[SMM Tin Midday Review: Weak Supply and Demand Combined with Macro Expectations Game, SHFE Tin Consolidates Above 420,000 yuan/mt]

Zimbabwe's lithium producers have committed approximately $1.45 billion to local beneficiation infrastructure, Lithium Producers Association Chairman Innocent Rukweza said at the Mine Entra Beneficiation and Value Addition Symposium in Bulawayo. The industry is seeking government flexibility on the pending unbeneficiated spodumene export restriction, citing limited sulphate-plant readiness. Of seven major producers, only Prospect Lithium Zimbabwe's Huayou-linked Arcadia plant ($400 million) has commissioned sulphate conversion, exporting Africa's first locally produced lithium sulphate in April 2026. Sinomine's $500 million Bikita plant remains under construction; Kamativi Mining Company's $200 million-plus project is at investment stage; a fourth facility targets end-2027. Concentrate exports move CIF China via Beira and Durban, with grid instability and border/rail congestion at Machipanda, Beitbridge-Durban and Maputo adding risk to export timelines. The industry's effective tax burden is estimated at ~40% of sales revenue (10% export tax, 7% royalties, 3% community levy, 1% MMCZ fee, 15.5% VAT), which the association has called SMM View: Zimbabwe's beneficiation drive remains a one-plant story Arcadia commissioned, three others still at construction or investment stage. Grid reliability and corridor congestion, not capital alone, will determine how fast that gap closes.

SMM, August 12: Foxconn Industrial Internet’s semi-annual report posted substantial profit growth, further confirming the strong momentum of AI computing infrastructure and driving a recovery in sentiment across the high-speed interconnect industry chain. Against the backdrop of expanding computing power, high-speed copper cables have drawn attention from some market funds as a key short-range interconnect solution for AI servers. At the same time, SMM observes that the installation rush for power batteries, high prosperity in energy storage, and surging AI computing-end demand have collectively driven an ongoing climb in operating rates in the copper foil industry. As of the close on August 12, the high-speed copper cable concept rose 2.22%. Among individual stocks, Taichenguang and Hengdongguang jumped over 6%, while the biggest gainers included Xianying Technology, Ruikeda, Dingtong Technology, Far East, Changxin Bochuang, ZTT, and Zhaolong Interconnect. Market News [Shanghai: Build 100,000-card-level ultra-large intelligent computing clusters in Songjiang, Lingang, Qingpu, etc.] The Shanghai Municipal Commission of Economy and Informatization issued the “15th Five-Year Plan for the Development of the Software and Information Services Industry in Shanghai.” The plan mentions creating a tiered supply system that synergizes “large clusters + small clusters + edge computing,” building 100,000-card-level ultra-large-scale intelligent computing clusters in Songjiang, Lingang, and Qingpu, and constructing 1,000-card-level clusters in Baoshan, Pudong, and Jiading. It guides the transformation of traditional data centers and ICT server rooms into 100-card-level edge intelligent computing centers to meet ultra-low-latency computing demands from enterprises and individuals. Focusing on industries such as finance, education, healthcare, culture and tourism, and manufacturing, the plan supports building Model as a Service (MaaS) platforms, providing industry application marketplaces, model customization and hosting, agent building, low-code development, API interfaces, computing power provision and management, and AI inference services, thereby upgrading intelligent computing cloud service capabilities. It also highlights tackling next-generation model architectures and promoting exploration of multiple technology routes based on non-Transformer architectures such as state space models, recurrent neural network variants, and liquid neural networks. Efforts will be accelerated to lay out technology systems for cutting-edge foundation models, including physical intelligence, world models, quantum intelligence, and brain-inspired intelligence. The plan further addresses breakthroughs in networking technologies for ultra-large-scale intelligent computing clusters, focusing on core segments such as high-performance computing chips (GPU/NPU), quantum chips (QPU), high-speed optical interconnects (CPO), high-bandwidth memory (HBM), and heterogeneous servers to boost supply capacity for intelligent computing hardware and facilitate the deep integration of proprietary chips with mainstream large models. With an emphasis on new storage retrieval and