Breaking Through the AI ‘Cost Barrier’ to Build a New Base for AI Enterprises

After Kuhn, computer science giant and database expert Jim Gray summarized the history of scientific and technological development into four paradigm phases: The first phase is experimental science, where natural phenomena are described based on experimentation or empirical observations. The second phase is theoretical science, where mathematical models are built to summarize and inductively derive scientific theories. The third phase is computational science, where scientific experiments are simulated using computers, leading to research breakthroughs through the “human brain + computer” combination. And the fourth phase, which we are currently in, is data science. With the explosive growth of research data and the rise of big data and AI, a technology revolution based on data is gradually unfolding.As data science emerges, computing power has become the “fuel” driving the rapid growth of the digital economy, and it has become a battleground for global tech giants. As the computing power industry chain continues to improve, the empowering effects of computing power are becoming increasingly evident. To ensure the ability to process vast amounts of data, the infrastructure for computing power has entered an accelerated development phase. We are now entering the “Intelligent Computing Era.”

From July 29 to 31, 2022, the “2022 China Computing Power Conference” was successfully held in Jinan, Shandong Province, hosted by the Ministry of Industry and Information Technology and the Shandong Provincial People’s Government. As the first China Computing Power Conference with the theme “Empowering Hundreds of Industries, Driving the Future,” what insights about the AI industry’s development will it reveal? It’s worth contemplating.Behind the competition for computing power lies a multi-faceted revolution in AI systems integration. In his 2015 book The Inevitable, Kevin Kelly argued, “Technology has an inherent preference that drives it in a specific direction.” From the early steam engines to the advent of electricity, now that specific direction is focused on computing power.Last year, Google launched its next-generation AI ASIC chip, TPUv4, which powers a computing cluster capable of releasing up to 1 exaflop of computing power—surpassing the fastest supercomputers in the world. Tesla’s AI and autonomous driving vision head, Andrej Karpathy, revealed Tesla’s Dojo supercomputer, which has 720 nodes and 5,760 NVIDIA GPUs, with a theoretical total computing power of 1.8 exaflops. In January this year, SenseTime’s AI Computing Center officially began operations, and as Asia’s largest AI computing power hub, it can process a total computing power of 3.74 exaflops per second.

It can also simultaneously connect to 8.5 million video feeds and complete the equivalent of 23,600 years’ worth of video processing in a single day.In March this year, the 2021-2022 Global Computing Power Index Assessment Report compiled by IDC and Tsinghua University’s Global Industry Research Institute revealed that every 1-point increase in the computing power index could lead to a 3.5‰ growth in the digital economy and a 1.8‰ increase in GDP. Thus, the importance of computing power is not only reflected at the enterprise level but has also risen to the level of national competitiveness.The significance of this first China Computing Power Conference, hosted by the Ministry of Industry and Information Technology and Shandong Provincial People’s Government, is evident. The conference featured two main forums, one press conference, 27 sub-forums, and an exhibition of “Computing Power Achievements.” It released China’s first comprehensive computing power index, the first “Computing Power White Paper,” and the first “Computing Power Facility Industry Map,” among other results. The event attracted numerous enterprises in the big data, AI, and related fields, including SenseTime, which actively participated in the conference and hosted the “AI Computing Power • Seeing the Future—SenseTime AI Infrastructure Innovation Forum.”In addition, SenseTime co-founder and AI Big Devices Business Group President, Yang Fan, delivered a keynote speech at the main forum on “AI Infrastructure: A New Engine for Industrial Development.”

