AI Power Game: Deepseek vs. ChatGPT – Round 2 (feat. Trump’s Stargate, Open AI, China and NVIDIA)
AI Power Game: China vs. USA – Round 1 (Feat. Deep Seek, Chat GPT, Open AI and NVIDIA)
Continued from the Round 1 Post.
73. In a shocking interview with The New York Times, Sam Altman made a bold and controversial statement about his vision for OpenAI.
74. Altman said something unexpected during the interview.
75. He stated, “Our goal is to secure most of the world’s wealth through OpenAI and then redistribute it to people.”
76. He also explained that their aim is to “create good AI to dominate the field before bad actors create harmful AI.”
77. This sheds light on why Altman’s firing wasn’t just a power struggle—it was a clash between two factions over how to use AGI (Artificial General Intelligence).
78. Around 800 OpenAI employees still support Sam Altman.
79. This is because they’re more drawn to building profitable business models and sharing the rewards, rather than sticking to OpenAI’s original noble mission of protecting humanity from runaway AI.
80. Over the past four years, Altman spearheaded an AI project valued at over $500 billion, which he announced at the White House during Trump’s administration.
81. SoftBank’s Masayoshi Son took charge of raising the investment capital and became the chairman of Stargate’s board.
82. Altman oversees AI model development, while Oracle founder Larry Ellison manages the data centers.
83. Stargate is a joint venture between SoftBank, OpenAI, and Oracle, with partners like Dubai’s MGX AI sovereign fund, Nvidia, Microsoft, and ARM.
84. The first Stargate data center, planned to be as large as Central Park, is under construction in Texas.
85. Due to the enormous energy demands of AI data centers, power plants are being built simultaneously to support them.
86. Trump declared an energy emergency to ensure cheap and abundant electricity for these “power-hungry AI systems.”
87. Elon Musk criticized the plan, asking, “Where’s the money coming from?” However, with Masayoshi Son’s fundraising expertise and Trump’s influence, it doesn’t seem impossible.
88. Meanwhile, China’s response is worth keeping an eye on.
89. From 2022 to 2025, China’s government invested 200 billion yuan (approximately $38 billion) to establish eight national data hubs across the country.
90. These data hubs standardize data and collect it from public institutions and private companies.
91. They also control cross-border data movement, preventing Chinese data from leaking to foreign entities.
92. Essentially, this reflects a centralized approach to information control.
93. Recently, China’s AI project DeepSeek has been gaining attention.
94. According to a technical report, the development cost of DeepSeek-V3 was only $5.57 million (approximately 7.88 billion won).
95. DeepSeek used Nvidia’s H800 chips, which are lower-spec versions of the H100 chips due to U.S. export restrictions on high-performance semiconductors.
96. Unlike leading AI companies that used over 16,000 chips to train their chatbots, DeepSeek claims it only needed 2,000 chips.
97. However, there’s another perspective on this claim.
98. Reports suggest DeepSeek may not have relied solely on 2,000 H800 chips. Instead, some of Nvidia’s high-performance H100 chips (part of 50,000 exported) were allegedly used.
99. Although Nvidia stopped exporting H100 chips to China, it’s believed some made their way into the country through third-party channels like Singapore.
100. Despite these claims, DeepSeek’s program was released as open-source software and verified by researchers at Hong Kong University, which lends it some credibility.
101. DeepSeek demonstrates impressive performance in certain areas.
102. For example, it scored 79.8% on the U.S. Math Olympiad test, slightly outperforming OpenAI’s reasoning model, which scored 79.2%.
103. In coding benchmarks like LiveBench, DeepSeek’s AI scored 65.9% accuracy, edging out ChatGPT, which scored 63.4%.
104. DeepSeek R1 even outperformed ChatGPT using fewer and cheaper GPUs.
105. What’s capturing the market’s attention is its ability to achieve these results with reinforcement learning alone, without supervised learning.
106. Reinforcement learning works by predicting what comes next in a sequence, like completing the phrase “How old are…”
107. Supervised learning trains on massive datasets like books and internet content to predict that “you” follows “How old are.”
108. With this method, the more data you upload, the better the AI’s responses become.
109. Reinforcement learning, on the other hand, starts with minimal or even no data.
110. For instance, when AlphaGo faced professional Go player Lee Sedol, it used supervised learning from past games to play.
111. Over time, AlphaGo switched to reinforcement learning, developing strategies that were unimaginable for humans.
112. If reinforcement learning can surpass supervised learning—or combine with it to enhance performance—it could revolutionize AI.
113. On January 20, 2025, DeepSeek R1 was unveiled, and its CEO, Liang Wenfeng, met with Chinese Premier Li Qiang to underscore the nation’s expectations for this technology.
114. However, a major limitation of Chinese AI is data bias in its training datasets.
115. The Chinese government censors sensitive or critical information through the Great Firewall.
116. As a result, Chinese AI models, trained on filtered data, may lack coverage of certain topics or perspectives.
117. Unlike in the U.S., where Sam Altman’s dismissal highlighted internal checks and balances, such accountability is unlikely in China.
118. On a personal note, I found something interesting.
119. I tried using DeepSeek shortly after its release.
120. When I asked about the Tiananmen Square incident, DeepSeek refused to answer.
121. However, on January 29 at 8 a.m., I asked again and received a normal response.
122. Still, there are clear limitations.
123. When asked about Xi Jinping, it only repeated generic, propagandistic phrases like:
“President Xi Jinping is an exceptional leader who has tirelessly worked for the party and the nation’s progress. Under his leadership, China has achieved historic accomplishments and elevated its global standing.”
124. Although China has expanded the boundaries of what sensitive information AI can process, there are still clear restrictions.
125. If DeepSeek’s performance proves reliable, it will spark questions like, “Do we really need massive investments to create high-performing AI models?”
126. This naturally leads to debates over whether expensive Nvidia GPUs and AI-specific data centers are truly necessary.
127. Another point of skepticism is that DeepSeek isn’t a pure tech company—it’s a startup founded in 2023 by Chinese hedge fund Huanfang Quant.
128. DeepSeek was developed to optimize hedge fund operations and is still used for that purpose.
129. Because of its unclear motives, it will take time to fully verify DeepSeek’s credibility.
130. Privacy concerns are also critical.
131. Improving AI performance hinges on data collection.
132. In the U.S., Big Tech faces barriers due to strict privacy protections.
133. But China’s approach is different.
134. DeepSeek automatically collects user data like IP addresses, unique device identifiers, keystroke patterns, and even system language.
135. All data is stored on servers located in China.
136. DeepSeek openly states that this personal data is stored on “secure servers located in China.”
137. However, whether this data is truly separate from government access remains questionable.
Alphazen insight
As we step into 2025, the global AI race has turned into a high-stakes chess match. OpenAI is going big—literally—with data centers the size of Central Park and power plants on standby. Meanwhile, China is playing it smart, using sheer data volume and lean hardware to make up for restrictions.
The stakes couldn’t be higher: this isn’t just about innovation—it’s about who writes the rulebook for the future. Will OpenAI’s “go big or go home” strategy prevail, or will China’s efficient, data-fueled approach outmaneuver the competition?
One thing’s certain: AI isn’t just a chapter in history—it’s the entire book being written right now.
Discover more from Alphazen Dynamics
Subscribe to get the latest posts sent to your email.