data-model collaboration, the plan aims to achieve breakthroughs in high-precision heterogeneous processing, native multimodal fusion, and dynamic value alignment, and to build automated complex reasoning covering the full life cycle of corpus data. [Foxconn Industrial Internet: H1 2026 net profit up 95.99% YoY; AI computing demand continued to surge during the reporting period] Foxconn Industrial Internet announced on August 11 that its H1 2026 revenue was 557.861 billion yuan, up 54.63% YoY. Net profit attributable to shareholders of the listed company was 23.74 billion yuan, up 95.99% YoY. Net profit attributable to shareholders of the listed company after deducting non-recurring profit or loss was 22.984 billion yuan, up 96.99% YoY. Basic earnings per share was 1.2 yuan. The company plans not to distribute cash dividends, not to issue bonus shares, and not to convert capital reserve into share capital. The change in operating revenue was mainly due to the benefit from the continued surge in AI computing power demand, steady increase in market share among major clients, and strong performance of cloud service business, driving overall revenue growth. The change in net profit was mainly due to the benefit from the continued surge in AI computing power demand, with the company's main business operations achieving steady improvement in profitability. (Jin10 Data) [CoreWeave Second-Quarter Revenue Doubles, Shares Surge 12% After Hours] CoreWeave (CRWV.O) rose 12% in after-hours trading on Tuesday after reporting second-quarter revenue of $2.58 billion, up 112% YoY and surpassing Wall Street expectations, indicating that demand for AI computing power is still growing rapidly; net loss was $626 million, compared to $290 million in the same period last year; order backlog reached $104 billion, with projects under construction totaling 1.5 gigawatts of capacity. CoreWeave is accelerating the expansion of its data center business, competing with cloud computing giants such as Amazon, Google and Microsoft to capture the market for data centers equipped with chips capable of running generative AI models. However, CoreWeave has yet to achieve profitability. As of the end of the quarter, its debt on the balance sheet reached $35 billion, used to cover NVIDIA GPU and other equipment procurement costs. This quarter, Meta said it would invest an additional $21 billion in CoreWeave. Additionally, CoreWeave announced a multi-year cooperation agreement with Anthropic and received a $6 billion commitment from quantitative trading firm Jane Street. (Jin10 Data) [Axera Next-Generation High-Power AI Chip Completes Tape-Out, Supports Multi-Chip Cascading for Full-Fledged Large Model Inference on the Edge] From the earnings call of Axera's 2026 semi-annual report, it was learned that the company's next-generation high-performance, high-power AI chip has completed tape-out, with a significant increase in computing power specifications, equipped with high bandwidth, and supports two-chip or four-chip cascading, enabling high-performance inference of full-fledged large models on the edge. [Strategic Cooperation Intent Reached, Huawei to Provide Ascend Computing Equipment to Beijing Data Group] According to Beijing Data Group, on August 7, Beijing Data Group and Huawei held working talks and reached a strategic cooperation intent. Next, the two sides will focus on deepening cooperation in computing power clusters and city-level computing infrastructure construction. Beijing Data Group’s subsidiary Tongniu Information will participate in the construction of Beijing's city-level computing infrastructure, coordinating the deployment, daily operations, and computing services of the group's self-innovated computing clusters. Huawei will fully support Beijing Data Group in advancing the city-wide layout of self-innovated computing, providing advanced Ascend computing equipment, comprehensive technical solutions, and service support to jointly build a trusted city-level computing foundation in Beijing, continuously releasing the value of computing engines and offering stable, reliable computing support for the development of “Digital Intelligence Beijing.” [Nvidia Announces Partnership with Six Financial Giants to Arrange $500 Billion AI Infrastructure Financing System] Nvidia (NVDA.O) announced on the 10th local time that it has established a strategic partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to create an independent computing financing platform, aiming to mobilize over $500 billion in third-party capital over the long term for building artificial intelligence infrastructure. Nvidia stated that the new financing platform transforms Nvidia’s computing and