DeepSeek Funding Doubles to 100 Billion Yuan, ByteDance Seed Establishes LLM Safety Department

The AI industry's funding boom and safety governance are accelerating in tandem. On one side, DeepSeek's funding scale has doubled to 100 billion yuan within half a month; on the other, ByteDance has become the first in China to establish a standalone safety and alignment department within its large model team. Meanwhile, progress in mass production of Huawei's Kirin chips and Tesla's Optimus robot is redrawing the boundaries of hardware competition.

DeepSeek Funding Doubles: Why Industrial Capital Is Rushing In

According to Caijing Magazine, DeepSeek^Ad will complete its second round of funding in the near future, with the planned amount increased from the previously planned 50 billion yuan to approximately 100 billion yuan—doubling in less than half a month. The pre-money valuation remains at approximately 500 billion yuan.

New investors in this round include Shixi Capital, SMIC Juyuan, CPE Yuanfeng, Legend Capital, Geely, and others, primarily industrial investors. CATL, Tencent, IDG Capital, Lisi Capital, and Shixiang Capital, which joined the first round, are also on this round's list. Among them, CATL plans to add approximately 10 billion yuan, having contributed about 5 billion yuan in the first round.

An LP from a participating fund stated that the situation is basically consistent with the facts, but the current funding transaction is still dynamically changing. The aggressive injection of industrial capital indicates that large model capabilities are shifting from the technology validation stage to the industry chain integration stage—investors value not just the model itself, but the efficiency gains from combining the model with their own business operations.

ByteDance Seed Establishes Safety and Alignment Department, a Key Step in China's LLM Governance

ByteDance Seed recently sent an internal email announcing the establishment of the "Seed Model-Alignment & Preparedness" department, led by Yan Lin, reporting to Seed head Wu Yonghui. The original Responsible AI team and some Chat team members will transfer to the new department.

The department has two core goals: first, to align model behavior with human intent and safety constraints, enhancing the safety capabilities of Agent Models; second, to guard against frontier capability risks such as CBRN (Chemical, Biological, Radiological, Nuclear). One detail worth noting: previously, safety and alignment work in domestic large model teams was mostly incorporated into the company's overall safety compliance system. ByteDance Seed is the first frontier AI lab in China to establish a standalone safety and alignment department.

This organizational restructuring sends a clear signal: as model capabilities rapidly approach the AGI threshold, safety is no longer just a compliance issue but a core engineering problem requiring independent research efforts and an independent reporting line. Spinning the safety team out of the compliance system and having it report directly to the Seed head means safety considerations will intervene earlier in the model R&D process rather than being remedied after the fact.

CBRN is an abbreviation for Chemical, Biological, Radiological, and Nuclear—four categories of high-risk capabilities. The international security community commonly uses this framework to assess the proliferation risks of weapons of mass destruction. In the AI context, CBRN risks specifically refer to the potential misuse of large models to obtain or deduce sensitive knowledge in these high-risk domains, such as assisting in the synthesis of toxic chemicals, enhancing pathogen characteristics, or optimizing nuclear device designs. Mainstream frontier labs typically conduct CBRN capability evaluations after pre-training is complete and before model release, employing layered controls through red-team testing, knowledge threshold filtering, and refusal fine-tuning. As Agent capabilities strengthen, models can autonomously invoke tools, decompose and chain multi-step tasks, and the safety boundary expands from "whether it will reveal dangerous knowledge" to "whether it can autonomously complete dangerous processes." This is precisely one of the backgrounds behind ByteDance including CBRN prevention among the new department's core objectives.

Huawei Kirin Chips and Tesla Optimus: Hardware Mass Production Race Accelerates

At the 2026 Huawei Terminal Innovation Technology Communication Meeting, Xu Songyu, Director of Huawei's Semiconductor Business Wireless Terminal Chip Product Management Department, revealed that the next-generation Kirin 2027 chip is ready and all test results meet expectations. Terminal products based on Kirin chips have cumulatively shipped over 100 million units since 2023, and the Kirin 9050 series has successfully entered mass production.

