DeepSeek Ignites Open Source vs. Closed Source Debate in Silicon Valley, Splitting Tech Giant Alliances

DeepSeek's Open Source Model Shakes Silicon Valley, AI Route Debate Reaches Fever Pitch

When DeepSeek released its open source large language model, the shockwaves through the American tech industry far exceeded expectations. The model not only put pressure on U.S. tech stocks — with some key companies seeing significant market cap evaporation — but more importantly, it tore open a long-existing yet never-so-publicly-exposed rift within Silicon Valley: Should artificial intelligence follow an open source or closed source path?

DeepSeek's impact was so powerful because of its breakthrough technical approach. The model employs cutting-edge technologies such as Mixture of Experts (MoE) architecture, achieving performance close to top-tier closed source models while dramatically reducing training costs. The core idea behind MoE is splitting the model into multiple expert sub-networks, activating only a subset during each inference pass. This significantly reduces computational resource consumption while maintaining a large total parameter count. This means high-quality large language models can be trained even in compute-constrained environments — a point particularly crucial for Chinese AI companies facing chip export controls.

According to the New York Times, China started later in frontier AI but is rapidly catching up through the open source route. A U.S. tech think tank previously estimated the AI gap between the U.S. and China at roughly six to seven months; it has now narrowed to approximately four months. This rapid pursuit has left U.S. regulators and some companies anxious.

Here's a telling detail: China's open source strategy isn't purely a technical choice — it carries clear strategic intent. By using open source models to attract global developers — especially those in the Global South — into its ecosystem, whoever defines the content within these models gains the initiative in future information discourse.

OpenAI Holds Firm on Closed Source; Microsoft and NVIDIA Champion Open Ecosystems

The open source versus closed source debate has persisted in Silicon Valley for years, but the accelerating catch-up of Chinese open source models like DeepSeek has brought it fully into the open. The opposing camps are clearly drawn.

To understand this debate, one must first clarify the fundamental difference between open and closed source models. An open source model publicly releases its weight parameters — all the numerical values the model learned through training — allowing anyone to download, use, modify, and redistribute them. In a stricter sense, open source also includes publishing training data and training code. Closed source models only provide services through API interfaces, with users unable to access underlying parameters. Open source models offer advantages in transparency, customizability, and vibrant community innovation, but they also carry risks of misuse — once model weights are public, developers cannot control downstream applications. Closed source models are easier to govern but also imply potential technology monopolization and lack of external auditing.

The Closed Source Camp: Restricting Openness in the Name of Safety

One side, represented by OpenAI and Anthropic, argues that certain advanced models pose too high a risk to be publicly released and should be strictly controlled by companies. They have expressed concerns about Chinese open source AI models to Washington regulators, with U.S. Treasury Secretary Bessent and Trump administration tech advisors participating in related discussions.

The Open Source Camp: Building a Healthy AI Ecosystem

The other side includes Microsoft, NVIDIA, Meta, and Palantir, among other tech giants. On the 24th (local time), NVIDIA CEO Jensen Huang posted for the first time on social platform X: "The world needs both the most advanced closed source models and the most advanced open source models."

Just nine minutes later, Microsoft's CEO immediately responded: "Open source software is essential for a healthy AI ecosystem." The two also jointly endorsed an industry open letter supporting the development of open weight models, which received co-signatures from NVIDIA, Microsoft, Meta, Palantir, and other leading companies.

Notably, the "Open Weight Model" mentioned in the letter represents a compromise between fully open source and fully closed source. It publicly releases a model's inference weights, allowing others to use and fine-tune the model, but typically does not disclose complete training data, training code, or training processes. Meta's LLaMA series is a classic example of open weight models. This approach is seen as a pragmatic choice balancing innovation promotion with risk control, and represents the greatest common denominator that mainstream Silicon Valley tech companies can currently agree on.

Apple vs. Micron Chip Standoff Reflects Deeper Silicon Valley Contradictions

The open vs. closed source debate is not Silicon Valley's only internal conflict. In the memory chip sector, Apple and Micron Technology are engaged in an intense power struggle — with Trump caught squarely between the two tech giants.