Yuval Noah Harari, in his book Homo Deus, defined “Artificial Intelligence” as: AI, mainly embodied in observation, cognition, thinking, language, computation, and movement, with the advantages of extremely fast computing speeds, huge storage capacities, networked information sensing, and almost infinite iterative accumulation due to the lack of biological death, far surpassing human intelligence in many domains.In reality, AI is not just a single technology, but a complex system built from multiple technologies. This operational system is becoming increasingly refined. The SenseCore AI system, displayed at the SenseTime sub-forum at the conference, is a prime example. The SenseCore AI infrastructure consists of three parts: the model layer, deep learning platform, and computing infrastructure. The model layer includes the OpenDILab open-source platform, OpenMMLab open-source framework, and Model Factory, which has released over 20 general-purpose decision-making AI algorithms and developed over 34,000 commercial AI models. The deep learning platform, built on algorithm innovation, leverages SenseTime’s SenseParrots training framework to efficiently use GPU clusters, achieving over 90% acceleration efficiency when training a single large model on 1,000 GPUs. The computing infrastructure includes the AI computing center with 3.74 exaflops of computing power, AI chips supporting large vision models with over 10 billion parameters, and sensors and ISP chips capable of completing the training of a 100 billion-parameter model in a single day, thus forming an open and rapidly developing ecosystem.

Through the establishment of a new type of AI computing center, SenseCore has connected SenseTime’s leading algorithm training platform and data processing platform, opened the algorithm model framework to the community, and enabled customized production through the algorithm model factory for enterprises. This greatly reduces the cost of AI production elements and enables high-efficiency, low-cost, and scalable AI innovation and application. For instance, at the West Coast Group in Shanghai, the AI model provided by SenseTime continuously monitors over 6,800 commercial assets in a complex area, including trees, street lights, sidewalks, and other facilities. By June 30, 2021, the system could process about 200 work orders monthly, with over 98% resolved within 20 minutes, greatly improving safety and the experience for residents and visitors.Chen Wei, a member of the Party Leadership Group and Deputy Director of the Big Data Bureau of Zibo, stated in a roundtable forum that, in the digital economy era, computing power should be considered the most essential productivity. The role of computing power far exceeds what we had imagined. As computing power improves, it will qualitatively enhance our lives in areas such as digital economy, digital society, and digital government. From a long-term perspective, AI, under the multi-faceted integration revolution, may become a key node in the growth of AI technology and industry.

For the industry, this also means that the large-scale commercialization of AI technology has laid the foundation.Beyond the Three Pillars of AI: Solving the “Know-How” Dilemma in Industries.From the three hundred years of development in the technology-economy landscape, we can observe that any large-scale technological-economic paradigm shift generally follows the same pattern. These shifts go through four stages: Explosion, Frenzy, Coordination, and Maturity.Explosion Phase: Due to stagnation or deflation, the lack of investment opportunities and the emergence of new products and technologies ignite people’s enthusiasm for investment, triggering a technological revolution. This phase leads to the rapid emergence of innovations and entrepreneurs.Frenzy Phase: Productivity experiences explosive growth, infrastructure construction intensifies, and bubbles begin to form.Coordination Phase: After the technological frenzy, the bubbles gradually clear. The technology starts to diffuse on a large scale, meeting the long-tail demand across various industries, leading to broad economic growth.Maturity Phase: As technology matures, profit margins decline. To maintain profitability, companies often resort to mergers and acquisitions and begin expanding into new markets.Looking at the development of the AI industry, it is clear that AI is currently in the critical Coordination Phase. In this stage, AI’s technological innovations have entered a new phase of fusion and diffusion.

This means that, on one hand, as AI expands into various industries, more focus needs to be placed on the three foundational pillars of AI: computing power, algorithms, and data. This focus is reflected in the trends within AI companies’ development. For example, SenseTime’s newly established AI computing center has a massive AI model production capacity, enabling more general-purpose large-scale model training. The algorithm training platform achieves nearly 90% parallel processing efficiency, accelerating computations by 900 times, equivalent to 1 GPU performing at the efficiency of 1,000 GPUs. The algorithm model factory can leverage the training platform to generate large-scale data models and open them up to more platforms. The “2022 China Computing Power Empowerment Excellence Case” was unveiled at the conference, where SenseTime’s AI Computing Center was successfully selected.On the other hand, while strengthening the AI foundation, it is essential not only to focus on technological advancements but also to address the issue of integrating AI with industry-specific Know-How. This is because AI can only achieve widespread technological diffusion when it is integrated with the characteristics of specific industries, bringing real value by helping businesses reduce costs and improve efficiency. AI needs to combine with industry knowledge, refining industry skills within application scenarios to meet the long-tail needs of specific sectors and truly create value in the business world.

