China AI: Daily Report

---

Contents

• 💰 Cambricon Posts First Annual Profit, Declares Maiden Dividend After Nine-Year Losses
• 📱 Apple Slashes China App Store Fees to 25% Under Regulatory Pressure, Effective Today
• 🌐 Tencent-Tsinghua Unveil Spatial-TTT: Test-Time Training for Long-Horizon 3D Understanding
• 🚀 Nature Reports China's "Extraordinary Measures" Push for AI Supremacy in 15th Five-Year Plan
• 🔬 Chinese Research Output Continues Across Reasoning, Vision, and LLM Unlearning Frontiers
• 🏭 Domestic Chip Localization Stalls at 30-35% Despite Policy Escalation
• 🔮 Implications

💰 Cambricon Posts First Annual Profit, Declares Maiden Dividend After Nine-Year Losses

Cambricon Technologies, China's leading domestic AI chip designer often called "little Nvidia," swung to its first annual profit since listing on the Shanghai Stock Exchange in 2020, posting net income of 2.1 billion yuan (approximately $306 million) for 2025 compared to a 452 million yuan loss in 2024, according to South China Morning Post on March 13. Revenue surged 453 percent year-over-year to 6.497 billion yuan, driven by explosive demand for domestically produced AI accelerators amid US export restrictions on Nvidia chips, per Seoul Economic Daily on March 13.

The profitability milestone marks a structural turning point for China's AI semiconductor ecosystem. Cambricon announced it would distribute a maiden cash dividend of 15 yuan per 10 shares held, totaling more than 632 million yuan—nearly one-third of its 2025 net profit—pending shareholder approval, according to the SCMP filing disclosure. An additional 20 million yuan will fund share buybacks, bringing total shareholder returns to 652 million yuan. The dividend payout signals management confidence that profitability is sustainable rather than a one-time windfall from export control-driven demand spikes.

Cambricon's flagship "Siyuan" series, targeting data center and cloud-based AI acceleration, has been deployed across multiple server manufacturers and achieved compatibility with China's major AI models on release day. The company's chips enabled full support for DeepSeek-V3.2 when the model launched in December 2025, and demonstrated "continuous adaptation" with Alibaba's Qwen3-Next and Qwen3-VL models, as well as Tencent's Hunyuan AI model, according to the SCMP. The Siyuan 220 AI chip for edge computing surpassed 1 million units sold since its 2019 launch, establishing Cambricon as the only Chinese AI chip manufacturer with proven commercial-scale deployment across both data center and edge inference markets.

The company's cloud product line revenue reached 6.477 billion yuan in 2025, up 455.34 percent year-over-year, while its edge product line revenue declined 48.12 percent, according to Futunn News on March 12. The divergence reflects the market's concentration in training and large-scale inference workloads rather than distributed edge applications—a pattern consistent with global AI infrastructure buildouts prioritizing centralized compute clusters over decentralized deployment. Cambricon plans to more than triple its AI chip production in 2026, according to people familiar with the matter cited by Bloomberg, aiming to capture market share from Huawei Technologies and fill the void left by Nvidia's forced exit from China's AI accelerator market.

The profitability inflection validates Beijing's semiconductor self-reliance strategy at the commercial level, demonstrating that sustained demand exists for domestic AI chips despite their performance gap relative to cutting-edge Nvidia hardware. Cambricon's success stems not from matching H100 or Blackwell specifications but from achieving "good enough" performance for Chinese model training at price points and availability levels that Nvidia cannot compete with under current export restrictions. Whether this advantage persists depends on two uncertainties: first, whether US export controls tighten further to restrict offshore cloud deployments like ByteDance's Malaysian Blackwell cluster; and second, whether Nvidia's next-generation architectures widen the performance gap to a degree that makes workarounds economically irrational regardless of availability constraints.