full-stack AI infrastructure into an investable asset class for global capital, expands access to AI factories, achieves long-term revenue tied to usage, and supports Nvidia’s ecosystem growth in hardware sales and software applications. [Tesla: Terafab Plan Launches in Texas, Targeting Over 1 Terawatt of Computing Capacity Annually] On August 6, Tesla (TSLA.O) stated that earlier this year, SpaceX and Tesla announced the launch of the “Terafab” project—the world’s largest chip manufacturing initiative, integrating logic chips, memory chips, and advanced packaging technology within a single facility. In April, Tesla broke ground on a new R&D fab at the northern campus of its Texas Gigafactory, which served as the predecessor to Terafab. And today, we officially announced that Terafab will be located in Grimes County, Texas. This facility will be an advanced semiconductor wafer fab, designed to bridge the vast gap between current global chip supply capability and future computing demand. The combined chip demand of SpaceX and Tesla is expected to exceed 1 terawatt (TW) of computing power, far surpassing the current global supply capability. We greatly appreciate our existing chip suppliers and encourage them to expand capacity where possible, but the widening supply-demand gap in the future is the core reason for the Terafab project. Terafab’s goal is to manufacture new computing capacity at unprecedented scale and speed. The project plans to build a vertically integrated factory with a manufacturing area exceeding 100 million square feet. The facility will cover the manufacturing, packaging, and testing of advanced logic chips and memory chips. Concentrating these processes at a single location will facilitate rapid iteration and accelerate the deployment of new computing power. [ZTE Partners With Sky47 to Build Pakistan's Largest Intelligent Computing Data Center] Recently, the inauguration ceremony of Sky47 Karakoram-01, the largest integrated general-purpose and intelligent computing data center in Pakistan, jointly built by ZTE and Pakistan's leading cloud service provider Sky47, was held in Islamabad. As Pakistan's first customized AI-native Tier III data center, Sky47 Karakoram-01 has a total power supply capacity of 8.5 MW. The center will provide robust cloud computing, data hosting, and advanced digital service support across Pakistan, fully meeting the computing power needs of government and enterprises in fields such as artificial intelligence (AI), machine learning (ML), and high-performance computing (HPC). Power Battery Cell Installation Rush and Robust AI Industry Chain Demand Keep Copper Foil Operating Rate Climbing According to SMM, In July, the operating rate of the copper foil industry continued to climb, supported by strong end-use demand from downstream. In the lithium battery sector, production schedules of major Chinese lithium battery companies hit another record high in July. The installation rush for power battery cells boosted production schedules, and demand for lithium battery copper foil remained positive. In the electronic circuit segment, AI industry chain-related demand remained strong; capacity continued to shift toward high-end products, and demand for electronic circuit copper foil across all specifications stayed robust. Voices From All Parties CSC Financial’s research report notes that the scaling of frontier models has entered a stage of parallel multi-path development. Anthropic Mythos 5 and Fable 5 are estimated by the industry to have 8 trillion and 5 trillion parameters respectively; Kimi K3 has a total of 2.8 trillion parameters; and ByteDance is reportedly pretraining a model with up to 10 trillion parameters. Post-training is further extended to million-token agent trajectories, thousands of tool calls, and complex tasks lasting several hours. The RSI review published on July 8 covered 1,250 papers, 74% of which were published in 2026, indicating a clear acceleration in AI R&D automation. We believe that model competition is shifting from single-parameter expansion toward coordinated evolution involving pretraining, reinforcement learning, inference-time computing, RSI, and long-term agent capabilities. Computing power demand will expand from training to inference and agent execution. We remain bullish on the Capex ecosystem of major players, domestic chips and super-nodes, computing services, Pre-AI, B-end AI applications, and local inference. Founder Securities' research report indicates that the market's oversold rebound has entered a critical phase, with divergence unfolding across tech and cyclical growth sectors. Continue to focus on allocation opportunities in three areas. First, tech