# ESSAY: DeepSeek V4, Open Source as a Geopolitical Weapon

> On April 24, 2026, DeepSeek released on Hugging Face two models under an MIT license: DeepSeek V4-Pro, a Mixture-of-Experts architecture with 1.6 trillion total parameters and 49 billion active parameters, and DeepSeek V4-Flash, a lighter variant. The MIT license means anyone — any individual, any company, any government, any army — can use, modify, or deploy these models witho

- Author: Maxime Marquette
- Publisher: MadMax (https://mad-max.co)
- Section: Essay
- Published: 2026-06-27
- Canonical URL: https://mad-max.co/en/article/essai-deepseek-v4-l-open-source-comme-arme-geopolitique

## Introduction: April 24, 2026 and what it means

### A date that will be studied in AI history courses

On **April 24, 2026**, **DeepSeek** released on **Hugging Face** two models under an **MIT license**: **DeepSeek V4-Pro**, a **Mixture-of-Experts** architecture with **1.6 trillion total parameters** and **49 billion active parameters**, and **DeepSeek V4-Flash**, a lighter variant. The MIT license means anyone — any individual, any company, any government, any army — can use, modify, or deploy these models without paying **DeepSeek**, without reporting to the Chinese government, and without technical restrictions.

Two benchmarks from that release deserve to be cited and remembered. On **SWE-bench**, the software engineering evaluation, **V4-Pro scores 80.6%** — beating **Claude 3.7 Sonnet**. On **GPQA Diamond**, the advanced scientific reasoning benchmark, it scores **90.1%**. At **$3.48 per million tokens**, **V4-Pro** is between four and seven times cheaper than its American equivalents at **$15 to $25 per million tokens**. This is not a marginal improvement. This is a structural reorientation of the global **AI** cost curve. And it was engineered on **Huawei Ascend chips** — not on **Nvidia**.

### What this release changes about the global AI landscape

*I want to be honest about my reaction when I read these numbers: it was not the reaction of a technology journalist covering a product launch. It was the reaction of someone watching a geopolitical event. The fact that a Chinese **state-backed** company can release a **frontier-tier** AI model under MIT license, cheaper than anything American, on non-Nvidia hardware — that is not just a tech story. That is a signal.*

## The political economy of open source as a weapon

### Why "open source" is not a neutral term here

Open source software has a long tradition as a commons — code freely shared for the benefit of all, built on principles of collaborative improvement and democratic access to technology. The MIT license is one of the most permissive in existence. But when an **open-source model** is released by a company backed by a **$4 billion round from China's state AI fund** (documented by Tech Times in May 2026), the "open" in open source acquires a different valence.

The political economy of **DeepSeek V4**'s release is this: **China** absorbs the development cost — through state-backed capital, through access to talent from Chinese universities, through government coordination of compute resources — and then gives the result away for free to the rest of the world. The effect is to undercut the subscription model of American AI companies, to flood the global market with capable free models, and to make **Chinese AI** the default infrastructure for governments and companies that cannot afford **GPT-5** or **Gemini**. Open source as market capture. Open source as **soft power**. Open source as a weapon.

### The strategic logic of free distribution

*I am not claiming that every engineer at DeepSeek is thinking about geopolitics when they write code. I am claiming that the institutional logic of this release — state-backed, MIT-licensed, priced to undercut — is strategically coherent in a way that goes beyond commercial optimization. Whether it is consciously designed as a **geopolitical weapon** or whether it operates as one regardless of intent does not change the strategic effect.*

## Technical performance: what do these numbers actually mean?

### Reading benchmarks with appropriate skepticism

**SWE-bench** measures the ability of a model to solve real software engineering tasks — actual GitHub issues, actual code fixes. An **80.6% score** means that in 80.6% of test cases, **V4-Pro** produces a working solution. This is not a toy benchmark. It reflects something real about the model's engineering capabilities. **GPQA Diamond** tests graduate-level scientific reasoning — questions that require synthesizing multiple domains. A **90.1% score** is competitive with human expert performance in those domains.