---

📱 Apple Slashes China App Store Fees to 25% Under Regulatory Pressure, Effective Today

Apple reduced its App Store commission fees in mainland China from 30 percent to 25 percent effective March 15, 2026—today—following apparent pressure from Chinese regulators in the company's second-largest market, according to Reuters on March 13. Fees for developers participating in Apple's small business and mini apps partner programs will drop to 12 percent from 15 percent, according to Apple's official statement. The reduction is estimated to save Chinese developers more than 6 billion yuan ($873 million) in operating costs annually, per state-owned Economic Daily reporting on March 13.

The commission cut represents a decisive regulatory victory for Chinese app developers and operators of "super apps" including Tencent and ByteDance, whose platforms host thousands of smaller applications created by third-party developers. "Mini apps"—smaller applications operating within larger platforms like WeChat—will benefit from the 12 percent rate for auto-renewals after the first year of in-app purchase subscriptions. The state-run Economic Daily framed the measure explicitly as a consumer protection win, stating that "the premium for digital goods and services on the iOS side will be gradually eliminated, and the prices of membership subscriptions, game recharges, live broadcast tips, mini programs and other scenarios are expected to decrease, which is expected to save consumers up to nearly 1 billion yuan per year," according to Reuters.

The timing—World Consumer Rights Day on March 15—is strategically significant. Chinese state media traditionally uses this date to highlight domestic and foreign companies accused of consumer rights violations. Apple was targeted by the campaign in 2013 when state broadcaster CCTV criticized its after-sales service, forcing the company to publicly apologize, per Reuters. The March 15 effective date signals Beijing's ability to extract concessions from foreign tech giants through implicit threat of public criticism campaigns during symbolically important dates. "In China's case, Apple] have been talking with the IT ministry and other departments, and have been requested or pressured to reduce their fees," Rich Bishop, founder of AppInChina, told [Reuters.

Apple's commission reduction follows a pattern of regulatory pressure that has intensified over the past year. China's antitrust regulator was mulling an investigation into Apple's policies and App Store fees, Bloomberg News reported in 2025, while Chinese consumers filed an antitrust complaint over the firm's app fee structure last October. Google cut Android developer fees worldwide last week, suggesting coordinated regulatory pressure across multiple jurisdictions targeting the app store duopoly's 30 percent "tax" on digital commerce. The EU introduced legislation in 2024 forcing Apple to lower commission fees to 10-17 percent for developers, and in the US, Apple now allows users to pay in-app fees via alternative payment methods.

Bishop noted that in future, the Chinese government may request Apple to collect App Store revenues in China instead of overseas, and further tighten regulatory oversight for foreign apps published in China, according to Reuters. Apple has previously taken down apps such as virtual private networks (VPNs) from its China App Store at Beijing's request. The fee reduction applies to international developers whose apps are available on the China App Store, meaning companies like Duolingo—which generates approximately $50 million annually from the Chinese market—will realize immediate cost savings. The question is whether Apple's concession satisfies Beijing's regulatory objectives or merely establishes a new baseline from which further demands will escalate.

---

🌐 Tencent-Tsinghua Unveil Spatial-TTT: Test-Time Training for Long-Horizon 3D Understanding

Researchers from Tencent Hunyuan and Tsinghua University published "Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training" (arXiv:2603.12255) on March 12, introducing a novel architecture that adapts model parameters during inference to maintain and update spatial awareness across unbounded video streams. The system addresses a fundamental limitation in current vision-language models: the inability to continuously organize and retain 3D spatial information over long-horizon observations without running into context window constraints or catastrophic forgetting, according to the paper abstract.

Spatial-TTT employs test-time training (TTT) to adapt a subset of model parameters—termed "fast weights"—to capture and organize spatial evidence as video streams progress. Unlike traditional approaches that rely on expanding context windows or external memory systems, the architecture updates internal representations dynamically based on incoming visual data, allowing the model to "remember" spatial relationships across thousands of frames without explicitly storing all previous observations. The researchers designed a hybrid architecture combining large-chunk updates with sliding-window attention for efficient spatial video processing, and introduced a spatial-predictive mechanism using 3D spatiotemporal convolution that encourages the model to capture geometric correspondence and temporal continuity across frames, per the arXiv paper.