Memory chips are indispensable core components in all electronic devices, primarily divided into two categories: DRAM (Dynamic Random Access Memory) and NAND Flash. DRAM handles temporary data storage during device operation, while NAND Flash provides persistent data storage. The global memory chip market has long been dominated by South Korea's Samsung and SK Hynix, with U.S.-based Micron Technology as the only domestic company capable of competing. As AI large models drive surging demand for High Bandwidth Memory (HBM), memory chips have evolved from traditional consumer electronics components to strategic materials for AI infrastructure, with supply chain security directly affecting a nation's position in the AI race.

According to the Wall Street Journal, Apple executives have been actively lobbying Trump, Commerce Secretary Lutnick, Treasury Secretary Bessent, and other officials in recent weeks, seeking permission to use Chinese-manufactured memory chips in Apple products sold outside the United States. Apple argues this would help alleviate the global memory chip supply crunch and reduce costs for American consumers.

However, Micron — America's only major memory chip manufacturer — strongly opposes this, warning that allowing Chinese companies to supply U.S. tech giants could repeat the historical devastation of America's steel and manufacturing industries, destroying the domestic memory chip industry at a critical stage of development.

The essence of this standoff is the deep contradiction within American tech companies between commercial interests and industrial protection. Apple needs a stable, low-cost supply chain, while Micron needs policy barriers to maintain its survival space.

AI and Historical Truth: A Cognitive War That Cannot Be Ignored

Beyond technological competition, AI is being drawn into a more covert cognitive war. China's Ministry of State Security published a commentary on its WeChat official account titled "AI Cannot Fabricate Truth; Tampering Cannot Wash Away Blood Stains," directly calling out Japanese right-wing forces for using AI to fabricate disinformation such as "the Nanjing Massacre is a lie" and other falsifications of Japan's wartime history in China.

The article cites Japan's Asahi Shimbun reporting that Japanese right-wing forces are using major crowdsourcing platforms like Crowd Works to recruit paid workers at scale, mass-producing "criticize China" videos using AI. Clients provide detailed operational guides, and workers need only input simple prompts into AI software to produce a history-distorting video in minutes.

Data shows that since 2015, Japan has invested over 56 billion yen under the banner of "overseas strategic information dissemination" for anti-China propaganda.

This phenomenon reveals a deep-seated danger of the AI era: If training data is contaminated, AI will carry bias and distorted perceptions when answering historical questions. From a technical mechanism standpoint, large language models derive their knowledge from training data — typically text corpora scraped at scale from the internet. Models lack the independent ability to judge factual truth; they essentially learn statistical patterns and linguistic regularities from data. If the proportion of false narratives about a particular historical event increases significantly in training corpora, the model will tend to generate content consistent with those false narratives. This phenomenon is known as "Data Poisoning" and is an important research direction in AI safety. More dangerously, once AI-generated false content spreads widely across the internet, it gets scraped as training data by subsequent models, creating a vicious cycle of "disinformation → AI generation → redistribution → retraining."

The internet is a primary channel through which younger generations learn about history. Once false historical content proliferates at scale through AI technology, it can easily mislead adolescents whose historical knowledge is still incomplete, causing historical memory to be diluted and distorted across generations. Even more concerning is that false information generated abroad could flow back into China's domestic internet space, potentially contaminating the training data of domestic large models.

The Strategic Game Behind Technical Routes Is Far From Over

From DeepSeek igniting Silicon Valley's open vs. closed source debate, to the Apple-Micron chip standoff, to AI being drawn into historical cognitive warfare — these seemingly disparate events all point to the same core proposition: In the AI era, technical route choices are never merely technical issues; they are extensions of strategic competition.

Open source versus closed source is fundamentally a contest over discourse power and ecosystem dominance; chip supply choices tip the scales between industrial security and commercial interests; and AI's ability to shape historical truth reminds us to guard against technology being misused to distort collective memory.

For ordinary users, AI will likely become a "dictionary" and core information source in daily life. Therefore, in this age of information explosion, maintaining independent thinking and discerning truth from falsehood will be more important than ever before. History should be verified, and verifying history requires the support of evidence piece by piece — a responsibility that no AI technology can replace.