Attack on the Nvidia monopoly: Why the AI ​​prodigy DeepSeek is now building its own chips

US sanctions backfire: How DeepSeek is turning the global tech order upside down

Cheaper, smarter, independent? This is what's behind DeepSeek's radical chip plan

The Chinese AI startup DeepSeek has already shaken up the global tech world with its extremely efficient and unprecedentedly affordable software models. Now comes the next logical and explosive step: According to insider reports, the company is secretly working on its own AI chip. What initially sounds like a mere technical detail for hardware nerds is, in reality, a geopolitical and economic earthquake. Driven by US export controls and the pursuit of ultimate cost control in the mass market of AI inference, DeepSeek is increasingly freeing itself from dependence on giants like Nvidia. Equipped with record funding in the billions and government backing, China's flagship lab is preparing for a paradigm shift. This move could not only threaten Nvidia's dominance but also fundamentally alter the entire global semiconductor industry and the balance of power in the race for artificial intelligence. An analysis of a strategic masterpiece.

DeepSeek develops its own AI chip: When software is no longer enough: China's flagship AI lab reaches for hardware sovereignty

From model to machine: What Reuters revealed

On July 7, 2026, Reuters, citing three sources familiar with the matter, reported that the Chinese AI startup DeepSeek was working on its own AI chip. This news, initially appearing to be a footnote in the global technology discourse, reveals itself upon closer examination as a strategic move with far-reaching economic, geopolitical, and industrial consequences. The chip is intended primarily for inference tasks—that is, for calculations where a pre-trained model generates answers to user queries—and not for training new models. This sounds like a technical specification, but is in reality a precise economic decision: Inference is the mass market of the AI ​​industry, the phase in which scaling translates into tangible costs.

According to multiple reports, development efforts are still in their early stages. DeepSeek has contacted external partners and held discussions with chip design companies, semiconductor fabricators, and memory manufacturers. Particularly revealing is the fact that the company has been selectively hiring chip design engineers in recent months – without public job postings on common platforms, but exclusively through discreet channels. This operational secrecy suggests a strategy that prioritizes strategic surprise over transparency and aims to give competitors no lead time to countermeasures.

According to one source, the chip project began about a year ago. This coincides precisely with the period in which DeepSeek gained international attention with its V3 model, while simultaneously its increasing reliance on Nvidia chips became a political and operational risk. Although DeepSeek increasingly emphasizes Huawei hardware in its public communications, solid evidence has emerged that the company has also used Nvidia's Blackwell chips for its latest models – chips whose export to China is officially prohibited.

The anatomy of an AI start-up: Who is behind DeepSeek?

To properly assess the significance of this chip project, one must understand DeepSeek's origins. The company is no ordinary startup that emerged from a garage. It is the ambitious side project of a quantitative hedge fund. Founder Liang Wenfeng, born in the 1980s in the southern Chinese province of Guangdong and a graduate of Zhejiang University, co-founded the quantitative fund High-Flyer in 2015. High-Flyer used mathematics and artificial intelligence for algorithmic stock trading and at one point grew to $14 billion in assets under management.

In 2021—even before the tightened US export restrictions—Liang began systematically buying Nvidia GPUs. A business partner described him at the time as a tech geek who talked about a 10,000-chip cluster for model development and was taken seriously by no one. In fact, by 2022, High-Flyer had amassed around 10,000 A100 chips—a resource that, in retrospect, seems like a strategic stroke of genius. In May 2023, Liang then founded DeepSeek as a spin-off from High-Flyer, with the stated goal not of maximizing profit, but of being at the forefront of global AI progress. In a widely quoted interview, Liang articulated his credo: neither to incur losses nor to generate excessive profits, but to advance the entire ecosystem.

In February 2025, Liang Xi Jinping met with DeepSeek personally at a meeting with technology entrepreneurs in Beijing. DeepSeek was thus no longer a private research project – it had become a national symbol of technological self-assertion. This symbolic status has practical consequences: access to state resources, protection from regulatory obstacles, and implicit support in procuring scarce hardware are privileges granted to only a few Chinese technology companies.