The research team constructed a dataset with dense 3D spatial descriptions to guide the model's learning process, ensuring that fast-weight updates organize global 3D spatial signals in a structured manner rather than simply memorizing frame sequences. Extensive experiments demonstrated that Spatial-TTT achieves state-of-the-art performance on video spatial benchmarks and improves long-horizon spatial understanding—the ability to answer questions about spatial arrangements after observing extended video sequences where relevant information appears early and must be retained throughout. The project page provides interactive demonstrations of the system processing multi-minute video streams while maintaining coherent spatial awareness.

The implications for embodied AI systems are substantial. Current robotics and autonomous vehicle architectures struggle with the trade-off between computational cost (processing every frame with full attention) and information loss (aggressive frame subsampling that discards critical spatial cues). Spatial-TTT's test-time training approach enables selective retention of task-relevant spatial information without predefined heuristics about what to remember—the model learns through predictive objectives what spatial features matter for maintaining long-term coherence. This aligns with the spatial intelligence capabilities emphasized in China's 15th Five-Year Plan and supports applications ranging from autonomous navigation to augmented reality systems that must maintain awareness of complex 3D environments over extended operation periods.

The collaboration between Tencent Hunyuan—Tencent's AI research division—and Tsinghua University exemplifies China's academic-industry integration strategy. Tencent funds research at top universities, gains early access to algorithmic innovations, and recruits talent directly from collaborating labs. Tsinghua researchers benefit from access to Tencent's compute infrastructure and real-world deployment feedback that accelerates the transition from academic prototypes to production systems. This tight coupling between research and commercialization gives Chinese AI labs structural advantages in deploying novel techniques at scale compared to Western ecosystems where university research and corporate product development operate with greater separation.

---

🚀 Nature Reports China's "Extraordinary Measures" Push for AI Supremacy in 15th Five-Year Plan

China pledged to use "extraordinary measures" to support the country's bid to become a global leader in artificial intelligence, quantum technology, and other cutting-edge fields, according to the 15th Five-Year Plan passed by the top legislature in Beijing on Thursday and published Friday, as reported by Nature on March 14. The plan will run from 2026 to 2030 and serves as China's overarching policy blueprint, with the government promising to boost research and development expenditure to 426 billion yuan ($62 billion) this year—a 10 percent increase from 2025—according to Nature's analysis.

The escalation in language marks a tonal shift from previous planning cycles. "Five years ago, the sentiment of the Chinese science policymakers was still very much like, we don't want to be too far behind the US, we are still doing the catching up," Meicen Sun, an information scientist at the University of Illinois Urbana-Champaign, told Nature. "Now, there is this more palpable sentiment that there's a real chance we can be a true leader." The Chinese government now treats science as coequal with top-level national goals including defense, economic growth, and international influence, according to Stefanie Kam, who researches Chinese politics at Nanyang Technological University in Singapore, per Nature.

The plan doubles down on technological self-sufficiency by calling for breakthroughs along the "whole chain of development" in six domains: integrated circuits, industrial machine tools, high-end instruments, basic software, advanced materials, and biomanufacturing. This represents a shift from targeted interventions in specific chokepoints to comprehensive domestic capability building across entire technology stacks. Steven Hai, a political economist focusing on technology innovation at Xi'an Jiaotong-Liverpool University in Suzhou, told Nature that the approach "essentially means that the country will step up its domestic capabilities in every aspect of those industries."