stocks also require selective positioning by structure. At the index level, the Sci-Tech Innovation Board and ChiNext have rebounded about 10% from their lows, and there remains upside relative to the typical oversold rebound amplitude of major themes. The AI narrative has seen some shifts after US CSP earnings reports, with competitive capex marginally weakening; cloud business and healthy cash flow are decisive factors. Therefore, within AI, hardware and applications will become more balanced. Focus on core overseas computing power names with low crowding, as well as domestic computing power segments with high earnings visibility such as semiconductor equipment and materials; relatively undervalued AI applications and Hang Seng Tech deserve attention. Second, watch for opportunities in HALO assets, as expectations for US Fed interest rate hikes are easing. Beyond core resource-related non-ferrous metals and chemicals, oversold old and new energy, including power grids and electrical equipment, coal and petrochemicals, etc. Third, leading pharmaceutical names with improving fundamentals, low crowding, and abating headwinds. CITIC Securities' research report notes that since 2023, the rapid development of AI has driven the iterative upgrade of optical module technology. New technologies such as optical chip speed upgrades, silicon photonics integration, and CPO architecture are jointly driving the iterative upgrade of optical module testing equipment. Combined with the rapid expansion of AI computing power infrastructure, this is driving a "volume and price increase" for optical module testing equipment. Currently, international players remain relatively ahead in the 1.6T high-end market, but domestic enterprises are accelerating their catch-up, with the gap steadily narrowing. We are bullish on the long-term development of the domestic optical module industry and the trend of import substitution for high-end optical module testing equipment. China Merchants Securities, reviewing nine sharp A-share market corrections since 2015, found that sharp declines were mostly triggered by external shocks or liquidity risks, with stabilization marked by policy responses. The average rebound window after a correction is 34 trading days, with the Wind All A-Share Index rebounding by more than 19% on average, and the larger the prior decline, the greater the subsequent rebound tends to be. Sector performance exhibits "two-phase" rotation: in the first 10 trading days of a rebound, high-beta, oversold sectors such as electronics and computers lead the gains; after 20 to 60 trading days, the market shifts to themes with fundamental support, such as electrical equipment and food & beverage. For the current cycle, a two-step allocation is recommended: initially, prioritize TMT and other oversold, high-beta sectors (with focus on computing power leaders in China and overseas); after 10 to 20 trading days, return to a rebalancing of fundamentals, focusing on electrical equipment, chemical pharmaceuticals, coal, and non-bank financials. Along sector themes, key opportunities to capture are the catch-up potential in the overseas computing power price-hike chain, the elasticity of domestic computing power hardware, and gold's value as a safe-haven and rebalancing asset. The overall allocation revolves around three main themes: technology innovation, enterprises going global, and rebalancing of traditional low-valuation sectors. According to CICC research, since mid-to-late June, global AI chains experienced notable pullbacks, with South Korea—characterized by high leverage, high crowding, and high retail participation—being the most severe. Behind this were the amplifying effects of high crowding and high leverage, disturbances from macro factors (such as rising expectations for US Fed interest rate hikes and the renewed blockade of the Strait of Hormuz driving up oil prices), and market concerns over a re-emerging AI bubble (e.g., Meta renting out computing power, declining token spending). In fact, before the bubble finally burst in March 2000, the tech stock market saw at least four rounds of large-scale, prolonged corrections. The triggers for these declines are highly similar to the current adjustment: short-term setbacks in industry trends, headwinds in the macro environment, and overheated valuation sentiment. The eventual rebound in tech stocks was also due to the easing of these three pressures. Therefore, corresponding to the present, for the market to stabilize and even start a new round of increases, these three factors are needed: the digestion of high crowding and high leverage (largely achieved), the easing of expectations for US Fed interest rate hikes or the actual announcement (watch the July FOMC meeting), and more importantly, new catalysts from earnings reports and industry developments (the July-August earnings season). Recommended reading:

