Brandsit

Tag: DeepSeek

  • DeepSeek and Chinese AI – Why is the State Department warning allies?

    DeepSeek and Chinese AI – Why is the State Department warning allies?

    US diplomacy is entering a new phase of offensive against Chinese artificial intelligence leaders. The State Department has issued global guidelines to its outposts, ordering them to warn foreign governments about the practices of companies such as DeepSeek, Moonshot AI and MiniMax. The crux of the dispute is no longer just access to processors, but the process of so-called distillation, which Washington explicitly calls the theft of American technological thought.

    From a business perspective, distillation is a tempting shortcut. It allows smaller, cheaper-to-operate models to be trained on the results generated by powerful systems such as those from OpenAI. For Chinese startups, it’s a way to erode the US advantage at a fraction of the research cost. However, according to the US administration, this process not only copies intellectual architecture, but is done without authorisation, hitting Silicon Valley’s commercial foundations.

    DeepSeek’s situation is key here. The startup, which recently electrified the market with its V3 model, has just unveiled the V4 version, optimised for Huawei hardware. This is a clear signal of building an independent ecosystem that challenges the hegemony of Nvidia and Microsoft. While DeepSeek has consistently denied using synthetic data from OpenAI, US lawmakers have received reports suggesting the opposite: deliberately replicating the behaviour of models in order to clone them.

    Washington alerts that ‘distilled’ models often lack built-in fuses and controls, making them unpredictable for corporate use. At the same time, many Western institutions are already banning the use of DeepSeek tools, citing data privacy concerns.

    The timing of this escalation is no coincidence. The escalation in rhetoric comes just weeks before President Donald Trump’s planned visit to Beijing. The dispute over AI intellectual property becomes a bargaining chip in a broader technology war, which, after a brief period of relaxation, is again gaining momentum. The choice of AI model supplier is ceasing to be a purely technical decision and is becoming a statement in a growing geopolitical conflict.

  • DeepSeek V4: New AI model optimised for Huawei chips

    DeepSeek V4: New AI model optimised for Huawei chips

    DeepSeek, the Chinese startup that destabilised the AI market last year with its low-cost models, has just made a move of a strictly strategic nature. The release of a familiarisation version of the V4 model demonstrates that the Chinese AI ecosystem is preparing for a permanent disconnect from Western infrastructure.

    A key differentiator of V4 is its strict optimisation for the Huawei Ascend processor architecture. While the Hangzhou-based startup has historically based its success on Nvidia chips, the current turn to domestic solutions is a response to growing regulatory pressure from Washington. Huawei has confirmed that the entire Ascend ‘super node’ product line already supports the new DeepSeek architecture, suggesting deep integration at the hardware-software level to minimise performance losses from not having access to the latest H100 or Blackwell units.

    In terms of content, V4 Pro positions itself at the top of the world. According to the manufacturer, the model outperforms other open-source solutions in general knowledge tests, second only to the closed model Gemini-Pro-3.1 from Google. The strategy of providing a flash and preview version allows the company to collect real-time feedback data, which is essential for calibrating parameters prior to final deployment.

    The market reaction to the launch was immediate and painful for competitors. The stock market listing of rivals such as Zhipu AI and MiniMax saw significant declines, confirming DeepSeek’s dominant position in China’s open-source sector. At the same time, the company finds itself at the centre of a geopolitical cyclone. The White House openly accuses Beijing Labs of systemic intellectual property theft, and DeepSeek itself faces allegations of misuse of data from its OpenAI and Anthropic models.

    For investors, however, DeepSeek remains one of the most promising assets in Asia. The company, controlled by High-Flyer Capital Management, is aiming for a valuation in excess of $20 billion. Interest in taking a stake from giants such as Alibaba and Tencent suggests that Chinese Big Tech sees DeepSeek not just as a technology provider, but as the foundation of a national technology stack.

  • DeepSeek V4 relies on Huawei chips

    DeepSeek V4 relies on Huawei chips

    When DeepSeek released the V3 and R1 models late last year, financial markets reacted nervously. Investors began to ask uncomfortable questions about the point of Nvidia’s multi-billion-dollar infrastructure spend, when the Chinese lab had proven that high performance could be achieved at a fraction of the cost. The upcoming release of DeepSeek V4 is not just another technical update – it is a manifesto of Beijing’s technological independence.

