China is accelerating its investment in artificial intelligence (AI) and has an ambitious plan to have its core AI industry output increase to RMB 150 billion (US$22 billion) by 2020, a ten-fold increase in a 3-year time period. That amount is expected to grow to RMB 450 billion (US$66 billion) by 2025. Core AI technology refers to areas of research like deep learning, core AI technologies, the development of basic software and hardware, such as chips and sensors, and applied research in areas like computer vision and cybersecurity. Hardware is a key component of China’s AI plans.
China has been pushing for domestic hardware players since 2014, having raised close to $50 billion for its Integrated Circuit Fund supporting semiconductor investment. In 2017, China took the fight directly to NVIDIA and the United States, announcing a chip project that aims to beat NVIDIA’s M40 chip performance and energy efficiency by 20X. Recently, the Chinese AI chip startup Cambricon, valued at $1 billion, was estimated to have graphics processing units (GPUs) that are 6X faster than NVIDIA’s. With regard to semiconductor fabs, most of which are based in Taiwan and South Korea, China already has major plans to accelerate its fabrication industry to support 40% of semiconductor demand by 2020.
Hardware: A Key AI Battleground
Hardware is becoming a major battleground in the race toward AI dominance, as China sees its hardware capabilities lagging behind other countries, especially the United States. Hardware is also part of the recent trade war between China and the United States, with the United States blocking Chinese attempts at acquiring U.S. semiconductor companies. While China is being stifled on trade, AI hardware startup activity is a key area where China sees itself having an upper hand, especially in terms of creating and owning intellectual property (IP) and following the fabless semiconductor model that has propelled the United States to hardware dominance in the last few decades. Apart from Cambricon, which is possibly China’s best-known AI hardware company, the large internet giants like Alibaba and Baidu are also expanding their AI hardware efforts, reducing reliance on third-party and foreign suppliers. Baidu announced its own Kunlun AI chip at its recent Baidu Create event, while Alibaba has been promoting its Ali-NPU chip. Both Kunlun and Ali-NPU are in the same category as Google’s tensor processing unit (TPU) or Graphcore’s intelligence processing unit (IPU), all of which are accelerators targeted at speeding up AI training and inference. While Google’s TPU and Graphcore’s IPU are initially targeting AI training operations in cloud-based servers, Ali-NPU and Kunlun are targeting inference operations such as image processing or object recognition.
Chinese Focus on AI Edge Hardware
While details are sparse, the Kunlun chip is also known to work in edge scenarios such as autonomous vehicles, which means they are likely to have different power-performance versions for cloud and edge applications. Kneron, a Chinese AI hardware startup, is specializing in AI edge applications like mobile devices, the Internet of Things (IoT), and cameras. Horizon Robotics is another Chinese AI hardware company focused on vision-based AI applications using embedded chips performing edge-based AI.
The particular focus of Chinese AI hardware startups around edge computing, specifically imaging, is interesting because China is the leading market for video surveillance applications, and it is also leading with AI-based video surveillance with AI software companies like SenseTime and Megvii having raised more than $2.5 billion in financing. Apart from video surveillance, drones, and autonomous cars, the robotics and mobile device sectors are expected to be among the leading adopters for AI edge hardware. China has a leading position in robotics and is well placed to lead in consumer and enterprise service robotics as well with companies like Ubtech and Sanbot leading the charge. DJI is the number one drone manufacturer in the world, and Xiaomi is among the top mobile smartphone vendors.
AI Edge at Odds with Chinese Governance Model
From a philosophical and architectural standpoint, having AI processing done at the edge is antithetical to the centralized governance mechanisms in China. AI at the edge promotes a decentralized hive mind, with individual agents making their own decisions locally rather than being governed by a central computing infrastructure of authority. One of the core strengths of China in terms of its AI capabilities is its control of data, and its companies’ compliance with central government in terms of data sharing. By having a central focal point of data collection and data governance, China is in an unprecedented position to drive and excel at AI, especially the deep learning branch of AI that is data hungry and where AI performance is directly correlated to the size of AI models and the amount of training data used. Once that control and data collection is performed at the edge, with the data staying at the edge, it would become harder to control for any central authority, especially when it comes to areas like video surveillance and security.
So why have we not seen any major pushback from the Chinese government around AI edge processing? We are in the very early stages of the AI technology cycle, where the market is going through a hype phase and the Chinese central government is keen to push all AI technology, without closer scrutiny. In other words, better technology and efficient engineering techniques that use edge-based processors for decentralized governance are being promoted and funded by a centralized governance authority, i.e., the Chinese government.
In the longer run, with the advancements in AI techniques like meta learning or simulation-based learning, the reliance on collection, management, and storage of large datasets and training in a centralized fashion will become less important. AI models will exist in a decentralized agent-based architecture, with individual nodes or agents running edge-based training and inference using embedded hardware, while communicating with a central node in terms of updating the weights and model parameters. Think of Google’s Federated Learning model with the Chinese government being the federation!
AI meta learning architectures are likely to prioritize models rather than data, and so, in a sense, we will see the return of centralized AI management. Ultimately, edge-based AI hardware should not have any major impact on the ability of a central authority to govern AI itself.
It would be surprising if the Chinese government is thinking on these lines or has started to question the impact of edge-based AI architectures of AI governance. We are clearly going to see some interesting contradictions come to the surface as China deploys and develops AI.