The AI Power Struggle: Should Superintelligence Be Monopolized or Decentralized for All?

A Single Tweet Sparks a Deeper Conversation About AI Power

Recently, a tweet about the future power dynamics of AI ignited widespread discussion on social media. Its core message struck at the most sensitive nerve of today's AI development:

"The power of the future will either be held by a few companies that use and control superintelligence, or it will be held by the people in a decentralized way. We believe the latter is far better, and our goal is to put truly general AI into the hands of as many people as possible."

Though brief, this statement precisely outlines the most fundamental power struggle of the AI era — when superintelligence becomes reality, who gets to control it? This is not merely a technical question; it is a political and ethical proposition that concerns the trajectory of human civilization.

To grasp the weight of this proposition, we first need to clarify what "superintelligence" actually means. The concept was first systematically articulated by philosopher Nick Bostrom in his 2014 book Superintelligence: Paths, Dangers, Strategies, referring specifically to an AI system that significantly surpasses the best human minds in virtually all cognitive domains — including scientific creativity, social skills, and general reasoning. It differs from today's "narrow AI" (which excels only at specific tasks) and "Artificial General Intelligence (AGI)" (which can perform all human cognitive tasks), representing instead an intelligence leap with the capacity for self-improvement and recursive enhancement. Organizations like OpenAI and DeepMind widely believe that AGI could be achieved within the 2030s, with superintelligence potentially following closely behind. It is precisely this approaching "capability singularity" that has transformed the question of control from a philosophical discussion into an urgent real-world issue.

The Reality of AI Power Centralization: Computing and Data Are Rapidly Concentrating

The Matthew Effect of Computing, Data, and Talent

The current state of the AI industry is sobering. The cost of training a frontier large model has skyrocketed from millions to billions of dollars. A handful of players — OpenAI, Google DeepMind, Anthropic, Meta — have near-monopolies on the most advanced model capabilities. This centralization trend is no accident — barriers in computing power, data, and talent have together built a moat that is nearly impossible to cross.

This Matthew Effect has deep economic roots. The computing power required to train frontier large models is growing exponentially: GPT-3's training cost approximately $4.5 million, GPT-4 is estimated to have exceeded $100 million, and the next generation of models may break the $1 billion mark. The dual effects of "economies of scale" and "data flywheels" create a positive feedback loop where the strong get stronger — more users generate more data, more data trains better models, and better models attract more users. Meanwhile, the AI chip market is highly concentrated (NVIDIA's H100/H800 dominate), and cloud computing infrastructure is divided among AWS, Azure, and Google Cloud. This means even well-funded challengers face supply-chain-level constraints. This structural centralization makes "winner-take-all" the default logic of the AI era.

What's even more noteworthy is that these companies don't just control the models themselves — they also control API pricing, usage policies, and safety strategies. When millions of developers and businesses worldwide depend on a handful of companies' AI interfaces to build products, a new form of "digital dependency" has quietly taken shape.

What Does Control of Superintelligence Actually Mean?

If we take the concept of "superintelligence" seriously — an AI system that surpasses humans in virtually all cognitive tasks — then the entity controlling it would wield unprecedented power. That power could be used for scientific discovery, curing diseases, and solving climate issues, but it could equally be used for surveillance, manipulation, and suppression.

History has repeatedly shown that the concentration of power never automatically leads to benevolent use. The question of who controls superintelligence is, at its core, an extension of humanity's age-old power distribution logic into the AI era.

The Decentralized AI Vision: Ideals and Challenges of the Open-Source Movement

Open-Source Models Are Changing the Game

"Putting general AI into the hands of as many people as possible" — the most direct path to realizing this vision is AI open source. The AI open-source movement traces back to around 2015, when the open-sourcing of deep learning frameworks like TensorFlow and PyTorch laid the foundation for technological democratization. The real turning point came in February 2023, when Meta released LLaMA (Large Language Model Meta AI), making model weights approaching GPT-3's level publicly available for the first time. Subsequently, open-source models like Mistral 7B, Falcon, Qwen, and DeepSeek^Ad emerged in succession, forming an ecosystem that stands toe-to-toe with closed-source models.

Meta's LLaMA series, Mistral, Stability AI, and other projects have already proven that open-source models can approach or even match the performance of closed-source models. When model weights are publicly available, any individual or organization with sufficient hardware can run, fine-tune, and deploy their own AI systems, free from the policy constraints of a single company. It's worth noting that "open source" in the AI field is contested — some models only release weights (weights-available) rather than complete training data and code, drawing criticism as "faux open source." Truly open-source AI requires the full disclosure of training data, training code, model architecture, and weights. Projects meeting this standard remain rare, with EleutherAI's Pythia series and BLOOM being notable examples.

This decentralization is not just a technical distribution — it's a redistribution of power. When a startup in Africa can build a localized medical diagnostic tool using an open-source model, or when an independent researcher can explore AI safety issues on their own GPU, technological democratization truly happens.

The Real-World Challenges Facing Decentralized AI

However, the vision of decentralized AI also faces formidable challenges:

• Amplified Safety Risks: Broadly distributing powerful AI capabilities inevitably increases the potential for malicious use. Risks such as bioweapon design, deepfakes, and automated cyberattacks cannot be ignored.
• The Computing Divide Persists: Even if models are open-sourced, the computing power required to run frontier models remains expensive. Decentralization may simply shift from "model monopoly" to "compute monopoly," still leaving ordinary users and small teams out in the cold.
• AI Governance Vacuum: Decentralization means no single entity is accountable for AI behavior, creating enormous gray areas in legal and ethical terms. Once an open-source model is used for illegal purposes, assigning responsibility becomes an extremely thorny problem.

A Third Way: Finding a Governance Balance Between Openness and Safety

The real answer may not be a binary either-or choice. What we need is a form of "orderly openness."