This will position AI as the bridge and connector between technology and various industries.In fact, this is also the goal of SenseCore, the AI infrastructure developed by SenseTime. It provides a robust foundation for AI’s future development. Through an end-to-end architectural system, it is capable of dissecting and colliding massive amounts of data to uncover potential value, breaking through the boundaries of cognition and application. The stronger the commercial impact AI companies have in specialized fields, the more favorable it is for AI’s diversified application. As AI technology continues to permeate various vertical industries, a business ecosystem spanning multiple sectors will take shape.”AI Mother Machine”: The Battle Between Technology and Costs.Over two hundred years ago, British economist William Stanley Jevons pointed out that reducing the cost of technology would enhance its popularity, thereby expanding the market scale. During the steam engine revolution, British coal miners feared that steam engines would increase efficiency, thereby reducing coal consumption and negatively affecting their businesses. However, the opposite occurred: the steam engine reduced costs and increased its use, ultimately leading to a significant increase in coal consumption.In a similar vein, SenseTime’s co-founder and Chief Scientist of Big Devices, Lin Dahua, stated during a speech that when artificial intelligence enters an industry, it often finds that the entire production chain is extremely long.

In response to every specific demand from long-tail scenarios, the costs and challenges are high. Therefore, to make AI the cornerstone of national economic development, we need to break through a new “red line”—the cost barrier.In the ongoing “computing power war,” the emergence of SenseCore by SenseTime signifies a key moment where the cost of AI technology is finally starting to decrease. Machine tools are often called the “mother machine” of the industry because they are used to manufacture other machines and are central to modern industry. If machine tools are considered the “mother machine” of the industrial age, then SenseCore, the AI infrastructure by SenseTime, can be thought of as the “AI mother machine.” In the future, it will become the heart of AI technology.SenseCore extracts the commonalities of AI and manufactures AI to reduce research and development labor costs and computing resource costs, which in turn helps expand the market scale. It acts like a software production line that elevates the entire industry, transforming the traditional “craft workshop” model. For instance, Qingdao FAW relies on SenseTime’s AIDC (Artificial Intelligence Data Center) to provide powerful computing support. By using SenseTime’s DeepSpring AI algorithm training platform, the company conducts deep learning on a large number of stamped sheet metal quality defect samples, generating exclusive algorithm models.

During actual production, specialized cameras used in the inspection process are pre-configured with the defect detection algorithms for real-time alerts. This process has enhanced stability in detection, increased defect detection rates, and effectively addressed customer pain points.Computing power is production power. According to officials from the Ministry of Industry and Information Technology, China’s computing power has grown at an average annual rate of over 30% in the past five years, ranking second globally. With such rapid growth in computing power, AI will inevitably evolve from a single-path, generalized computing model to a more diversified form, combining with various industries and promoting the development of the digital economy across sectors.Policies such as the “14th Five-Year Plan for the Development of the Information and Communication Industry” and the “14th Five-Year Plan for the Development of the Digital Economy” point out that “the digital economy is entering a new phase of deepening applications, standardized development, and inclusive sharing.” By 2025, “a new type of digital infrastructure that is high-speed, ubiquitous, integrated, intelligent, green, secure, and reliable will be basically built.” The host province of the forum, Shandong, is also actively driving the digital economy. It has outlined a plan to accelerate the construction of data and computing power infrastructure to provide new opportunities for AI companies’ development.