分类： technology

In a groundbreaking revelation, Chinese AI developer DeepSeek disclosed that it spent a mere $294,000 to train its R1 model, a figure significantly lower than the costs reported by its U.S. counterparts. This disclosure, published in a peer-reviewed article in the journal Nature on September 18, 2025, is poised to reignite discussions about China’s role in the global AI race. The Hangzhou-based company, which has largely remained out of the public eye since its January 2025 announcement of lower-cost AI systems, detailed that the R1 model was trained using 512 Nvidia H800 chips over 80 hours. The article, co-authored by DeepSeek founder Liang Wenfeng, also revealed that the company utilized Nvidia A100 GPUs in the preparatory stages of development, a fact that had not been previously disclosed. This revelation comes amidst ongoing scrutiny from U.S. companies and officials regarding DeepSeek’s access to advanced AI chips, particularly after the U.S. imposed export controls on high-performance chips to China in October 2022. Despite these challenges, DeepSeek has managed to attract top talent in China, partly due to its operation of an A100 supercomputing cluster, a rarity among domestic firms. The company’s cost-effective approach to AI development has already had a significant impact on global markets, prompting investors to reevaluate the dominance of established AI leaders like Nvidia.

In a landmark move, Nvidia has announced a $5 billion investment in Intel, solidifying its position as one of the largest shareholders in the U.S. chipmaker. This strategic partnership, unveiled on September 18 in San Francisco, comes weeks after the U.S. government acquired a 10% stake in Intel, signaling a concerted effort to bolster domestic semiconductor capabilities. The collaboration aims to enhance AI and computing technologies, posing a significant challenge to industry giants like TSMC and AMD. Nvidia’s investment, priced at $23.28 per share, reflects confidence in Intel’s potential despite its recent struggles. The deal excludes Intel’s foundry business but focuses on joint development of PC and data center chips, leveraging Nvidia’s proprietary technology for faster chip-to-chip communication. This alliance could reshape the competitive landscape, particularly in AI servers, where Nvidia and Intel’s combined offerings may outpace rivals like AMD and Broadcom. Intel’s CEO, Lip-Bu Tan, has pledged to streamline operations and align factory capacity with demand. The partnership underscores a broader trend of U.S. tech firms uniting to counter global competition, with Nvidia and Intel poised to drive innovation in the next era of computing.

In a groundbreaking move, Chinese tech giant Huawei has broken its silence to reveal a comprehensive product roadmap for its chips and computing power systems, marking its first public strategy to compete with global leader Nvidia. The announcement, made on September 18, 2024, outlines Huawei’s plans to introduce three new Ascend series chips over the next three years, starting with the Ascend 950 in the first quarter of 2025. The Ascend 950 will come in two variants: the 950PR, optimized for inference and recommendations, and the 950DT, designed for model training and decoding. Huawei also revealed that the Ascend 960 and 970 will significantly boost computing power and memory capacity, with the 970 expected to surpass Nvidia’s offerings. Despite U.S. export restrictions limiting Huawei’s access to TSMC, the company has developed its own high-bandwidth memory (HBM) technology, which it claims is more cost-effective than SK Hynix and Samsung’s HBM3E and HBM4E. Huawei’s cluster computing systems, such as the Atlas 900 A3 SuperPoD, already rival Nvidia’s advanced products, and the company plans to launch the Atlas 950 SuperPod in Q4 2026, boasting 6.7 times more computing power than Nvidia’s NVL144 system. Additionally, Huawei’s Kunpeng CPU chip series, first introduced in 2019, will see new iterations in 2026 and 2028, accompanied by the TaiShan 950 SuperPod for general-purpose computing. This bold strategy underscores Huawei’s determination to establish itself as a major player in the global semiconductor and AI markets.

SoftBank Group Corp. and OpenAI’s highly anticipated joint venture to deliver artificial intelligence (AI) services to corporate clients in Japan has encountered significant delays, according to an insider familiar with the matter. Initially slated for launch this summer, the venture, named SB OpenAI Japan, is now expected to provide an update on its progress in November. The source, who requested anonymity due to the confidential nature of the details, cited prolonged preparations as the primary cause for the setback. SoftBank confirmed that preparations are ongoing but refrained from commenting further, while OpenAI has yet to respond to inquiries. The venture was announced in February by SoftBank CEO Masayoshi Son and OpenAI CEO Sam Altman, with ownership shared between OpenAI and a newly established SoftBank entity. At a June shareholder meeting, Junichi Miyakawa, CEO of SoftBank’s telecom unit, had targeted the end of July for the venture’s launch, though specific product offerings remain under discussion. This initiative marks a resurgence in Son’s bold investment strategy, following a period of retrenchment due to underperforming tech investments. Meanwhile, SoftBank’s $500 billion Stargate project, aimed at developing data centers in the U.S., has also faced delays due to protracted negotiations and location-related decisions, as disclosed by CFO Yoshimitsu Goto last month.