The outsider's business model: Efficiency as system critique

Before DeepSeek's chip ambitions can be assessed economically, it's essential to understand the underlying business model. DeepSeek has methodically challenged the rules of the AI ​​industry by demonstrating that peak performance doesn't require astronomical training costs. When the company revealed in December 2024 that training its R1 model had cost only around $5.6 million—compared to hundreds of millions for OpenAI's GPT-4—it sent shockwaves through global stock markets. Nvidia's stock lost nearly 17 percent of its value in a single trading day, wiping out $589 billion in market capitalization—the largest single-day drop in stock market history.

The technological basis for this efficiency lies in the architecture of the DeepSeek models: They use a Mixture of Experts (MoE) structure, in which not all parameters of a model are activated for each query, but only a relevant subset. This drastically reduces the computational effort per inference operation. In addition, there are further algorithmic innovations such as Multi-Head Latent Attention (MLA), which significantly reduces the memory requirements when processing long contexts. DeepSeek has thus demonstrated that algorithmic creativity can compensate for some of the hardware deficit – a finding that calls into question the effectiveness of the entire Western chip export strategy.

The consequences for corporate economics are remarkable: DeepSeek offers its services at prices that undercut Western competitors by up to 90 percent. Although the model is available as open source, this pricing structure enables aggressive market penetration based not on the classic venture capital model of "growth before profitability," but on structurally lower operating costs. This is precisely the key to understanding the chip project: Whoever controls their own hardware controls the longest cost lever in the AI ​​value chain.

The shadow of Nvidia and Huawei: Why DeepSeek wants to break the dependency

DeepSeek's current chip situation is the result of an extraordinary mix of geopolitical pressure, technological compromises, and strategic self-reliance. The company has long relied on Nvidia's hardware, whose CUDA software ecosystem is still considered the most powerful and developer-friendly in the world. Chinese authorities and a US government official have confirmed that DeepSeek's V4 model was trained on Nvidia's Blackwell chips—currently the company's most powerful chip—despite their export to China being officially prohibited. The infrastructure in question is reportedly located in a data center in Inner Mongolia.

This reliance on prohibited or at least legally questionable hardware is not a sustainable foundation for a company that aspires to define China's national AI infrastructure. Huawei offers an alternative with its Ascend chip family, but the performance gap is significant: DeepSeek's own tests show that the Ascend 910C achieves only 60 percent of the inference performance of Nvidia's H100. For training tasks, the gap is even wider. Huawei manufactures its chips using SMIC's 7-nanometer process – a technology that corresponds to TSMC's state of the art from 2019/2020, not the current state. The reason for this is structural: To date, China has not received a single EUV lithography machine from ASML, the Dutch monopolist for the production of the highest-resolution semiconductor layers.

A revealing turning point occurred in February 2026: Reuters reported that DeepSeek had not granted US chip manufacturers—including Nvidia—early access to its new flagship model, the V4, despite this being industry standard practice. Instead, Huawei received exclusive early access to optimize its software for running the model. In April 2026, DeepSeek then released the V4 model, which for the first time incorporated both Nvidia's GPUs and Huawei's Ascend NPUs within a shared hardware validation framework. Huawei confirmed that its Ascend 950 chips had contributed to the V4's development.

An analysis by the Wall Street research firm SemiAnalysis revealed an even more fundamental connection: DeepSeek V4 and Huawei's Ascend 950DT were co-designed – meaning they were jointly developed from the outset, rather than the model being adapted for Huawei hardware later. The 950DT architecture, with its HiZQ 2.0 memory (144 GB capacity, 4 TB/s bandwidth) and specialized execution units, was designed from the beginning to target DeepSeek's inference patterns. The market reaction to the V4 announcement was clear: SMIC's stock rose by 10 percent on the day of the announcement, while shares of other Chinese contract manufacturers in Hong Kong climbed between 9 and 15 percent.

The economics of your own chip: Between strategic rationality and technological risk

Why is DeepSeek developing its own chip now, when the co-development with Huawei has already progressed so far? The answer lies at the intersection of corporate economics, strategic autonomy, and a sober risk analysis.

First: cost structure and margin optimization. In the AI ​​industry, inference isn't the glamorous part, but it is the profit-driven business. Every user query to a DeepSeek model generates computational costs that depend on the hardware used. Those who rely on purchased chips—be it Nvidia or Huawei—are also paying the hardware supplier's margin. A proprietary inference chip, optimized for the specific characteristics of DeepSeek models (MoE architecture, MLA mechanism, long context windows of up to one million tokens), could significantly reduce inference costs per token and thus sustainably defend the structural cost advantage that underpins DeepSeek's market position.