That said, benchmarks are not the whole picture. **DeepSeek's own admission** places V4-Pro at **three to six months behind Gemini 3.1-Pro and GPT-5.4** on the most demanding frontier tasks. The **1 million token context window** is competitive but not unique. **V4-Flash**, trained entirely on **Huawei Ascend chips**, is particularly significant because it demonstrates that frontier-grade model training is now possible without access to **Nvidia**'s restricted export hardware. The benchmark story is impressive. The hardware independence story may be more important.

### What hardware independence actually means

*The Nvidia dependency has been one of the core assumptions of Western AI policy. Export controls on advanced chips were supposed to slow China's AI development. DeepSeek V4-Flash, trained on Huawei Ascend, is a direct empirical challenge to that assumption. It does not mean **export controls** are useless. It means they are not sufficient. That is a significant strategic finding.*

## Huawei infrastructure: technological independence on the march

### What the Ascend ecosystem represents

**Huawei's Ascend** chip ecosystem is the most advanced Chinese alternative to **Nvidia's H100 and A100** series, which have been restricted for export to **China** since 2022. The fact that **DeepSeek V4-Flash** was trained on **Ascend chips** is the most concrete evidence to date that Chinese AI labs have successfully adapted their training pipelines to non-Nvidia hardware at a level that produces competitive frontier models.

The strategic implication is significant. Western export controls on advanced semiconductors were designed on the premise that without access to the most capable chips, Chinese AI would fall progressively further behind American frontier systems. The **DeepSeek V4** release suggests that this lag is narrower than export control architects hoped — and that the Chinese AI ecosystem is developing the engineering capacity to extract frontier performance from domestically available hardware. Whether this trend continues, and how quickly, is one of the most important questions in the geopolitics of technology over the next decade.

### The export control assumption challenged

*I want to be careful not to overstate this. Huawei Ascend chips are not equivalent to Nvidia H100s at the highest levels of performance. Training the most capable models at the absolute frontier still has an Nvidia advantage. But the gap is closing — and V4-Flash proves that you can build a model that beats Claude on SWE-bench without touching a single Nvidia chip. That matters.*

## State funding and the "startup" fiction

### What "private company" means in the Chinese AI ecosystem

**DeepSeek** presents itself as a private AI research company, founded in 2023, based in Hangzhou. This framing is accurate in the narrow corporate sense. But the May 2026 reporting by **Tech Times** on a **$4 billion funding round backed by China's state AI fund** adds essential context. In **China**'s technology ecosystem, the line between private company and state instrument is systematically blurry in ways that have no Western equivalent.

State-backed capital means state-aligned priorities. It does not mean that **DeepSeek** researchers are writing code in response to daily instructions from the **Zhongnanhai**. It means that the company's strategic direction — what models to build, at what price to release them, under what license — is shaped by a capital structure that has national strategic interests built into it. The **MIT license release of V4** is consistent with a strategy of market capture through free distribution. That strategy serves Chinese national interests. Whether the company executives consciously frame it that way is secondary to the structural reality.

### The structural relationship between capital and strategic direction

*There is a conversation worth having in the West about whether our own AI companies are as "private" as we claim — when they receive Pentagon contracts, when they shape national AI strategies, when their infrastructure is treated as critical national infrastructure. The Chinese case is more explicit. Ours is more euphemistic. Neither is entirely the independent private enterprise we sometimes pretend.*

## The impact on the global AI ecosystem

### What frontier-tier open-source models do to the market

The release of **V4-Pro** under an **MIT license** at **$3.48 per million tokens** has immediate competitive effects on the global AI market. Companies that were paying **$15 to $25 per million tokens** for equivalent American models face a direct cost competition they cannot ignore. Startups, government agencies, and universities in the **Global South** that have been priced out of frontier AI now have access to a competitive free alternative.

The longer-term effect is more structural. If **DeepSeek V4** and its successors become the default AI infrastructure for governments and companies that cannot afford American models, then the **data flows**, the **fine-tuning practices**, and the governance assumptions built into **Chinese AI** become embedded in global technology infrastructure. This is the **network effect** of open-source dominance: once you are the default tool, you shape how the tool is used, extended, and understood. **Google** understood this with Android. **DeepSeek** may be understanding it with AI.

