I. Pre-market News
Macro:
The Minister of Finance, Lan Fu'an, stated at a press conference held by the State Council Information Office on October 12th that there is a plan to increase the debt limit on a one-time basis to a larger scale, to replace the existing implicit debt of local governments, and to intensify efforts to support local governments in resolving debt risks. The relevant policies will be detailed to the public after the legal procedures have been completed. It needs to be emphasized that this upcoming policy, which is about to be implemented, is the most significant measure to support debt resolution in recent years. Undoubtedly, it is a timely policy rain that will greatly alleviate the pressure of local government debt issuance, freeing up more resources for economic development, boosting the confidence of business entities, and consolidating the basic guarantees at the grassroots level. Hot
China's CPI in September rose by 0.4% year-on-year, lower than the expected 0.6%, and remained unchanged month-on-month. The PPI decreased by 2.8% year-on-year and fell by 0.6% month-on-month. Hot
On October 13th local time, a reporter from the central station learned that the Ministry of National Defense of North Korea stated that the General Staff of the Korean People's Army has issued instructions to front-line troops to prepare for shooting. Hot
Industry:
After Shenzhen optimized its real estate market policy on September 29th, the local commercial housing market has seen "daylight discs". The Shenye Shangcheng Xuefu, located in Longhua, opened for sale today. The project offered 332 housing units this time, and the final number of registered customers who made the cut was 744 batches, with a registration ratio of 1:2.24. It is understood that the average selling price of the project is 76,000 yuan/square meter, which opened at 3 pm today and was sold out in three and a half hours, with the total sales amount of this batch of housing units being about 2.6 billion yuan.
Advertisement
Company:
The latest wholesale reference price disclosed by today's liquor price shows that the price of the 2024 Feitian Moutai loose bottle is reported at 2,240 yuan/bottle, unchanged from the previous day; the 2024 Feitian Moutai original box is reported at 2,300 yuan/bottle, down another 20 yuan from the previous day.
On the evening of October 13th Beijing time, the fifth test flight of Starship (S30+B12) was conducted: (1) The first stage of the super-heavy booster was launched and returned in 6 minutes, successfully recovered by the launch tower mechanical arm "chopsticks"; (2) After the second stage spacecraft entered orbit, it successfully made a controlled splashdown in the Indian Ocean; (3) The real-time footage of the test flight was transmitted back to the ground through Starlink.II. Forward Look
10.14 - China's import and export data for September
10.17 - Initial jobless claims in the United States for the week ending October 12, and the monthly rate of U.S. retail sales in September
10.18 - China's GDP for the third quarter, and China's economic data for September
III. Public Opinion Hotspots
Debt-to-equity AMC - National Development and Reform Commission: It is necessary to ensure adequate fiscal expenditure, strengthen the positive role in promoting economic development, increase support for local debt swaps, and resolve debt risks; Hot
Huawei HarmonyOS - Huawei's Mate 70 series of smartphones have entered full-scale mass production, equipped with a pure HarmonyOS system, and may be officially launched in November;
Private Economy/Double 11: JD.com's "Double 11" starts early at 8 PM on October 14; direct spot sales; recently, the Ministry of Justice and the National Development and Reform Commission have published the "Private Economy Promotion Law (Draft for Comments)" on their official websites, openly soliciting opinions from society. The formulation of the Private Economy Promotion Law is an important deployment made by the Third Plenary Session of the Party's Twentieth Central Committee;
GB200 Supply Chain/Copper Connections - Blackwell's demand is astonishing! Nvidia's orders for the next year have all been sold out; (Kaiwang Technology, Shenyu Shares, Kapaiyi, Luxshare Precision, Megalong Interconnect, etc.);
Commercial Spaceflight - Starship (S30+B12) conducts its fifth test flight, the first stage of the heavy-duty booster returns after 6 minutes of launch and is successfully retrieved by the launch tower's mechanical arm "chopsticks".IV. Hotspot Research Reports and Minutes
1. **Open Source Communications** Blackwell Servers Ship in an orderly manner, Tesla CyberCab Officially Launched, Optimistic about the AI Computing Power Industry Chain
**Events**
• Blackwell Servers Ship in an orderly manner, promoting a new round of AI development. The shipment of Blackwell chip servers is expanding the computing power industry chain.