Secondly: supply chain security and export control risk. Dependence on Nvidia hardware is existentially risky given escalating US export restrictions. While the Trump administration temporarily authorized the export of Nvidia's H200 chips to China, not a single H200 device reached a Chinese buyer as of July 2026 – blocked by ongoing diplomatic disputes over trade terms. Goldman Sachs analysts expect the shift of Chinese companies to domestic chips to accelerate significantly between 2026 and 2028. Those who achieve independence early on safeguard their operational capability against political uncertainties.

Third: Market positioning and ecosystem control. A proprietary chip creates the opportunity to establish a proprietary software ecosystem that binds other developers to the DeepSeek platform. According to the unanimous assessment of the Chinese semiconductor industry, Nvidia's CUDA ecosystem is the decisive competitive barrier for domestic alternatives: Moore Threads described Nvidia's ecosystem in its December 2025 IPO prospectus as "not easily overcome." Another strategy would be to integrate the software stack directly into the model ecosystem – precisely what DeepSeek is attempting through its co-development with Huawei and now with its own chip project.

Fourth: Political context and state support. China's 15th Five-Year Plan (2026–2030) mentions artificial intelligence 52 times, compared to 11 mentions in the previous plan. The plan aims for a 90 percent AI adoption rate in Chinese industry by 2030, relying exclusively on domestic providers. The National AI Investment Fund has invested directly in DeepSeek—as the sole investor with voting rights and without a lock-up period. DeepSeek's chip project is thus implicitly backed by the state and part of a national strategy of technological self-assertion.

The financing framework: $7.4 billion for the next step

The economic framework for DeepSeek's chip ambitions has been defined by its latest funding round. In June 2026, the company closed its first external funding round, raising more than 50 billion yuan – approximately US$7.4 billion – at a valuation of US$50 to US$59 billion. It is the largest AI investment in China to date.

The structure of the funding round is both unusual and revealing. Liang Wenfeng himself contributed 20 billion yuan – roughly 40 percent of the total – thereby securing control of the company. Investors must deposit their capital into a limited partnership managed by Liang, not directly into DeepSeek. They are subject to a five-year lock-up period and have no voting rights. Tencent is expected to invest around 10 billion yuan, and CATL – the world's largest battery manufacturer – around 5 billion yuan. Other investors include NetEase, JD.com, IDG Capital, and Monolith Management, with the total number of investors expected to remain below ten.

This financing structure sends a clear signal. A founder who personally finances 40 percent of a billion-dollar funding round isn't maximizing their own exit payout—they're securing the operational independence of a long-term project. CATL's involvement is particularly noteworthy: A battery manufacturer investing in an AI company signals an expectation that AI infrastructure and energy systems will be inextricably linked in the future. China's approach of envisioning artificial intelligence as national infrastructure—not as a consumer product—is evident here in the capital structure.

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

Geopolitical tectonics: Export controls as an innovation accelerator

It is one of the most remarkable ironies of recent technology policy: US export restrictions, designed to slow China's AI development, may have had precisely the opposite effect. This thesis deserves a nuanced economic analysis.

On the one hand, the restriction is real and measurable. China cannot import EUV lithography machines from ASML. According to ASML, it has not yet delivered a single EUV machine to China. Furthermore, the MATCH Act legislation currently being debated in the US Congress would further restrict the export of older DUV machines. SMIC, China's leading chip contract manufacturer, produces using a 7-nanometer process—but only through a complex multi-patterning process that increases production costs and reduces yield. China's semiconductor self-sufficiency reached approximately 28 percent in the fourth quarter of 2025—compared to 16 percent in 2024—driven by government subsidies equivalent to US$150 billion since 2020. By comparison, the US CHIPS Act amounts to only US$52 billion.

On the other hand, sanctions without full enforcement create substitution pressure, which fuels innovation. DeepSeek's R1 shock in early 2025 proved that Chinese algorithm engineers turned hardware scarcity into a virtue of efficiency. Because no H100 chips were available, architectures were developed that delivered more performance with less hardware. This forced efficiency innovation is now a global competitive advantage in the form of DeepSeek's MoE architecture.

