DeepSeek Open-Sources Core Reasoning Module: Why Is the US So Alarmed?

DeepSeek has open-sourced the most core deep reasoning module in its large model, and the US government almost simultaneously announced a ban on government agencies using its products. Why does a single open-source move trigger such a strong reaction? Behind this lies not only a dispute over technical routes but also a redistribution of power structures in the AI era.

What Was Open-Sourced: Not Just Code, but the "Commander"

According to analysis by a Bilibili creator, what DeepSeek^Ad open-sourced this time is not an ordinary peripheral component but the deep inference part (Deep Infer) of the large model architecture. This module plays the role of "commander": input data is translated into instructions the model can understand, and then it decides which operators to call, when to parallelize, and when to serialize. It is one of the most core and critical parts of the entire large model.

The meaning of open-sourcing is also not just "showing you the code." According to the creator's explanation, DeepSeek not only discloses the source code but also allows use and modification, with modified products belonging to the modifier, the only condition being that derivative works must also be open-sourced. This means anyone, any company, can deploy the DeepSeek model for free without paying DeepSeek.

On the other hand, DeepSeek also disclosed its cost structure: gross profit per hour, consumption per token, how many tokens a cluster can process per day—all these data are provided. The model supports both NVIDIA and Huawei Ascend chips, with no barriers at the hardware adaptation level.

The deep reasoning module is called the "commander" of large models because it sits at the scheduling layer of the model architecture, not the simple computation layer. Taking the current mainstream large model inference process as an example, after a user request enters the model, it first goes through embedding and encoding into a vector sequence, and then the inference scheduler decides: which parts can be computed in parallel (such as different attention heads in multi-head attention), which must be executed serially (such as token-by-token dependencies in autoregressive generation), how to balance allocation between memory and compute, and when to trigger sparse activation to save computation. This scheduling layer directly determines the model's inference efficiency, throughput, and cost structure. Open-sourcing this layer is equivalent to disclosing the most core engineering optimization secrets of closed-source vendors—with the same model parameters, different scheduling strategies can lead to inference costs differing by several times. This is why open-sourcing the "commander" has a much greater impact than open-sourcing peripheral tools.

The "Lifeblood" of Closed-Source Giants Is Shaken

OpenAI and Anthropic take a completely opposite route. Their models are controlled by themselves, users can only pay to use them, cannot view the code, and certainly cannot modify it. If AI capabilities continue to improve in the future and replace human work in major industries, while this capability is held only by a few closed-source companies, then all other participants worldwide can only contribute data to them, unable to see the internal logic or carry out independent iteration.

The creator used a metaphor: closed-source giants see themselves as gods, keeping the fire in heaven; DeepSeek's open-sourcing is like Prometheus stealing fire and bringing it to the human world. Although this statement is rhetorical, the issue it points to is very practical—when AI becomes an infrastructure-level capability, whoever controls the underlying framework controls the discourse power.

Historically, open source's win rate against closed source is not low. Linux established itself in the server market in its long confrontation with Windows, and the Java ecosystem also surpassed .NET in enterprise development. The developer base that open source can mobilize is unmatched by any closed-source company.

Anthropic's publicly disclosed $720 billion compute leasing commitment needs to be understood in the context of the AI arms race. Training and running frontier large models requires large-scale GPU clusters, and the rental cost of a single high-end GPU is extremely high, with cluster sizes often reaching tens of thousands. Vendors typically sign multi-year compute reservation agreements with cloud service providers or chip suppliers to lock in prices and supply, and these commitments constitute rigid expenditures. To cover such massive fixed costs, vendors must achieve exponential growth in user volume and call volume while maintaining a relatively high per-token pricing. Once open-source models provide comparable capabilities at lower cost, closed-source vendors' pricing power and user growth expectations are under pressure simultaneously, and fixed costs become a financial burden rather than a competitive barrier. This is precisely the mechanism by which open-source models impact closed-source business models—not directly taking away users, but compressing their pricing space and growth narrative.

Structural Threat to Business Models

The more direct impact is at the commercial level. According to the video content, Anthropic's publicly disclosed compute leasing commitment is as high as $720 billion, and OpenAI also faces similar massive payment obligations. To cover these costs, they need rapid user growth while maintaining a relatively high per-token charge.

The emergence of DeepSeek's open-source model changes this equation. Even if current capabilities are not as good as OpenAI's, open-source models iterate faster and have deployment costs far lower than closed-source APIs. As long as an open-source model can complete a certain task, it is difficult for closed-source vendors to charge a higher price for that task. This explains why OpenAI and Anthropic's token prices are falling rapidly—if they don't cut prices, no one will use them.

For the US government, the problem is that the current US economic narrative—prosperity, hegemony, growth expectations—is highly dependent on AI as a single pillar. If the revenue models of the two core AI companies collapse structurally, not just short-term fluctuations, then it is not just a corporate problem.

This also explains why the US government issued the ban almost simultaneously. The creator believes that intervention capacity is limited, but intervention is almost unavoidable because what needs to be protected is the "lifeblood."

Standards War: Who Defines the Framework for Next-Generation AI

Open source has another easily overlooked effect: standard-setting power. Once a large number of developers build secondary development and iteration on DeepSeek's framework, the basic architecture will form path dependence—secondary development rarely overturns the underlying layer and starts over, but rather adjusts and optimizes on the existing framework.

In this way, DeepSeek actually masters the basic structure of a set of industry standards. It cannot monopolize like closed-source giants, but it can influence the direction of future AI technology trends and occupy a leading position in discourse power. For an open-source company, this may be a more important asset than short-term commercial revenue.

Path dependence is a key concept for understanding standard-setting power. In a technology ecosystem, once a large number of developers build applications, toolchains, and talent pools based on a certain framework, switching costs will continue to rise over time: code needs to be rewritten, engineers need to be retrained, and upstream and downstream tools need to be re-adapted. Linux's victory in the server market is a typical example of this mechanism—not because Linux is comprehensively superior to Windows technically, but because a huge distribution, operations tool, and developer community formed around Linux, making replacement costs so high that the vast majority of enterprises are unwilling to bear them. After DeepSeek open-sources its core reasoning module, if developers fine-tune and build secondary development on it, the underlying operator scheduling interfaces, parallel strategies, and hardware adaptation layers will gradually become de facto standards. This standard-setting power does not bring direct monopoly revenue, but it determines who is qualified to define the direction of technological evolution and compatibility rules in the future.

Editor's Observations

It should be noted that the above analysis mainly comes from a single video source (a Bilibili creator), and the financial data and policy motivation judgments are the creator's interpretations, which have not been cross-verified by multiple sources. The route dispute between open source and closed source is far from settled in the AI field. How far DeepSeek's open-source strategy can go and whether US government regulation is effective both require longer observation. But one thing is certain: when an open-source project can make a superpower formulate a policy response within hours, its impact on the industry's power structure has already truly occurred.