Huawei also stated that the logic folding evolution path over the next 10 years is sustainable, but which specific technology to adopt at each stage is still under gradual discussion.

The robotics sector is also accelerating. Tesla's core supplier Xinjian Transmission explicitly stated in its response to the ChiNext listing review inquiry letter that the hardware design of Tesla's Optimus has been basically frozen and finalized, with plans to begin mass production before the end of 2026. Tesla has converted the former Model S and Model X production lines at its Fremont factory into Optimus V3 production lines, with a long-term capacity target of 1 million units per year. The Texas Gigafactory is planned to break ground in 2027, with a long-term capacity target of up to 10 million units per year after completion. A small number of Optimus units are already operating internally at the Fremont factory, mainly for AI training and data collection.

Logic folding is an evolutionary approach in advanced process technology and chip design, referring to the coordinated optimization of transistor structures, metal interconnects, and packaging methods to continue compressing logic unit area and energy consumption beyond a single process node. Unlike simply relying on lithography precision to shrink line widths, it emphasizes joint innovation across design, process, and packaging layers, with typical directions including backside power delivery, gate-all-around transistors, 3D stacking, and chiplet interconnects. Huawei's statement that "the logic folding evolution path over the next 10 years is sustainable" is essentially expressing a performance improvement route that does not fully depend on the most advanced lithography equipment: even if process nodes are constrained, energy efficiency and integration density can still be gained through continuous folding at the architecture and packaging levels. The difficulty of this route lies in the fact that process selection, yield control, and ecosystem adaptation at each stage require long-term validation, which is why Huawei also acknowledged that which technology to adopt at each stage is still under gradual discussion.

Yin Zhiyao Warns of AI Bubble: 90% of Domestic Semiconductor Equipment Makers Will Be Eliminated

Yin Zhiyao, Chairman and General Manager of AMEC, issued a warning about the current AI investment boom. He pointed out that looking back at the annual increase and decrease curve of global semiconductor industry CAPEX from 1984 to 2026, the fluctuations are extremely dramatic—industry capital expenditure can increase by 100% in a year or decline by 30% in a year.

Regarding the industry landscape, Yin Zhiyao judged that the number of domestic semiconductor equipment companies is enormous—conservatively estimated at no fewer than 500, and at most close to 1,000. Referring to the U.S. landscape where only a few dozen remain after multiple rounds of consolidation, ultimately only a few dozen may survive in China, meaning 90% of companies will be eliminated. He believes leading companies need to transform from single-product companies to platform companies, possessing etching, thin film, and metrology/inspection capabilities simultaneously to provide complete solutions.

This judgment forms an interesting contrast with DeepSeek's funding boom: capital is concentrating toward the top, and leading companies need not just funding but also the technological depth to weather industry cycles.

CAPEX stands for Capital Expenditure, referring to spending by enterprises on acquiring, expanding, or upgrading long-term assets. In the semiconductor industry, it mainly manifests as investments in building factories, purchasing equipment such as lithography machines, and production line renovations. Semiconductors is a typical strongly cyclical industry, and the annual fluctuation of CAPEX often far exceeds changes in end demand: during upswings, manufacturers race to expand capacity, with annual increases reaching 100%; once demand peaks, capital expenditure quickly contracts by more than 30%, forming a cycle of "overcapacity—price decline—investment contraction." Yin Zhiyao cited the global semiconductor CAPEX curve from 1984 to 2026 precisely to illustrate that this industry has never escaped dramatic fluctuations over four decades. For equipment makers, orders surge when downstream fabs expand, while during contractions they face delivery delays, extended receivables, and even order cancellations—this is the underlying logic behind his judgment that small and medium-sized manufacturers will struggle to weather the cycle and most will ultimately be eliminated.