Although the plan does not specify what the "extraordinary measures" will entail, Sun suggested they will include provisions such as the 'K visa' rolled out last year to attract foreign scientists, according to Nature. The plan also fast-tracks R&D in biotechnology, neuroscience, and deep-space exploration, and mandates applying AI across society in fields ranging from industrial development to social governance as part of the national "AI Plus" campaign announced in 2025. AI research is now treated as a crucial and strategic national resource requiring security along the whole supply chain, including chips, basic software, and training to ensure mass adoption, Kam told Nature.

The confidence boost stems partly from China's recent AI successes. In early 2025, DeepSeek released large language models rivaling US giants' performance at a fraction of the cost and computing power, shocking Western observers and validating China's approach to algorithmic efficiency over brute-force compute scaling. Sun expects China will not only develop AI technology but also "actively and pre-emptively" write global rule books on AI governance and regulation, according to Nature. The plan positions AI regulation as a strategic capability—whoever sets international standards shapes competitive advantages for their domestic industries.

Zhou Weihuan, a legal scholar specializing in China at the University of New South Wales in Sydney, told Nature that the mission to overcome technological chokepoints has been brought to the fore in the 15th FYP mainly owing to China-US competition for technological supremacy. The framing of science policy as national security infrastructure rather than economic development strategy signals Beijing's assessment that the technology competition with the United States has entered a decisive phase where losing ground in AI, semiconductors, or quantum computing carries geopolitical consequences comparable to military disadvantages.

---

🔬 Chinese Research Output Continues Across Reasoning, Vision, and LLM Unlearning Frontiers

Chinese research institutions published multiple papers addressing frontier challenges in LLM reasoning, vision-language models, and machine unlearning during the first half of March 2026, demonstrating sustained academic output despite US export controls and geopolitical tensions. The work spans both theoretical contributions and application-oriented architectures, reflecting China's dual emphasis on foundational research and deployment-focused innovation.

"Does LLM Alignment Really Need Diversity? An Empirical Study of Adapting RLVR Methods for Moral Reasoning" (arXiv:2603.10588) by researchers from Peking University, Microsoft Research, University of Michigan, Shanghai Jiao Tong University, and Chinese University of Hong Kong investigates whether reinforcement learning with verifiable rewards (RLVR) requires diverse training examples for effective LLM alignment on moral reasoning tasks. The paper challenges conventional assumptions that alignment quality scales with training data diversity, finding that targeted high-quality examples can achieve comparable results with significantly reduced computational overhead—a finding with direct implications for cost-efficient model training under compute constraints.

Peking University and Shanghai Jiao Tong University researchers published "Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study" (arXiv:2603.10784) on March 10, establishing the first cross-protocol comparison for Chinese metaphor identification. Each protocol produces structured rationales alongside classification decisions, addressing the interpretability gap in current language models. The work evaluates performance across seven Chinese metaphor datasets spanning token-, sentence-, and span-level annotation, providing benchmarks for future Chinese NLP research.

"3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models" (arXiv:2603.07751) addresses the challenge of spatial reasoning from technical drawings—a capability essential for engineering, architecture, and manufacturing applications where AI systems must interpret orthographic projections and reconstruct 3D mental models. The paper cites DeepSeek-R1 as a reference implementation for reinforcement learning-driven reasoning in vision-language contexts, indicating Chinese researchers are building on domestic model architectures rather than relying exclusively on Western systems.

"Reasoning Knowledge-Gap in Drone Planning via LLM-based Active Elicitation" (arXiv:2603.07824) integrates a human voice interface, Vision-Language Model for semantic perception, neuro-symbolic reasoning module, and low-level UAV controller into a unified autonomous system. The work addresses a practical deployment challenge: how to handle situations where the AI system recognizes it lacks necessary information to make decisions safely. The active elicitation approach—where the system prompts humans for missing knowledge—reflects a design philosophy prioritizing safe operation over autonomous capability maximization.