Zimbabwe's lithium producers have collectively committed approximately US$1.45 billion to local beneficiation infrastructure, according to Lithium Producers Association Chairman Innocent Rukweza, who also heads Mutapa Energy Resources. Speaking at the Mine Entra Beneficiation and Value Addition Symposium in Bulawayo, Rukweza said the industry has sought government flexibility on the pending export restriction on unbeneficiated spodumene concentrate, citing limited sulphate-plant readiness across the sector. SMM notes the ban's effective date and any adjustment to it should be confirmed against the latest government notice rather than symposium commentary, and that the US$1.45 billion figure is association-reported and has not been independently verified. Of seven major producers, only the Huayou-linked plant at Prospect Lithium Zimbabwe's Arcadia site has completed and commissioned sulphate conversion capacity to date. The US$400 million facility exported Africa's first locally produced lithium sulphate in April 2026 and is now operating. Sinomine's US$500 million sulphate plant at Bikita remains under construction, Kamativi Mining Company's US$200 million-plus sulphate project is at investment stage with commissioning status unconfirmed, and a fourth, unnamed facility is targeted for end-2027. The gap between committed capital and commissioned output is the operative supply constraint behind the industry's request for timeline flexibility: the remaining producers are still moving concentrate rather than sulphate and would be directly exposed if the unbeneficiated export restriction takes effect before their plants come online. Logistics add a further layer of risk to that timeline. Spodumene concentrate is currently exported CIF China, routed predominantly via Beira in Mozambique and Durban in South Africa, and sulphate output from commissioned and future plants will likely need to move through the same corridors absent dedicated arrangements. ZETDC grid instability continues to affect processing-plant uptime, bearing directly on how quickly the remaining sulphate facilities can ramp to nameplate capacity once commissioned, while congestion at the Machipanda border post and along the Beitbridge Durban and Maputo rail corridors affects turnaround times independent of on-site production readiness. Even where sulphate capacity comes online on schedule, these constraints will shape whether the sector can convert commissioned capacity into consistent export volumes ahead of any tightened concentrate-export policy. On the fiscal side, the industry's effective tax burden is estimated at around 40% of sales revenue, comprising a 10% export tax on unbeneficiated lithium, 7% royalties, a 3% community development levy, a 1% MMCZ marketing fee and 15.5% VAT, before corporate income tax, payroll and foreign currency retention requirements are added. The association has called for a review, noting Zimbabwe's rate sits above global peers. Beyond lithium, the association has pointed to unrecovered value in tantalum, niobium and caesium from lithium operations, and government has since introduced mandatory mineral declaration and on-site assay laboratory requirements at producing mines. Specific loss and investment figures cited for this by-product stream were not independently attributed at the symposium and are not carried here pending primary sourcing. SMM View: Zimbabwe's beneficiation drive is currently a one-plant story, with Arcadia commissioned and exporting sulphate while Bikita, Kamativi and the fourth facility remain at construction or investment stage. The binding constraints on closing that gap are not purely financial grid reliability and corridor congestion at Machipanda and along the Beitbridge–Durban and Maputo routes will shape how quickly commissioned capacity converts into consistent export volumes. The roughly 40% effective tax burden remains a separate and likely more durable point of friction with Harare. SMM will confirm the committed investment figure and current export-policy status once further primary disclosures are available.

DeepSeek V4 Pro triggered a global price war among large language models with its ultra-low pricing. OpenAI slashed prices by 80%, Anthropic pulled back from markets outside China, Microsoft is evaluating integration, and Nvidia’s share of the Chinese GPU market plunged from 95% to 8%. Chinese models accounted for 60% of total call volume on OpenRouter, leading US models for 14 consecutive weeks.