### The network effect logic of open-source dominance

*I am genuinely uncertain about the long-term competitive dynamics here. Open source has historically tended to raise the floor of quality while compressing margins — benefiting users, challenging incumbents. Whether that dynamic works the same way when the open-source actor is state-backed and the incumbents are private companies competing for shareholder returns is not obvious. The rules of this game are still being written.*

## The security risk question

### What the MIT license permits that most people have not considered

The **MIT license** on **DeepSeek V4** permits modification without restriction. This means that any actor — including state intelligence agencies, including non-state armed groups, including cybercriminal organizations — can take the model, remove its **safety training**, and deploy a version optimized for purposes that the original model's safety constraints were designed to prevent. This is not a theoretical risk. Multiple AI safety researchers have documented that **safety training can be stripped from open-source models** with modest computational resources and technical expertise.

The European regulatory response has been uneven. **Italy** blocked the **DeepSeek API** but not self-hosted deployments — a distinction that illustrates the fundamental challenge of regulating open-source AI: once the weights are public, distribution is essentially impossible to control. Any server anywhere in the world can host and serve a modified version. The European response is catching up to a reality that the Chinese release has already created. The policy conversation about what to do with state-backed, MIT-licensed frontier AI is just beginning.

### The regulatory gap in European AI governance

*I will say clearly what the security analysis implies: an MIT-licensed frontier-tier AI model trained with Chinese state backing, with safety training that can be removed by anyone with a few hundred dollars of compute, is not a neutral technical contribution to open-source culture. It is a **dual-use tool** at a scale that the open-source community has not previously had to reckon with. The conversation about this needs to happen faster.*

## Sino-American rivalry from below

### How AI is becoming the new terrain of great-power competition

The **Sino-American technological rivalry** has been analyzed primarily through the lens of hardware — chip export controls, **semiconductor supply chains**, advanced manufacturing capacity. The **DeepSeek V4** release suggests that the decisive contest is shifting to software — specifically to models, weights, and the data and training practices that produce them. **China** cannot easily get the most advanced chips. But it can build models that are competitive with those trained on those chips, and it can release them free to the world.

This is competition "from below" in the sense that it bypasses the terrain where Western advantages are strongest — hardware manufacturing, semiconductor design, Nvidia's architecture — and instead contests the terrain where barriers to entry are lower: **training efficiency**, architectural innovation, and distribution strategy. The **CFR analysis of April 29, 2026** described this as "a new phase in the U.S.-China AI rivalry" — one defined less by absolute capability gaps and more by strategic ecosystem building. **DeepSeek** is building an ecosystem. And it is doing so by giving it away.

### Why distribution strategy matters as much as capability

*The American AI industry built its competitive moat around capability and compute. DeepSeek is building its competitive moat around ubiquity and cost. These are different moats, and they interact in unexpected ways. The company that wins may not be the one with the best model — it may be the one whose model is running on the most infrastructure by 2030.*

## The Mixture-of-Experts model: efficiency as strategy

### Why activating 49 billion out of 1.6 trillion parameters is significant

The **Mixture-of-Experts** (**MoE**) architecture that underlies **DeepSeek V4-Pro** is one of the most important technical decisions in the model's design. In an **MoE** system, not all parameters are active for every inference — only a subset of "expert" modules activates depending on the input. With **1.6 trillion total parameters** but only **49 billion active** per inference, **V4-Pro** achieves performance comparable to much more computationally expensive dense models at a fraction of the **inference cost**.

This architectural efficiency is not just a technical achievement. It is a strategic one. It means that **DeepSeek V4-Pro** can be served at **$3.48 per million tokens** while maintaining frontier-tier performance — because the per-inference compute cost is dramatically lower than a comparably performing dense model. The **MoE** architecture is also why training on **Huawei Ascend chips** — which are less capable than **Nvidia H100s** at pure dense computation — was sufficient to produce a competitive model. Efficiency as the path around hardware disadvantage. This is elegant engineering in service of a strategic objective.