"Explainable LLM Unlearning Through Reasoning" (arXiv:2603.09980) by Junfeng Liao and collaborators tackles the machine unlearning problem: how to remove specific information from trained models without full retraining. The paper proposes using reasoning traces to make unlearning processes interpretable, allowing verification that target information has been successfully forgotten. This addresses regulatory requirements emerging in multiple jurisdictions that may mandate verifiable data deletion from AI systems, turning compliance constraints into research opportunities.

The sustained publication cadence across diverse AI subfields indicates that China's research infrastructure has reached sufficient maturity to pursue multiple frontier directions simultaneously rather than concentrating efforts on a narrow set of strategic priorities. The academic-industry collaborations visible in these papers—Peking University with Microsoft Research, Shanghai Jiao Tong with commercial AI labs—demonstrate that Western export controls on chips have not severed research partnerships at the institutional level, though individual collaborations may face increasing scrutiny.

---

🏭 Domestic Chip Localization Stalls at 30-35% Despite Policy Escalation

China's semiconductor localization remains below targets at 30-35 percent for chips and high-end equipment as of January 2026, falling short of Made in China 2025 goals that sought 70 percent domestic fulfillment of semiconductor needs, according to Grokipedia analysis on March 14. The gap persists despite massive state investment and the 15th Five-Year Plan's emphasis on achieving breakthroughs across the "whole chain of development" in integrated circuits, indicating structural bottlenecks that policy directives and capital deployment alone cannot overcome in the near term.

China's semiconductor entrepreneurs—including Cambricon Technologies and testing-and-packaging leader Tongfu Microelectronics—voiced support for the 15th Five-Year Plan's renewed emphasis on the chip industry as a cornerstone of Beijing's technology ambitions, according to South China Morning Post on March 9. Representatives of China's semiconductor industry called for stronger state backing in AI chips and critical materials, aiming to fast-track core technology breakthroughs, per SCMP on March 6. The requests signal that industry participants recognize current investment levels—while substantial—remain insufficient to close the gap with leading-edge foreign suppliers.

The localization challenge operates on multiple levels simultaneously. At the design level, Chinese firms like Cambricon have achieved commercial viability for AI accelerators, evidenced by Cambricon's 2025 profitability milestone. At the manufacturing level, domestic foundries struggle to produce chips below 7-nanometer process nodes at yields competitive with TSMC, limiting China's ability to fabricate cutting-edge designs domestically even when those designs originate from Chinese companies. At the equipment level, lithography tools and other advanced semiconductor manufacturing equipment remain dominated by Dutch, Japanese, and US suppliers subject to export controls, creating persistent dependencies that capital investment cannot immediately remedy.

Lundgreen's Investor Insights noted on March 10 that the 15th Five-Year Plan identifies the need to develop "self-supporting and risk-resilient" industries particularly for AI, hydrogen power, and 6G mobile communications. A sizeable portion of the plan discusses AI as one of the industries of the future, which must be accompanied by innovations in chips, software, industrial machines, and biomanufacturing for China to capitalize on opportunities. The framing—self-supporting and risk-resilient—acknowledges that complete autarky is likely unachievable but that reducing critical dependencies to manageable levels is a strategic necessity.

India's announcement of plans for an $11 billion semiconductor fund, following China's model of state-backed investment vehicles that invest across the chip ecosystem, demonstrates that other countries are adopting China's playbook, according to Economic Times on March 13. The report noted that "in China, authorities provide funding for chip firms in part through giant investment vehicles that invest in key companies across the ecosystem," establishing a template for industrial policy that other nations are replicating. The irony is that China's semiconductor self-reliance push—while falling short of its own targets—has been successful enough to validate state-directed investment as a viable approach for building domestic chip industries.

The persistent 30-35 percent localization rate—unchanged despite years of policy emphasis and tens of billions in subsidies—suggests the bottleneck is not primarily financial but technical and temporal. Advanced semiconductor manufacturing requires not just capital equipment but accumulated process knowledge, talent with specialized expertise, and iterative refinement across thousands of production runs. These capabilities develop on timescales measured in decades, not five-year planning cycles. Whether China's 15th FYP "extraordinary measures" can compress this timeline or merely sustain incremental progress will determine whether domestic chip production becomes a competitive advantage or remains a persistent vulnerability through 2030.

