ChinAI #312: New-type AI Storage Research Report (Part 2)
Greetings from a world where…
Andor is peak TV
…As always, the searchable archive of all past issues is here. Please please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay support access for all AND compensation for awesome ChinAI contributors).
Feature Translation: New-type AI Storage Research Report (part 2)
Context: We built this newsletter on white papers, so let’s dig back into the “New-type AI research report” (link to original Chinese pdf) jointly published by the China Academy of Information and Communications Technology and the China AI Industry Alliance. In the first part of the research report, the authors make a provocative claim: Of the triad of AI drivers (data, computing power, and algorithms), data is the most important.
Key Takeaways: The report’s last section doubles down on the importance of data, and hence, new-type AI storage: “At present, the development of global AI is accelerating from ‘model/computing power-centered’ to ‘data-centered’, and high-quality, large-scale, and diversified data have become key strategic elements for the development of AI.”
From the report’s best practices section, let’s examine some specific examples from the financial and manufacturing industries.
Project background in the financial industry: a bank wants to deploy scenario-specific AI models but its traditional object storage system faces issues with workloads spike and single-site failures. This bank adopts a new-type AI storage system. This reportedly decreases the retrieval latency of hundreds of billions of objects from >10 sec. to 50 milliseconds. It also supports deployment of scenario-specific AI models at 12 sites, with tolerance of two site failures.
See image below: The top row lists different AI-enabled banking applications — from the article: “800 AI models are used to empower 350 application cases, including customer experience, intelligent customer service, anti-money laundering, risk management, smart financial management and other business modules”; in the middle layer, there are big data computing clusters (lists tools and frameworks that help with data processing (Hadoop, Apache Spark, HBase) and AI clusters (lists Nvidia GPUs and Intel CPUs). Then the bottom layer is AI storage.
In the manufacturing industry case, a certain technology company makes a wide range of industrial storage products, but an increase in customers has brought increased pressures on business consultations (40,000 problems to address per year). This company worked with AI companies to integrate industry-specific knowledge bases with large models. Using long-term memory storage techniques, the integration of large models helped achieve 24/7 on-duty responses and improved user satisfaction by 95%. *Note: The point here is not to take these data points as incontrovertible truths (the companies and AI storage solutions are not named); rather, it is to get a sense of how this works in practice.
The report concludes with policy recommendations for China’s development of AI storage, including the incorporation of special projects in AI storage within national science and technology research plans.
One line that caught my eye: Drive the application of domestic storage chips through storage devices and solid state drive (SSD) controllers. When we think about AI infrastructure, we mostly think about GPUs. These SSD controllers “feed” GPUs by giving them fast and consistent access to the large volumes of data needed for AI applications.
One interesting player in this space is the Taiwanese firm Silicon Motion, which producers about 1/3 of the global market for SSD controller chips, supplying “80% of those chips to south China.” According to that linked article, US chipmaker MaxLinear was set to buy Silicon Motion, but the deal ultimately fell through.
FULL TRANSLATION: New-type AI Storage Research Report
ChinAI Links (Four to Forward)
Must-read: China built hundreds of AI data centers to catch the AI boom. Now many stand unused.
For MIT Technology Review, Caiwei Chen reports:
According to people on the ground who spoke to MIT Technology Review—including contractors, an executive at a GPU server company, and project managers—most of the companies running these data centers are struggling to stay afloat. The local Chinese outlets Jiazi Guangnian and 36Kr report that up to 80% of China’s newly built computing resources remain unused.
A good indicator of high-quality reporting on China’s AI landscape: the reporter has taken the time to read Chinese-language sources (in this case, Jiazi Guangnian [甲子光年]).
Should-read: [citation needed], a newsletter by Molly White
The pitch: keep up with happenings in the cryptocurrency industry without the boosterism. It’s cool to see another newsletter with the same funding model as ChinAI: “All content is free and available to all readers, and will continue to be that way…That said, paid subscribers are crucial to allowing me to continue doing this kind of research and writing, and so if you are able to support my work I would be immensely grateful…If you really love my work, there’s also a founding member option where you can pledge a custom amount for an annual subscription. This is also a great option if you happen to have, say, an employer-sponsored reading or education budget.”
Should-read: DeepSeek, Huawei, Export Controls, and the Future of the U.S.-China AI Race
By Greg Allen, an extremely well-researched CSIS report on the implementation and effectiveness of chip controls. Some really interesting details on how the Commerce Department could not predict Nvidia’s A800 modification:
Industry sources confirmed to CSIS that Nvidia blew fuses on A100 chips to reduce their interconnect speed (but not their processing power) below the export control performance thresholds, thus creating the A800 product lines.
Should-read: The Anatomy of Chinese Innovation: Insights on Patent Quality and Ownership
A new paper that analyzes innovation in China with some intriguing findings: “Chinese patenting has become narrower and less innovative over time. The role of overseas knowledge has also declined sharply.” To study all invention patents applied for at the China National Intellectual Property Administration, researchers from Leibniz Centre for European Economic Research, University of Toronto, and Central University of Finance and Economics use a LLM to derive information on patent abstracts and claims.
Two extra links: other Cool Corners of the Internet
I’ve been very slow to explore Substack’s recommendations function. It’s a great way to explore the Internet’s nooks and crannies.1 It’s gratifying to see all the China- and AI-related newsletters that recommend ChinAI, but it’s also fun to see some unrelated ones in the list. I wanted to spotlight two off-road newsletters in particular:
Rachel Cohn’s “Are You My Boyfriend” newsletter, which features stories about her own dates as well as her commentary on emerging trends in the online dating world.
Joseph Gordon-Levitt’s journal on a wide range of topics including Nirvana, his acting career, and mindfulness.
Thank you for reading and engaging.
These are Jeff Ding's (sometimes) weekly translations of Chinese-language musings on AI and related topics. Jeff is an Assistant Professor of Political Science at George Washington University.
Check out the archive of all past issues here & please subscribe here to support ChinAI under a Guardian/Wikipedia-style tipping model (everyone gets the same content but those who can pay for a subscription will support access for all).
Also! Listen to narrations of the ChinAI Newsletter in podcast format here.
One of my favorite spaces on the Internet was the comment section for Ta-Nehisi Coates’s blog for The Atlantic.