### The strategic implications of MoE efficiency

*The MoE architecture is not DeepSeek's invention — it has been used by Google, Meta, and others. What DeepSeek has done is combine MoE with Huawei hardware and MIT licensing in a single release that produces strategic effects none of those components would generate alone. It is not the technology that is the weapon. It is the combination and the distribution.*

## The DeepSeek ecosystem: far beyond V4

### What exists beyond the flagship release

**DeepSeek**'s release history is broader than the V4 moment. Prior to **V4-Pro** and **V4-Flash**, the company had released **DeepSeek-R1** in January 2026, which produced the famous **Nasdaq shock** when markets realized that a Chinese open-source model was competitive with **GPT-4o** on reasoning tasks. The **V4** release builds on a trajectory of accelerating capability that has surprised Western analysts repeatedly.

The company had already announced by the time of the **April 24 release** that **V4.1** was expected in **June 2026**. The broader ecosystem includes coding models, multimodal variants, and domain-specific fine-tunes being built by third-party developers on the **open-weight releases**. Once an open-source model achieves critical mass — as **Meta's Llama** demonstrated in the West — the ecosystem of derivative models and applications grows faster than the original developer's roadmap. **DeepSeek** is seeding that ecosystem globally, free of charge.

### The ecosystem momentum beyond the flagship

*I have been following AI releases for several years. The cadence of DeepSeek's releases — R1 in January, V4 in April, V4.1 expected in June — suggests an organization that is not racing to a single benchmark victory. It is building an ecosystem at a pace that implies sustained institutional commitment and resource availability. That is the signature of state-backed ambition, not a startup sprint.*

## The ethical dimension: Chinese AI, its values and its biases

### What "aligned" means when the aligner is the Chinese state

All large language models encode the values, biases, and political assumptions of their training data and human feedback processes. **American** models encode American assumptions about speech, democracy, sensitive topics, and institutional legitimacy. **Chinese** models encode Chinese state assumptions about the same topics. **DeepSeek V4**, trained with data and RLHF processes shaped by a company operating under **Chinese law**, reflects those assumptions.

This is not a claim that **DeepSeek V4** is a propaganda machine producing false information. It is an observation that the model's behavior on politically sensitive topics — **Taiwan**, **Xinjiang**, **Tiananmen**, **Hong Kong** — reflects the constraints of its training in ways that any informed user should understand. The MIT license allows removing safety training, including safety training that encodes Chinese political constraints. But it also allows deploying the model in its unmodified state, constraints and all, to populations who may not be aware of them. The ethical complexity here is genuine and has not been adequately addressed in the public conversation about the release.

### The values embedded in training data

*I apply the same standard to American models as to Chinese ones: they all embed assumptions. GPT and Claude embed American liberal assumptions about speech and politics that are themselves not universal. The difference is one of transparency and power: American companies are subject to democratic oversight mechanisms, however imperfect. DeepSeek is not. That asymmetry matters for global governance of AI.*

## DeepSeek's impact on financial markets and big tech strategies

### What happened to Nasdaq in January 2026 and why it matters

The **January 2026 release of DeepSeek-R1** produced a significant market reaction. **Nvidia** shares fell sharply in a single session — the market's interpretation being that if **Chinese AI** can produce competitive models on cheaper hardware, the demand trajectory for the most advanced **Nvidia GPUs** is less certain than previously assumed. The **April 2026 V4 release** was absorbed with less volatility, in part because markets had already partially repriced the competitive threat after the R1 shock.

The strategic response from **American big tech** has been multi-directional. **Google** has accelerated its own **Gemini** open-weight releases. **Meta** has continued its **Llama** series. **Mistral** in Europe has maintained a competitive open-source presence. The effect is a global acceleration of open-source frontier AI capability — an arms race of open-weight models in which **DeepSeek** is the catalyst but by no means the only actor. The competitive pressure from **DeepSeek** is forcing American AI companies to be more open than their historical business model preference would suggest.