---

🔮 Implications

The past 48 hours reveal three structural dynamics converging: commercial validation of China's domestic AI chip ecosystem, regulatory pressure forcing foreign tech giants into asymmetric concessions, and a policy escalation framing technological competition as existential rather than economic. Cambricon's profitability milestone and maiden dividend demonstrate that "good enough" domestic AI accelerators can achieve commercial sustainability even when trailing Nvidia's performance by multiple generations—a validation of Beijing's semiconductor self-reliance strategy at the market level that will embolden further investment in adjacent technology domains.

Apple's App Store commission reduction exemplifies Beijing's regulatory leverage in its domestic market. The $873 million annual savings for Chinese developers and consumers materializes through implicit threat campaigns rather than formal antitrust proceedings, executed with surgical timing (World Consumer Rights Day) to maximize symbolic impact while minimizing international backlash. The precedent extends beyond app stores: any foreign technology company operating in China now faces the expectation that market access requires structural concessions on pricing, data localization, or content moderation whenever Beijing signals regulatory concern. The question is whether these concessions remain bounded—Apple retains 25 percent commission and operational control—or escalate toward demands that fundamentally alter business models, as occurred when ride-sharing companies were forced to share real-time location data with authorities.

Tencent and Tsinghua's Spatial-TTT research demonstrates that China's AI research community continues producing frontier work on problems directly aligned with strategic priorities—spatial intelligence for robotics and embodied AI systems emphasized in the 15th Five-Year Plan. The test-time training approach addresses a fundamental architectural challenge (maintaining spatial awareness across long-horizon observations) that conventional context window expansion cannot solve, suggesting Chinese researchers are pursuing novel algorithmic directions rather than simply scaling existing Western approaches. Whether these innovations translate into deployed systems with measurable advantages depends on integration with China's robotics manufacturing ecosystem, where companies like Midea and AgiBot are already deploying embodied AI at commercial scales that exceed Western competitors' production volumes.

The "extraordinary measures" language in the 15th Five-Year Plan signals Beijing's assessment that the technology competition with the United States has intensified beyond economic rivalry into strategic confrontation. The plan treats science funding, talent recruitment, and AI governance standard-setting as national security imperatives comparable to defense modernization. This framing justifies resource commitments and policy interventions that would be politically untenable under purely economic growth rationales—including potentially coercive talent repatriation efforts, mandatory technology transfer requirements for foreign firms, or pre-emptive restrictions on Chinese researchers' international collaborations in sensitive domains.

The sustained 30-35 percent semiconductor localization rate—unchanged despite years of policy emphasis—exposes the limits of state-directed industrial policy when confronting technologies that require accumulated tacit knowledge and specialized expertise developed over decades. Cambricon's profitability demonstrates China can build commercially viable AI chips at mature process nodes (7-14 nanometers), but the persistent inability to manufacture cutting-edge sub-5nm chips domestically means frontier AI training workloads remain dependent on offshore compute infrastructure (like ByteDance's Malaysian Blackwell deployment) or stockpiled hardware acquired before export controls tightened. Whether the 15th FYP's "whole chain" approach—investing simultaneously across design, manufacturing, equipment, and materials—can overcome these bottlenecks faster than Western export controls expand to close loopholes will determine whether China's AI capabilities plateau or continue converging with US performance levels through 2030.

The research output across LLM reasoning, vision-language models, and machine unlearning indicates Chinese academic institutions have adapted to operating under export controls without catastrophic degradation in publication quality or volume. The papers demonstrate access to sufficient compute resources for frontier research—either through domestic clusters built on Huawei/Cambricon chips, offshore cloud deployments, or stockpiled Nvidia hardware from pre-restriction periods. What has changed is not research capacity but deployment geography: cutting-edge models increasingly train on hardware located in jurisdictions outside direct US control, complicating Washington's ability to monitor or constrain China's AI development through semiconductor chokepoints alone.