    The signal sent by the industry is unmistakable. Instead of the AI industry’s traditional solicitation to optimise for US chips, DeepSeek has completely bypassed US manufacturers. Instead, the lab’s engineers have spent the last few months working hand-in-hand with Huawei Technologies and Cambricon Technologies. The result of this symbiosis has been to rewrite key sections of the model’s base code to squeeze the maximum capability out of the domestic silicon architecture.

    The scale of this is reflected in the order books of the Chinese giants. Alibaba, ByteDance and Tencent have decided to make bulk purchases of Huawei’s latest chips, amounting to hundreds of thousands of units. This is a strategic move to secure operational continuity in the face of geopolitical uncertainty and tightening sanctions. The fact that the V4 is being developed in three different variants optimised for Chinese processors suggests that the era of default priority for Santa Clara hardware in China is coming to an end.

    From a business perspective, DeepSeek V4 represents a stress test for the thesis of the need for the most expensive infrastructure to build powerful AI systems. If the next-generation model, trained on local hardware with theoretically lower performance than top-of-the-range H100 or Blackwell units, keeps pace with the US lead, the roadmap for AI development will be permanently altered.

  • Hunter Alpha: Is this a new model of artificial intelligence from DeepSeek?

    Hunter Alpha: Is this a new model of artificial intelligence from DeepSeek?

    A mysterious, powerful artificial intelligence model that recently debuted on developer platform OpenRouter has caused a stir in the technology industry. The system, labelled Hunter Alpha, appeared without any indication of its creator, which immediately sparked speculation that Chinese startup DeepSeek was quietly testing its next-generation technology ahead of an announced official launch.

    From a business and operational cost perspective, Hunter Alpha is a market anomaly. The system offers an impressive architecture of one trillion parameters and a contextual window of as many as one million tokens, allowing gigantic blocks of information to be analysed in a single interaction.

    In the case of market-leading frontier models, maintaining such performance involves huge expenditure on computing power. Meanwhile, the new system makes these capabilities available for free, clearly indicating the backing of a powerful entity. The chatbot itself maintains secrecy, admitting only that it is a Chinese system with a knowledge base ending in May 2025.

    The coincidence of this data with the anticipated April release of the DeepSeek V4 model is striking. AI engineers studying the new system point out that the model’s inference pattern and chain of thought bear a striking resemblance to DeepSeek‘s methodology, which is extremely difficult to mask or copy. The development community, however, remains balanced in its assessments.

    Independent researchers point to different architectural patterns and token processing behaviour, which may suggest that a very different player from the growing Asian AI market is behind the model. DeepSeek, which has a strong and rather unusual financial backing in the form of a quant hedge fund, has consistently declined to comment.

    Such anonymous launches are part of a proven market strategy. Publishing ‘stealth’ models allows companies to collect unbiased telemetry and query data from millions of users, which is key to calibrating a product before commercial launch. Recently, a similar manoeuvre was employed by Chinese company Zhipu AI prior to the release of the GLM-5.

    Regardless of whose logo will ultimately be on the Hunter Alpha model, the market has received it with great enthusiasm. The system processed more than 160 billion tokens in just a few days.

    The rapid adoption, especially by developers building autonomous AI agents, proves that free and advanced access to such a large contextual window addresses a real and unmet industry need. This is further confirmation that the Asian arms race in the generative AI segment is just now entering a new and extremely capital-intensive phase.

  • DeepSeek has broken the embargo: Chinese AI trained on Nvidia Blackwell

    DeepSeek has broken the embargo: Chinese AI trained on Nvidia Blackwell

    A senior Trump administration official confirmed on Monday that the latest model of artificial intelligence from Chinese startup DeepSeek had been trained using Nvidia Blackwell processors. The information caused a stir in Washington, as Blackwell is currently the world’s most advanced AI chip, the export of which to China is strictly prohibited by US law.

    According to US government findings, the computing infrastructure used by DeepSeek is located in a data centre in Inner Mongolia. Officials speculate that the startup will attempt to remove technical signatures from the upcoming model to hide the fact that US chips were being used. While the administration refuses to disclose the source of this information and the route the processors took to China, the incident highlights the leakiness of current technological barriers.

