The artificial intelligence landscape is seeing significant shifts today as major hardware investments in China collide with rapid advancements in inference efficiency and robotics. While Chinese tech giants commit tens of billions of dollars to secure Nvidia silicon, the software ecosystem is evolving through the release of high-performance inference frameworks and highly optimized small-scale models. Simultaneously, the physical embodiment of AI is advancing, evidenced by successful long-duration endurance testing for humanoid robots and the expansion of autonomous vehicle fleets.
Chinese Tech Giants Place $35 Billion in Nvidia Orders
Major Chinese technology firms, including ByteDance, Alibaba, and Tencent, have placed orders exceeding $35 billion for Nvidia AI chips. This massive capital expenditure highlights the continued investment in AI infrastructure within the Chinese market, despite ongoing geopolitical tensions and export controls. Industry analysts suggest these orders, which include millions of units of the RTX 5500 Pro, are likely intended to bolster large language model (LLM) inference capabilities and large-scale AI video generation. This move underscores a significant divergence in market sentiment compared to some Western narratives regarding the current state of AI investment.
DeepSeek Releases DeepInfer Inference Framework
DeepSeek has officially released DeepInfer, an open-source, high-performance inference framework designed specifically for DeepSeek models. The framework is engineered to support large-scale deployments and provides an inference backend that can adapt to various API formats, such as Chat Completions and Messages. By open-sourcing this technology, DeepSeek provides the developer community with the tools necessary to study and improve upon existing custom engines, such as gufo, halogen, and STRATA, potentially accelerating the development of more efficient local inference solutions.
Figure F.03 Completes 348-Mile Endurance Test
The Figure F.03 humanoid robot has successfully completed a rigorous endurance testing phase, demonstrating significant progress in mechanical reliability. During the test, the robot walked continuously on a treadmill for nine days, completing approximately 1.2 million steps. This movement covered a total distance of 348 miles (560 km). The successful completion of this phase marks a critical milestone in testing the durability of humanoid actuators and battery management systems required for long-term autonomous operation in real-world environments.
Qwen 3.8 Flash Next and EAGLE-4 Performance Benchmarks
New performance metrics have been released for the Qwen 3.8 Flash Next model, which demonstrates increased precision and speed in complex reasoning tasks. In comparative benchmarks, the model showed significant advantages in coding and feature implementation. Complementing this, the EAGLE-4 inference acceleration framework has been released, showing substantial speedups over previous methods. In testing, EAGLE-4 achieved a 3.59× speedup over vanilla implementations, outperforming DFlash 2 and DSpark, which is particularly beneficial for running larger models like Qwen3.8-27B more efficiently.
China Increases AI Computing Allocation for Video Content
A recent report indicates that 55% of China’s total AI computing power is currently dedicated to the production of short dramas and video content. This high level of resource allocation reflects the massive scale of the AI-generated video market. As demand for high-fidelity video continues to grow, there is ongoing speculation that regulators may eventually shift compute allocations away from content generation to prioritize the training of more foundational large-scale models.
Elon Musk Warns of AI Persuasion Risks
Elon Musk has issued a warning regarding the potential risks associated with the persuasion capabilities of superintelligent AI. Musk noted that highly advanced models could potentially “one-shot” humans through sophisticated linguistic manipulation, making it difficult for individuals to maintain autonomy during arguments. This concern centers on the counterintuitive nature of AI persuasion, where a model might use psychological tactics to influence human behavior or decision-making processes in ways that are difficult to detect or resist.
STRATA Optimization and Local Model Performance
New performance metrics for the STRATA optimization framework have been released, highlighting its effectiveness in enhancing local model execution. Users have reported significant speed increases when utilizing STRATA on specific operating systems, achieving average speeds of 30 tokens per second on certain hardware configurations. This optimization is part of a broader trend toward making high-performance local LLM execution more accessible to users without massive enterprise-grade hardware.
Waymo Expands Robotaxi Operations and Safety Protocols
Waymo has announced an expansion of its robotaxi operations, aiming to grow its fleet from 4,000 to 20,000 vehicles by the end of 2027, supported by a $5 billion loan. The company also updated its safety protocols, reporting that its autonomous fleet has completed 271 million driverless miles with a rate of only 0.01 serious-injury crashes per million miles. This data suggests that Waymo’s current safety protocols result in significantly fewer serious accidents compared to human drivers.
Release of Small-Scale Models: Pocket TTS, Phonon, qev, and Spark-X2.6
A series of efficient, small-scale models have been released to facilitate edge computing and local decision-making. These include Pocket TTS, a 1.9 million parameter model capable of voice cloning, and the Phonon edge models (31M and 40M parameters) designed for low-latency text-to-speech. Additionally, the qev-0.8b decision model has been introduced as a new state-of-the-art tiny local decision model, alongside the Spark-X2.6-4B model, which is optimized to manage inference costs dynamically for resource-constrained environments.