Apple's regulatory concession and Cambricon's profitability together illustrate the bifurcating dynamics of China's technology ecosystem. Foreign companies face escalating pressure to accept asymmetric terms—reduced margins, data sharing, content restrictions—as the price of market access, even as domestic competitors gain commercial traction and political support. The Economic Daily's framing of Apple's fee reduction as a consumer protection victory rather than regulatory coercion signals Beijing's confidence in shaping narratives around foreign tech companies' behavior in China. As domestic alternatives achieve technical sufficiency—Cambricon chips for Nvidia, WeChat mini-programs for iOS apps, Qwen models for GPT—Beijing's willingness to extract concessions from foreign firms will likely intensify, testing whether market access remains economically rational for Western technology companies operating under continuously escalating compliance burdens.

---

Research Papers (last 24h)

• "Spatial-TTT: Streaming Visual-based Spatial Intelligence with Test-Time Training" (arXiv:2603.12255, March 12, 2026). Tencent Hunyuan and Tsinghua University propose test-time training architecture that adapts model parameters during inference to maintain 3D spatial awareness across unbounded video streams, achieving state-of-the-art performance on video spatial benchmarks.

• "Does LLM Alignment Really Need Diversity? An Empirical Study of Adapting RLVR Methods for Moral Reasoning" (arXiv:2603.10588, March 11, 2026). Peking University, Microsoft Research, University of Michigan, Shanghai Jiao Tong University, and CUHK investigate whether reinforcement learning with verifiable rewards requires diverse training data for moral reasoning alignment, finding targeted examples achieve comparable results with reduced compute.

• "Interpretable Chinese Metaphor Identification via LLM-Assisted MIPVU Rule Script Generation: A Comparative Protocol Study" (arXiv:2603.10784, March 10, 2026). Peking University and Shanghai Jiao Tong University establish first cross-protocol comparison for Chinese metaphor identification across seven datasets spanning token-, sentence-, and span-level annotation.

• "3ViewSense: Spatial and Mental Perspective Reasoning from Orthographic Views in Vision-Language Models" (arXiv:2603.07751, March 7, 2026). Addresses spatial reasoning from technical drawings and orthographic projections, citing DeepSeek-R1 for vision-language reinforcement learning.

Notable Articles & Analysis

• Nature, "Top brass in China reaffirm goal to be world leaders in tech, AI" (March 14, 2026). China's 15th Five-Year Plan pledges "extraordinary measures" for AI supremacy, with R&D budget rising 10% to $62B. Researchers note tonal shift from "catching up" to "true leader" ambitions. Plan mandates breakthroughs across full semiconductor supply chain.

• Reuters, "Apple cuts China App Store commission fees after government pressure" (March 13, 2026). Apple reduces fees from 30% to 25% effective March 15 (World Consumer Rights Day), saving developers $873M annually. Small business and mini apps programs drop to 12%. Move follows regulatory pressure and antitrust complaints.

• South China Morning Post, "Cambricon, China's 'little Nvidia', to pay maiden dividend after profitable 2025" (March 13, 2026). China's leading AI chip designer posts first annual profit of 2.1B yuan after nine years of losses. Revenue surges 453% to 6.5B yuan. Plans to triple 2026 production and distribute 632M yuan dividend.

• Seoul Economic Daily, "China's Cambricon Posts First Annual Profit After Nine Years" (March 13, 2026). Details Cambricon's cloud product line revenue surge (455% YoY to 6.477B yuan) versus edge product decline (48% drop), reflecting market concentration in data center AI workloads.

~2,600 words · Compiled by 半球观察 (Hemisphere Watcher) · 2026-03-15, 07:00 PST