### The competitive dynamic forcing open-source acceleration

*There is a genuine irony here: DeepSeek's open-source releases are pushing American AI companies to release more open-source models than they would have chosen to otherwise. The result is a global enrichment of the open-source AI commons. The mechanism is geopolitical competition. The output is, in some respects, beneficial for the global AI ecosystem. History rarely moves in straight lines.*

## AI in public institutions: when governments adopt DeepSeek

### The adoption pattern in the Global South and beyond

The price differential between **DeepSeek V4-Pro** at **$3.48 per million tokens** and American equivalents at **$15 to $25** creates a powerful incentive for governments and public institutions with limited AI budgets to adopt the Chinese model. This adoption pattern is already visible in parts of **Southeast Asia**, **Africa**, and **South America** — regions where the AI budget constraints are real and where the preference for cost-effective solutions overrides geopolitical alignment concerns.

When a government's digital infrastructure runs on **DeepSeek**, that government's data flows through — or is processed by — a model whose training, alignment, and ongoing development are shaped by a Chinese state-backed company. Even without any active data collection (and the MIT license for self-hosted deployments reduces some of this risk), the embedding of Chinese AI assumptions into government workflows has long-term implications for how public institutions understand and respond to politically sensitive topics. This is soft power in the most literal sense: the values built into the tools shaping the conclusions that institutions reach.

### The soft power of default infrastructure

*I am not saying that every government that deploys DeepSeek becomes an agent of Chinese interests. I am saying that the tools we use to process information shape the information we produce. This is true of American AI too. The difference is that when you deploy American AI, you are generally aware that you are deploying American AI. When DeepSeek becomes ubiquitous through free open-source distribution, the awareness of its Chinese origin may fade from practical consideration. That fading matters.*

## The challenge of international AI regulation: a framework still to be built

### Why existing frameworks fail to address state-backed open-source AI

The **European Union AI Act**, which came into force in 2024, establishes risk categories and compliance requirements for AI systems deployed in Europe. But it was designed in a world where the primary regulatory targets were commercial AI systems operated by identifiable corporate entities. A **MIT-licensed** model distributed as open weights, trained by a Chinese company, deployable by anyone on any server — this does not map cleanly onto the AI Act's enforcement architecture.

The international regulatory vacuum is more severe. There is no binding international framework for AI development, deployment, or export. The **Bletchley Park Declaration** of 2023 and subsequent AI safety summits produced non-binding commitments that reflect the priorities of the governments that organized them — primarily Western governments. **China** participated selectively. **DeepSeek V4**'s release is not a violation of any existing international norm. It is a demonstration that the existing norm architecture was not designed for this category of competitive action. Building the framework is the work of this decade.

### Why the regulatory architecture needs to be rebuilt

*I do not believe that regulation alone can solve the governance challenge posed by state-backed open-source frontier AI. Regulation that is too restrictive will simply push development to jurisdictions without it. Regulation that is too permissive will fail to address the genuine security and values risks. The answer probably involves a combination of technical standards, transparency requirements, and international coordination that does not yet exist. I do not know what it looks like. I know it is urgent.*

## Conclusion: open source as the new terrain of power

### What April 24, 2026 established

**April 24, 2026** established that **China** can produce frontier-tier AI models, train them on domestically available hardware, release them under the most permissive license in software, and price them at a fraction of American equivalents — all within a single product release. This is not a future capability. It is a demonstrated present one. The question for Western governments, AI companies, and policymakers is not whether this happened. It is what it means, and what to do about it.

### What the essay cannot predict but must name

What this essay cannot predict is whether **DeepSeek** will continue at this pace, whether American frontier labs will maintain their capability lead, whether export controls will be tightened or circumvented, or whether international AI governance will develop fast enough to be relevant. What it can name, with confidence grounded in verified facts, is this: the terrain of great-power competition has expanded to include the weights of AI models. The weapon is open source. The distribution is free. And the strategic logic — use capability to build ecosystem, use ecosystem to embed influence — is as old as great-power competition itself, applied to a technology that did not exist a decade ago. **April 24, 2026 is the day that became undeniable.**

**By Maxime Marquette, columnist**

## Columnist's transparency note

### Editorial stance

I am not a journalist but a **columnist and analyst**. My expertise lies in observing and analysing the geopolitical, economic and strategic dynamics that shape our world. My work is to take political strategies apart, to understand global economic movements, to put the decisions of international actors in context and to offer analytical perspectives on the changes that are redefining our societies.