Semiconductor analyst Kevin Xu of Interconnected Capital predicts that Chinese companies will still rely on Nvidia chips for another three to five years – but the direction is clear: Beijing has a systemic interest in ending this dependency as quickly as possible. Goldman Sachs confirms in a May 2026 analysis that DeepSeek V4 is compatible with eight different Chinese chip architectures, including products from Huawei, Hygon, and Alibaba's T-Head division. The Beijing Institute for Artificial Intelligence (BAAI) has already adapted DeepSeek V4 Flash for full inference operation on more than eight different AI chip architectures. This isn't dependency reduction – it's systematic platform independence as a corporate strategy.

Nvidia's position: Between market exclusion and strategic adjustment

For Nvidia, DeepSeek's chip project represents a further escalation of an already existential challenge. CEO Jensen Huang has described China's AI infrastructure market as a $50 billion market with 50 percent annual growth. KeyBanc analyst John Vinh estimates that under free trade conditions, Chinese companies would purchase around 1.5 million H200 chips this year—a potential revenue of approximately $30 billion. Actual shipments: zero.

The situation for Nvidia is more ambivalent than it initially appears. In the area of ​​model training, Nvidia's CUDA ecosystem still holds a dominant position that is unlikely to be challenged in the short to medium term. Chinese companies themselves acknowledge this internally: Shanxi Securities, in a stock analysis, described Nvidia's CUDA ecosystem as "the main obstacle" to the widespread adoption of domestic AI chips. The real shift is occurring in the inference domain, where switching costs are lower because software adjustments—not entirely new developments—are sufficient.

Nvidia has already reacted. The company is trying to maintain its market position through China's "Physical AI" sector, for example, through a collaboration with the humanoid robotics startup Unitree. But this is a niche pivot, not a strategic response to the structural decline of the AI ​​infrastructure market. The historical analogy being discussed in the industry is revealing: At the height of the dot-com era, Cisco represented four percent of the S&P 500—the market was right that the internet would change the world, but it was wrong that Cisco would own that change. The question of whether Nvidia could experience a similar misjudgment is no longer merely academic.

China's Semiconductor Strategy is undergoing a Paradigm Shift

Beyond the immediate corporate level, DeepSeek's chip project is part of a broader strategic realignment documented in China's 15th Five-Year Plan. The term "lithography machine" does not appear once in the 141-page planning document. This is not an oversight—it is a strategic signal. China no longer measures its success by how many chips it produces domestically, but by how deeply computing power is embedded in its economy. The goal is digital value creation at 12.5 percent of GDP by 2030.

The new strategic concept – in Chinese “模芯云用” (Model Chip Cloud Application) – defines the chip as one of four equally important layers in an integrated system. This conceptual shift has practical consequences: Instead of pursuing a hopeless catch-up in EUV manufacturing, Beijing is focusing its resources on chiplet design and advanced packaging – techniques that allow multiple legacy chips to be integrated into a more powerful system. Suzhou and Wuxi are being developed into national packaging hubs, supported by the National Integrated Circuit Industry Investment Fund.

This strategy of "overtaking by changing lanes" has a historical parallel in the Chinese mobile communications market: When China made the technological leap from 3G to 4G, it was able to move directly into the latest generation without the burden of outdated infrastructure – and today, with Huawei, it dominates a significant portion of global 5G development. A similar leap in the semiconductor sector – from addressing the manufacturing gap to system optimization – could fundamentally shift the geopolitical landscape. The key indicator will be whether China's industry can replace the CUDA software stack, which Chinese chip manufacturers themselves describe as "not easily overcome.".

Market structural implications: Bifurcation as a new paradigm

The economic world order of the semiconductor industry is facing its most consequential crossroads since the emergence of Silicon Valley. On one side, there is a US-centric supply chain dominated by Nvidia's CUDA ecosystem, TSMC as a manufacturing monopolist, and a software stack that has evolved over decades. On the other side, there is a consolidating Chinese alternative: Huawei Ascend as a hardware platform, DeepSeek as a model layer, Alibaba Cloud, Tencent Cloud, and Baidu AI Cloud as distribution channels, and increasingly, proprietary chip designs that do not rely on CUDA.

This bifurcation of the global AI infrastructure is no longer a theoretical possibility – it is happening in real time. Goldman Sachs predicts a strong shift towards domestically produced chips in China between 2026 and 2028. China's AI chip market is expected to grow to over one trillion yuan (around 140 billion US dollars) by 2028 – representing approximately 30 percent of the global market. The Huawei Ascend 950DT is slated for cloud deployment in August 2026, thus establishing the domestic inference infrastructure for the next generation of models.

