottleneck, and near-impossible edge deployment. A Deutsche Bank report indicates that AI high-bandwidth memory demand is expanding at a compound annual growth rate of approximately 40%, making memory chips a 'key variable determining AI progress.' The industry is pursuing two main routes: one is Nvidia's continuous upgrade of HBM bandwidth and capacity; the other is active full-stack compression to reduce data volume at the source. Toread Group's subsidiary, Shanghai Tongtu Semiconductor, has chosen the latter path.
Tongtu has built a rare domestic full-stack AI compression system covering three layers: input data, intermediate state, and model parameters. Front-end multimodal input can be compressed to one-fifth of its original size, significantly reducing pre-model transmission bandwidth pressure. The intermediate KV Cache can be further compressed by 50% based on Google's TurboQuant, alleviating the inference memory wall bottleneck. The model parameter layer can achieve 30% compression on trillion-parameter models without retraining, offering plug-and-play functionality. Tongtu CEO Wang Hongjian stated at the Zhongguancun Forum that Tongtu's full-stack compression aligns closely with Google's TurboQuant in core pain points and technical paths, confirming that 'using compression to liberate computing power' is a definitive industry direction.
Technical Strength Translates into Ecosystem Influence
In June 2026, Tongtu signed a strategic cooperation agreement with global ecosystem platform GNS, integrating core IPs for LLM weight compression and AI inference memory optimization into the GNS technical ecosystem. This marks the first time a Chinese self-developed compression IP has entered a global platform ecosystem. Subsequently, Tongtu signed a strategic agreement with a leading automotive-grade AI chip company to optimize inference bandwidth and reduce memory usage for intelligent driving computing platforms. This collaboration is a key step for edge AI to break through the 'memory wall,' advancing large-model compression toward edge industrialization at the system level.
These two strategic signings are the systematic outcome of Toread's 'outdoor + chip' dual-core strategy, established in 2021. The group has acquired five chip companies, building a full-chain chip product matrix covering sensing, interaction, display, and compression. Tongtu's core RISC-V image algorithm IP has been licensed to over 20 medium-to-large chip design companies, with gross margins exceeding 60%. Its display bridge chip shipments rank among the top in the industry. Compression IPs have been batch-imported into leading supply chains like HiSilicon and Will Semiconductor, with self-developed edge compression chips entering tape-out and mass production. Tongtu's compression capabilities can deeply synergize with Betterlife's touch chips and Beijing Xineng's display driver chips, creating a complete technical closed loop for AR glasses, smart vehicles, and humanoid robots—a rare cross-product synergy in A-share markets.
From Chips to End Products: Closing the Loop
Underlying chips are rapidly translating into terminal competitiveness. Toread's self-developed second-generation Crest C3 lower-limb intelligent exoskeleton achieves full-chain self-control from chip to equipment. During the 618 shopping festival, it sold over 100 units in a single day on Tmall, weighing only 1.8 kg with a push-to-weight ratio of 10:1. From compression IP to end products, the 'outdoor + chip' strategy is moving from industrial layout to commercial realization. Looking back, Toread is leveraging a driving model where 'outdoor provides stable cash flow, chips contribute growth elasticity, and compression technology connects the entire industry chain,' occupying a scarce niche in the edge AI track. As memory compression becomes a fundamental necessity for the AI industry, trillion-yuan markets in personal devices, smart vehicles, embodied robots, and AR/VR wearables are opening up.