    The situation exacerbates an ongoing dispute in Washington over strategy towards China’s AI sector. On the one hand, AI advisor David Sacks and Nvidia CEO Jensen Huang argue that allowing the sale of older or intentionally weakened units (like the H200) discourages Chinese companies from building their own alternatives. On the other hand, ‘hawks’ in the administration warn that the commercial success of startups such as DeepSeek directly supports Chinese military capabilities.

    Headquartered in Hangzhou, DeepSeek has already gained notoriety for creating models that rival leading Silicon Valley products at significantly lower operating costs. However, current accusations suggest that this success may be based not only on engineering, but also on so-called model distillation – a technique that involves training its own systems on results generated by OpenAI, Google or Anthropic models.

  • China’s DeepSeek accused of stealing data. How will this affect the race against the US?

    China’s DeepSeek accused of stealing data. How will this affect the race against the US?

    In Silicon Valley, admiration for the technical prowess of Chinese start-up DeepSeek is quickly giving way to a hardline defensiveness. According to a memo to US lawmakers accessed by Reuters, OpenAI has officially raised the alarm over the methods used by its Hangzhou-based rival. The creators of ChatGPT claim that DeepSeek has not only challenged US dominance, but has done so by systematically draining the intellectual value developed by US-based labs.

    The essence of the dispute centres around the distillation process. While this is a familiar technique in the open-source world, OpenAI gives it a predatory character in this context. According to the company led by Sam Altman, DeepSeek employees were said to use extensive infrastructure, including obscure third-party routers, to programmatically circumvent security and extract data en masse from OpenAI models. The aim was to ‘feed’ the company’s own algorithms with responses generated by more sophisticated systems, which in practice dramatically reduces the time and cost of training the model while maintaining high quality results.

    For policymakers in Washington and business leaders, the message is clear: Chinese models such as DeepSeek-V3 or R1, which until recently were praised for their cost-effectiveness, may be the product of clever reverse engineering, not just a breakthrough in algorithmics. OpenAI calls this ‘free riding’, suggesting that Chinese competitors are taking shortcuts not only in terms of development, but also in terms of AI security.

    From a business strategy point of view, the OpenAI movement – to proactively remove accounts linked to distillation attempts – marks the end of an era of naive openness. If the most powerful models are to serve as free tutors for third-party rivals, the boundary between public APIs and intellectual property will become a major point of legal and political contention. For investors, the key question remains whether DeepSeek’s success is evidence of a waning US advantage, or merely a signal that the barriers to entry into the top AI league are easier to leap over using data developed by competitors.

  • One year has passed since the launch of DeepSeek. Rivals go on the offensive

    One year has passed since the launch of DeepSeek. Rivals go on the offensive

    When Hangzhou-based startup DeepSeek unveiled a model to match the world’s best for a fraction of the cost of training in early 2025, global technology markets reacted with panic. Today, a year after that event, China’s artificial intelligence ecosystem is no longer just a backdrop for one player’s phenomenal growth. Ahead of the upcoming Chinese New Year, giants such as Alibaba and ByteDance are preparing a series of launches to prove that they have learnt lessons from their smaller competitor’s success.

    The AI landscape in China has undergone a fundamental shift: the low-cost open-source model has become the prevailing standard rather than an anomaly. According to data from research group RAND, Chinese systems now operate at between one-sixth and one-quarter the cost of their US counterparts. This undermines the previous Silicon Valley paradigm of assuming that only multi-billion dollar infrastructure investments guarantee the highest quality.

    Strategic divergence: Research versus product

    Although rivals have adopted DeepSeek’s open-source model, their business strategies are clearly beginning to diverge. DeepSeek, backed by hedge fund High-Flyer Quant, remains in a unique structural position. It can prioritise pure research over immediate commercialisation, avoiding pressure from external investors.

    The situation is quite different for Alibaba or ByteDance. These companies need to prove to shareholders that AI translates into real revenues. Alibaba is experimenting with integrating the Qwen chatbot into its e-commerce systems, enabling purchases directly via voice commands. ByteDance, on the other hand, is setting its sights on dominating the video and consumer app segment – their Seedance 2.0 model is expected to generate cinema-quality content in seconds, and chatbot Doubao, with 155 million weekly users, remains the most popular AI gateway in China.

    The upcoming wave of premieres

    February 2026 promises to be a moment of truth for China’s technology industry. DeepSeek is preparing the launch of the V4 model, while Alibaba is preparing the Qwen 3.5 series with improved mathematical reasoning. Zhipu AI has already managed to present a model capable of autonomously performing lengthy tasks without additional prompting.