For international companies seeking to operate in both markets—from automakers to pharmaceutical companies using AI models for R&D—this bifurcation increasingly means unavoidable strategic decisions. Technology platforms built on CUDA are incompatible with Chinese hardware. Companies developing on Huawei or DeepSeek infrastructure cannot scale their applications to Western hardware without significant adaptation. This isn't a hypothetical future—it's the current reality for any developer trying to operate on both sides of the technological divide.

Technological limitations and remaining uncertainties

A serious analysis cannot ignore the limitations of what is known. According to all available reports, DeepSeek's chip project is still in its early stages. The gap between a chip design that is in discussions with manufacturing partners and a marketable product is enormous. The technological hurdles are substantial: High-performance AI chips require high-bandwidth memory, advanced interconnect technologies, and a complete software stack. Manufacturing capacity in China—limited by the ASML embargo—imposes structural performance constraints.

It's significant that the chip is primarily designed for inference, not training. This reflects a realistic assessment of its own capabilities: Inference chips don't need to compete with Nvidia's H100 or Blackwell – they need to be efficient enough to reduce the operating costs of mass-producing model requests. This is an achievable goal, even with SMIC manufacturing technology.

Another uncertainty lies in assessing the scalability of the co-design model – the close integration of model architecture and hardware design. DeepSeek and Huawei demonstrated the viability of this strategy with the V4/Ascend-950DT project. Whether a completely in-house DeepSeek chip design can replicate or surpass these synergies, or whether co-development with an established chip designer like Huawei will remain more efficient in the long run, remains to be seen.

What this move means

DeepSeek's chip project is more than a technological investment – ​​it's a hypothesis about the future of the AI ​​industry. This hypothesis states that the crucial value creation is shifting from model development to hardware-software integration. Whoever controls both controls the costs, the performance, and ultimately, the market power.

It is no coincidence that other tech giants worldwide are pursuing the same hypothesis at the same time: Tesla has developed the AI5 chip for edge inference, Google is splitting its TPU line, Meta is committed to four generations of its own silicon development, Amazon operates Trainium, and Microsoft is developing Maia. DeepSeek is following a global trend that has gained particular urgency due to structural cost pressures and the strategic restrictions of US export policy for Chinese companies.

The economic irony remains: Had US export restrictions fully achieved their intended effect, there would be no DeepSeek as a global competitor, no independent Chinese AI chip ecosystem, and no strategic demand for a DeepSeek-proprietary inference chip. Instead, external pressure has triggered a surge of innovation that—if technologically successful—could permanently shift the initial asymmetry between US and Chinese AI infrastructure.

According to its 15th Five-Year Plan, China is pursuing national R&D spending at annual growth rates exceeding seven percent and has set a science and technology budget of 426.4 billion yuan (approximately US$59 billion) for 2026 – a ten percent increase over the previous year. These funds are channeled into a system in which DeepSeek, as a flagship company, is both the target and the catalyst of state technology policy. Within this framework, its own chip project is not the ambition of an individual company – it is the most capitalized form of state technology strategy, disguised as a startup.

The next twelve to eighteen months will show whether DeepSeek can cross the line from aspiring chip designer to fully functional semiconductor manufacturer. Its competitors—primarily Nvidia, but also Huawei—have a decisive lead in technology, ecosystem, and production infrastructure. However, DeepSeek has already proven its ability to translate resource scarcity into algorithmic ingenuity. The next proof will be more challenging—but the attempt has begun.

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs

From visibility to trust: Your scalable path with Xpert.Digital - Image: Xpert.Digital

In industrial B2B, sustainable business relationships rarely emerge overnight. They develop step by step – through visibility, professional relevance, recurring touchpoints, and growing trust. Xpert.Digital's 4-stage model addresses precisely this: It offers a structured path that begins with a manageable entry point and can evolve into deeper collaboration in business development if needed.

Instead of relying on loud marketing promises, this model puts the relationship at the forefront. Companies start with clearly defined, easily calculable measures and then decide, based on their own experience, how far they want to expand the collaboration. A key factor for this undisturbed trust-building process: The platform completely avoids annoying advertising ads, so the editorial focus remains solely on the companies' expertise.

More information here:

• From visibility to trust: Your scalable path with Xpert.Digital