Alibaba Qwen Launches on Two Fronts: Office E-Commerce Edition and Personal Agent

Alibaba's AI office product "Qwen Office" is developing an e-commerce version for multi-platform merchants, planned for release ahead of the Double Eleven pre-sale period. The project is coordinated by Qwen Office CEO Chen Yusen, with nearly 100 employees currently involved. The e-commerce version's homepage features 5 sections: Double Eleven zone, business analytics, product management, material production, and marketing promotion. Its current form is closer to a cross-platform analytics and production workstation, and it cannot yet directly adjust merchants' ad spending plans.

Meanwhile, the Qwen Personal Agent "Qwen Go!" is also about to be released, helping users complete tasks such as order claims, car maintenance, sports appointments, and shopping price comparisons. When users need it, Qwen can autonomously query and operate various authorized platforms, compare options, call service providers, and complete target tasks.

International News Roundup

Musk stated on social media that once China masters the manufacturing method for a certain product, its capacity expansion capability often exceeds that of the rest of the world combined, citing lithium batteries and the solar industry as examples. This discussion was built on AI analyst Melvin's analysis of ChangXin Memory—ChangXin Memory currently holds approximately 9.5% revenue share in the global DRAM market, making it the world's fourth-largest DRAM manufacturer.

OpenAI is in negotiations with a startup to transfer the shelved video generation AI Sora to that company. The startup was founded several months ago by Hollywood heavyweight Jeffrey Katzenberg, Bill Peebles who led Sora research, and former Dropbox CFO Sujay Jaswa.

Anthropic announced a $100 million investment to launch an FDE training program. Selected participants first attend a 4-day intensive on-site training, then undertake actual projects using Claude tools over 12 weeks. Anthropic hopes to train 10,000 engineers by the end of 2027.

Additionally, Chinese AI scientist Yao Shunyu, Senior Staff Research Scientist at Google DeepMind, is reportedly starting an RSI venture. The new company is positioned to develop customized models for enterprises, valued at approximately $2 billion, with major investors from overseas.

From funding to safety governance to hardware mass production, the competitive dimensions of this AI race are expanding from single-model capabilities to an all-around contest of organizational capability, supply chain depth, and safety infrastructure.

DRAM stands for Dynamic Random Access Memory, the mainstream memory type in computers and servers responsible for temporarily storing running data. Together with NAND flash, it forms the two pillars of memory chips. The DRAM market has long been dominated by Samsung, SK Hynix, and Micron, with their combined share persistently above 90%. It is extremely difficult for new entrants to break through the triple barriers of patent walls, process yield, and customer certification. ChangXin Memory, as mainland China's major DRAM manufacturer, reaching approximately 9.5% revenue share and ranking fourth globally means it has achieved large-scale shipment capability in standard DDR products and is approaching first-tier international levels at some process nodes. This share change is not only a breakthrough at the corporate level but will also affect global memory supply-demand and price cycles—which is why Musk's comments on manufacturing capacity expansion have sparked widespread discussion.

One-line Summary

DeepSeek funding doubles to 100 billion yuan, ByteDance establishes safety alignment department, Kirin and Optimus mass production accelerate—the AI race turns to multi-dimensional competition.

Paragraph Summary

This edition's highlights show the AI industry advancing simultaneously along three main lines: funding, governance, and hardware. DeepSeek's second funding round doubled from 50 billion to approximately 100 billion yuan in half a month, with a pre-money valuation of about 500 billion yuan; CATL plans to add 10 billion yuan, with industrial capital as the main force. ByteDance Seed established China's first independent safety and alignment department, reporting directly to the Seed head, focusing on Agent safety and CBRN risks. Huawei revealed that the next-generation Kirin 2027 chip is ready, with the Kirin 9050 series in mass production; Tesla's Optimus V3 design is near freezing, with plans for mass production by end of 2026. Yin Zhiyao warned of a bubble, predicting 90% of domestic semiconductor equipment makers will be eliminated. Alibaba Qwen is simultaneously advancing an office e-commerce edition and a Personal Agent. Internationally, news involves ChangXin Memory, Sora's fate, and Anthropic's training program.