I do not claim the cold objectivity of traditional journalism, which stops at reporting facts. **I claim analytical clarity**, rigorous interpretation and a deep understanding of the complex issues that concern us all. My role is to give meaning to the facts, to place them in their historical and strategic context, and to offer a critical reading of events.

### Method and sources

This text keeps the basic distinction between **verified facts and interpretive analysis**. The rule never changes: a factual claim is published only when a verifiable source supports it, and the sources this article actually used are listed under **Sources**, never here.

**Primary source families** used by the publication, when the story has them: official statements from governments and international institutions, public statements by political leaders, reports from intergovernmental organisations, dispatches from recognised international news agencies.

**Secondary source families**: specialised publications, internationally recognised news media, analyses from established research institutions, reports from industry organisations.

When an article cites statistical, economic or geopolitical data, it comes from the institutions that produce it (intergovernmental organisations, central banks, national statistics offices), and the exact institution is named under **Sources**.

### Nature of the analysis

The analyses, interpretations and perspectives in the analytical sections of this article are a **critical, contextual synthesis** based on the information available, the trends observed and the expert comments cited in the sources consulted.

My role is to interpret these facts, to set them in the context of today's geopolitical and economic dynamics, and to give them a coherent meaning within the larger story of the changes shaping our time. These analyses reflect expertise built through continuous observation of international affairs and an understanding of the strategic mechanisms that drive global actors.

This article describes a documented state of affairs on its publication date, not a prediction: later developments may shift its outlook. No update is promised in advance; when an article is corrected or completed, the change is dated in the text.

*I finish this essay with a conviction I hold strongly enough to state plainly: open-source AI as a geopolitical weapon is not a hypothetical future scenario. It is the present. DeepSeek V4 is not a warning sign — it is the event itself. Western governments that are still treating this as a future policy question are already behind. The framework must be built now, imperfect and incomplete, because waiting for a perfect framework means conceding the terrain entirely.*

## Sources

### Primary sources

- [Council on Foreign Relations — DeepSeek V4 Signals a New Phase in the U.S.-China AI Rivalry — April 29, 2026](https://www.cfr.org/articles/deepseek-v4-signals-a-new-phase-in-the-u-s-china-ai-rivalry)

- [Al Jazeera — China's DeepSeek unveils latest model a year after upending global tech — April 24, 2026](https://www.aljazeera.com/economy/2026/4/24/chinas-deepseek-unveils-latest-model-a-year-after-upending-global-tech)

- [Reuters — DeepSeek V4: Chinese AI model adapted for Huawei chips — April 24, 2026](https://www.reuters.com/world/china/deepseek-v4-chinese-ai-model-adapted-huawei-chips-2026-04-24/)

- [Tech Times — China's State AI Fund Backs DeepSeek $4 Billion Round — May 16, 2026](https://www.techtimes.com/articles/316717/20260516/chinas-state-ai-fund-backs-deepseek-4-billion-round-efficiency-challenge-nvidia-dependent.htm)

### Secondary sources

- [AI Blogs — DeepSeek V4-Pro and Flash: MIT License, Frontier-Tier Open — May 26, 2026](https://ai-blogs.org/news/2026-05-26-deepseek-v4-pro-flash-mit-license-frontier-tier-open-am.html)

- [Forbes — China's DeepSeek V4 and Qwen Reshape the Open-Source AI Race — April 28, 2026](https://www.forbes.com/sites/jonmarkman/2026/04/28/chinas-deepseek-v4-and-qwen-reshape-the-open-source-ai-race/)

- [Hokai — DeepSeek V4-Pro model card — April 24, 2026](https://hokai.io/hub/models/deepseek-v4-pro)

- [Pittsburgh Post-Gazette — China's DeepSeek unveils new AI model — April 24, 2026](https://www.post-gazette.com/news/world/2026/04/24/china-deepseek-ai-artificial-intelligence/stories/202604240064)

---

Source: Maxime Marquette, MadMax. When citing, name the author and link https://mad-max.co/en/article/essai-deepseek-v4-l-open-source-comme-arme-geopolitique