Kansai Electric Power Co Inc (9503.T) is set to resume feasibility surveys at its Mihama nuclear power station in western Japan this November, marking Japan’s first significant move towards constructing a new reactor since the Fukushima disaster. The surveys, initially announced in July, will assess the viability of building a successor unit at the Mihama site, which was previously halted following the 2011 catastrophe. The comprehensive studies, expected to continue until around 2030, will include geological investigations both inside and outside the plant premises to identify suitable locations, followed by detailed assessments of topography and ground conditions. Kansai Electric emphasized that the findings will be evaluated alongside advancements in light-water reactor technology, regulatory policies, and the overall business environment before making a final decision on the new unit. The initial surveys for a replacement reactor at Mihama began in late 2010 but were suspended after the Fukushima incident. The renewed efforts will adhere to stricter safety regulations implemented post-disaster. Currently, only the No. 3 unit at Mihama remains operational, as the No. 1 and No. 2 units are being decommissioned. Since 2022, Kansai Electric has been collaborating with Mitsubishi Heavy Industries (7011.T) and other utilities, including Kyushu Electric Power (9508.T), Shikoku Electric Power (9507.T), and Hokkaido Electric Power (9509.T), to develop a next-generation 1.2-gigawatt advanced light-water reactor, known as ‘SRZ-1200.’ This consortium aims to enhance nuclear technology while ensuring compliance with modern safety standards.

Jensen Huang, CEO of Nvidia, has voiced his disappointment following reports that China has instructed its leading technology firms to cease purchasing Nvidia’s artificial intelligence (AI) chips. Speaking to BBC News, Huang emphasized the importance of global access to advanced technology, stating, ‘The advance of human society is not a zero-sum game.’ Huang, who is among the tech executives accompanying US President Donald Trump on his state visit to the UK, acknowledged the competitive ambitions of both nations, noting, ‘President Trump wants America to win, and President Xi wants China to win, and it’s possible for both of them to.’ He expressed confidence that ongoing discussions would resolve the issue. This development comes after Nvidia, the world’s leading chipmaker, faced a temporary ban on selling its most advanced chips to China, which was later reversed by Trump in July. However, Nvidia must now allocate 15% of its Chinese revenues to the US government under a unique agreement. The Financial Times recently reported that China’s Cyberspace Administration has directed tech companies to halt the use of Nvidia chips specifically designed for the Chinese market, causing Nvidia’s shares to drop by over 1% in premarket trading. Huang reiterated his support for the US in addressing geopolitical challenges and pledged to convey the same message to Trump if questioned during the state banquet in the UK. The US and China have been engaged in trade talks in Europe this week, with China’s market regulator accusing Nvidia of violating anti-monopoly laws. As Nvidia continues to play a pivotal role in the global AI boom, China is striving to challenge US dominance in the AI sector by developing its own chips. Major Chinese tech firms like DeepSeek, Tencent, and Alibaba, which had previously ordered Nvidia chips, are now affected by the purchase halt. Meanwhile, Nvidia has announced significant investments in the UK, including supplying chips to the Stargate UK data center in collaboration with OpenAI, Arm, and NScale.

In a significant development, the United States and China have reached an agreement to ensure TikTok continues its operations in the US. President Donald Trump announced the deal, stating that he will confirm the details with Chinese President Xi Jinping during an upcoming call. TikTok, owned by Chinese company ByteDance, had faced the threat of a shutdown unless it sold its US operations. However, Trump has repeatedly postponed the ban since its initial announcement in January, with the latest extension pushing the deadline to December 16.

Under the negotiated deal, TikTok’s US business will be controlled by an investor consortium, including tech giant Oracle, private equity firm Silver Lake, and venture capital firm Andreessen Horowitz. The new US entity will see American investors holding approximately 80% of the stake, with US nationals dominating the board, including one government-appointed member. US users will transition to a new app, currently in testing, which will utilize content-recommendation algorithms licensed from ByteDance—a key factor in TikTok’s success.

Oracle will maintain its existing agreement to host TikTok servers within the US, addressing concerns about data security. The deal is expected to be finalized within the next 30 to 45 days. Earlier, a US trade delegation in Madrid announced a ‘framework’ agreement with China, which China confirmed but emphasized that no deal would compromise its firms’ interests. Wang Jingtao, deputy head of China’s cyberspace administration, highlighted that the agreement includes licensing algorithms and intellectual property rights, subject to Chinese government approval.

Trump’s reversal on TikTok marks a shift from his initial stance during his first term, when he called for the app’s ban. The US Supreme Court upheld a 2024 law banning TikTok unless ByteDance sold its US operations, citing national security concerns. ByteDance has consistently denied sharing user data with the Chinese government, maintaining that its US operations are independent. The deadline for a sale has been extended four times, with the latest delay set to expire on December 16.