On October 8, 2024, Foxconn's parent company, Hon Hai Precision Industry, held the 2024 Hon Hai Technology Day. The company's chairman stated that Hon Hai and its supply chain are ready to become "the first company to mass-produce and ship (NVIDIA chips) GB200," and are currently building the world's largest GB200 chip production base for NVIDIA in Mexico. It is expected that by 2025, the production capacity of NVIDIA NVL72 servers will reach 20,000 units. In addition, the chairman indicated that the GB200 server is planned to be shipped in the middle to late fourth quarter of 2024. On October 8, Microsoft Azure Cloud became the first cloud provider to run the Blackwell system, equipped with GB200 AI servers, paired with IB networks and liquid cooling for operation. On October 9, OpenAI announced that it had received the first batch of engineering version DGX B200 servers. With the orderly shipment of B-series AI servers, it is expected to stimulate a new round of high computing power infrastructure construction, supporting the continuous iteration of generative AI.
The Nobel Prize goes to the AI field, and AI has a profound impact on social development. The 2024 Nobel Prize in Physics was awarded to John J. Hopfield and Geoffrey E. Hinton for their fundamental discoveries and inventions in the use of artificial neural networks for machine learning. The Hopfield network and Hinton's Boltzmann machine are important foundations for the vigorous development of modern deep learning. Half of the Nobel Prize in Chemistry was awarded to Demis Hassabis and John Michael Krop, in recognition of their contributions to computational protein design, using AI technology to predict the structure of almost all known proteins. With the continuous iteration and development of AI models, the reliability in professional fields continues to improve, which is expected to accelerate the empowerment of various industries and drive the continuous growth of inference computing power demand.
• AMD announces MI325X, MI355X parameters, and releases the first UEC AI network card to help expand AI clusters. AMD releases MI325X GPU, server CPU, AI network card, and DPU.
On October 10, AMD released the AMD Instinct MI325X GPU at AMD Advancing AI 2024. It has 153 billion transistors, uses the CDNA 3 architecture, and 256GB HBM3E memory, with a memory bandwidth of 6TB/s. The FP8 peak performance reaches 2.6 PFLOPS, the FP16 peak performance reaches 1.3 PFLOPS, and the AI peak computing power reaches 21 PFLOPS. The memory capacity is 1.8 times that of H200, and the memory bandwidth, FP16, and FP8 peak theoretical computing power are 1.3 times that of H200. In 2025, the MI350 series using the CDNA 4 architecture will be launched, and the FP8 and FP16 performance of MI355X will be increased by 80% compared to MI325X, with the FP16 peak performance reaching 2.3 PFLOPS, the FP8 peak performance reaching 4.6 PFLOPS, and the FP6 and FP4 peak performance reaching 9.2 PFLOPS. The AI peak computing power will reach 74 PFLOPS, and the TDP will rise to 1000W. The MI400 series will use a more advanced CDNA architecture. The company expects MI325X to be fully produced in the fourth quarter of 2024, and MI355X is planned to be shipped in the second half of 2025.
AMD releases the first UEC backend AI network card. In terms of networking, AMD's vice president stated that networking has become the key to constraining AI system performance, with an average of 30% of training time spent waiting for networking. In training and distributed inference, communication accounts for 40%-75% of the time. From a TCO perspective, compared to IB networks, UEC Ethernet networks have decreased by 50%; in terms of scalability, Ethernet can support the deployment of ultra-large-scale cluster deployments with more than 1 million+ GPUs, far exceeding the 48,000 expansion limit of IB networks. AMD released the industry's first UEC ultra-Ethernet alliance backend network AI network card Pensando Pollara 400 and Pensando Salina 400 DPU, supporting 400G ports, both of which will provide samples in the fourth quarter of 2024 and are expected to be launched in the first half of 2025.
• Tesla CyberCab taxi debuts, accelerating the development of autonomous driving and humanoid robots.On October 11th, Tesla held the We Robot event, where the Tesla Robotaxi autonomous taxi, CyberCab, was officially launched. It features two seats and butterfly doors, without a steering wheel or pedals, and comes with inductive charging capabilities. CEO Musk stated that the cost of the CyberCab is expected to be below $30,000, and fully unsupervised autonomous driving (FSD) is anticipated to be realized in Texas and California by 2025. Production is expected to begin in 2026, with a large-scale rollout in 2027.
Currently, the operating cost of CyberCab is relatively high, at around $1 per mile, but it is expected to decrease to approximately $0.2 per mile in the future. There are also plans to introduce a CyberCab 2 model. In addition, Tesla unveiled the autonomous Model Y and RoboVan (autonomous van), which can carry up to 20 people and transport goods, significantly reducing travel costs to about 10-15 cents per mile.
The humanoid robot Optimus has also made significant progress and can handle many daily tasks, with individual costs expected to be between $20,000 and $30,000. We believe that the implementation of Tesla's FSD technology and embodied intelligent humanoid robots will require extensive data training, which is likely to continue to drive demand for AIDC and computational infrastructure.