Samsung unveiled the V10 Bonding V-NAND (BV-NAND) prototype and the industry’s first conceptual models of zHBM and zNAND-O architectures at the Future of Memory and Storage (FMS) conference held in Santa Clara, California, US. The BV-NAND chip features over 400 layers and adopts a novel wafer bonding architecture to enhance memory density and performance. Compared with the previous-generation V9 NAND, it delivers approximately 58% higher memory density, along with significant improvements in read, write, and I/O performance, meeting the demands of AI systems for higher-capacity, more energy-efficient memory solutions. For the zHBM and zNAND-O architectures, Samsung achieves greater data capacity in a smaller footprint by vertically stacking memory cells. Unlike conventional HBM, which is packaged side-by-side with AI processors, zHBM stacks memory vertically on top of the AI accelerator, thereby shortening data transfer distances, increasing bandwidth, and boosting energy efficiency. Samsung stated that its wafer bonding technology is expected to deliver more than 10 times the memory density of conventional HBM5, improve energy efficiency by 3 times, and reduce thermal resistance by more than half.

Independent tech journalist Tim Culpan pointed out that the A20 Pro processor wafers produced by TSMC’s N2 process are showing good yield and sufficient capacity, with processor production basically ready, and that this process is not a bottleneck affecting iPhone shipments. However, due to a severe shortage of DRAM supply, the completed processor wafers cannot be subsequently packaged, leading to inventory buildup and putting pressure on Apple and its supply chain’s stockpiling. With only weeks to go before the release of the iPhone 18 series and the foldable iPhone Ultra, Apple and its supply chain assemblers are urgently seeking more mobile DRAM chip supply. In terms of DRAM supply, Apple currently relies primarily on Micron Technology, while also sourcing from SK Hynix and Samsung.

Tesla (TSLA.O) announced that the "Terafab" project has officially landed in Grimes County, Texas. Jointly initiated earlier this year by SpaceX and Tesla, the project is the world’s largest chip manufacturing plan, aiming to bridge the massive gap between current global chip supply capabilities and future computing demand. In fact, the predecessor of Terafab broke ground in April at the north campus of the Texas Gigafactory—a new R&D wafer fab. Now the project has been officially upgraded and sited in Grimes County, where it will be developed into an advanced semiconductor wafer fab. SpaceX and Tesla’s combined chip demand is expected to exceed 1 terawatt (TW) of computing capacity, a scale that far surpasses current global supply capability. We deeply appreciate our existing chip suppliers and encourage them to expand their capacity wherever possible, but the widening supply-demand gap in the future is the core reason behind the existence of the Terafab project. Terafab aims to manufacture new computing capacity at unprecedented scale and speed. The project plans to build a vertically integrated factory with over 100 million square feet of manufacturing space, covering fabrication, packaging, and testing of advanced logic and memory chips. Consolidating these processes in a single facility facilitates rapid iterative improvements and accelerates the deployment of new computing capacity.

According to the Anhui Province Key Laboratory of Quantum Computing Chips, a joint research team from Origin Quantum and the University of Science and Technology of China has innovatively proposed a "parameter-space-extended controlled-phase gate (PSE-CZ)" scheme, successfully resolving the long-standing trade-off between speed and fidelity in superconducting quantum computing, and achieving a two-qubit quantum gate that significantly enhances accuracy while maintaining high-speed operation. The results have been published in Physical Review Letters, a top international physics journal, with experimental verification carried out on China's independently developed superconducting quantum computer "Origin Wukong".

According to "The Information," Nvidia (NVDA.O) is considering significant adjustments to the memory configuration of its next-generation graphics processor, Rubin Ultra, to address the shortage of high-end high-bandwidth memory chips. People familiar with the matter revealed that Nvidia has tested multiple versions of the Rubin Ultra graphics cards over the past few weeks. Some of these versions feature memory capacities lower than the initially announced specifications, primarily because the company may be unable to obtain sufficient high-end memory to meet the supply requirements of the original design. Although reducing the memory will have some impact on chip performance, Nvidia can compensate through other technical means. However, if other AI enterprises adopt this reduced-memory chip to run large AI models, they will need to use a larger number of chips than originally required.