    This evolution shows that Chinese companies have stopped merely chasing US leaders. Instead, they are building their own path based on cost efficiency and deep integration with digital services. Market expectations are huge. If the new models deliver the promised quality, 2026 could be the moment when the centre of gravity of AI innovation finally shifts to the East.

  • AI will take people’s jobs – DeepSeek’s surprising pessimism

    AI will take people’s jobs – DeepSeek’s surprising pessimism

    China’s AI ‘little dragon’, DeepSeek, has returned to the scene after almost a year’s public absence – but its message was far from industry optimism. At the government’s World Internet Conference in Wuzhen, the company’s senior researcher, Chen Deli, gave a surprisingly pessimistic assessment of the long-term impact of the technology he himself is developing.

    DeepSeek became a global sensation in January when its low-cost, open source (open source) models showed performance in key benchmarks that outperformed leading US counterparts. Despite this success, the company has adopted a strategy of near-total silence. Its founder and CEO, Liang Wenfeng, has only appeared publicly once, in February, in a televised meeting with President Xi Jinping.

    Chen Deli’s speech at Wuzhen was therefore the first substantive voice from inside the company in months. While the researcher acknowledged that AI will be a “great help” to humanity in the short term, his predictions quickly became bleak. Chen warned that as early as five to 10 years out, technology could threaten jobs by taking over “some” of the tasks performed by humans.

    His long-term vision is even more worrying. “In the next 10-20 years, artificial intelligence could take over the rest of the work,” – Chen said, adding that society will face a “huge challenge”. In this scenario, he stressed, technology companies must take on the role of protector, although he himself is very positive about the technology but sees the negative impact it could have on society.

    This publicly expressed caution contrasts with the strategic role DeepSeek now plays in China. The company has become a symbol of the country’s technological capabilities and resilience in the face of US sanctions, particularly those relating to access to advanced chips. DeepSeek has become a key player in building the domestic AI ecosystem, optimising its models for Chinese hardware.

    Its software is compatible with processors from local champions such as Cambricon and Huawei. When DeepSeek announced a model update optimised for Chinese chips in August, it triggered a surge in the share prices of domestic component manufacturers.

    The sudden success of open-source models such as DeepSeek demonstrates that the barriers to entry in advanced AI are rapidly falling, forcing companies to immediately review their own technology strategy and rate of adoption. At the same time, an unprecedented warning from inside a leading lab signals that an ‘automation at all costs’ strategy is risky and requires boards to prepare viable plans to manage labour market transformation.

  • The paradox of the Chinese military. They are building AI on DeepSeek, but still need Nvidia chips

    The paradox of the Chinese military. They are building AI on DeepSeek, but still need Nvidia chips

    When Chinese state-owned defence giant Norinco unveiled the P60, an autonomous combat support vehicle, in February, the key element was not the armour, but its decision-making system. Capable of autonomous operations at 50 km/h, the vehicle was powered by a language model from DeepSeek – the technological pride of China’s AI sector.

    The presentation, publicised by party officials, is much more than a one-off show of force. It is a public demonstration of Beijing’s systematic effort to integrate commercial artificial intelligence into the military, with the aim of closing the gap with the US in the technological arms race. A Reuters analysis of hundreds of tender documents, patents and research papers gives an insight into the scale of these ambitions, which include autonomous target recognition and real-time battlefield decision support.

    The centrepiece of this strategy is now becoming DeepSeek. A review of this year’s orders placed by affiliates of the People’s Liberation Army (PLA) shows the growing popularity of this model, which appears in documents far more often than the rival Qwen from Alibaba. Washington is under no illusions about the nature of this cooperation. A US State Department spokesperson stated that DeepSeek “has willingly provided, and is likely to continue to provide, support for Chinese military and intelligence operations”. This is part of Beijing’s broader strategy, described as “algorithmic sovereignty” – independence from Western technology.

    AI applications go far beyond autonomous vehicles. Chinese researchers are working on drone swarms, psami-robots and advanced simulations. Researchers at Xi’an Technological University described in May a DeepSeek-based system capable of assessing 10,000 battlefield scenarios in 48 seconds. This task, they claim, would take a conventional planning team 48 hours.

    However, this technological leap faces a fundamental obstacle: hardware. Despite the move towards self-sufficiency, the Chinese military and related research institutions still rely on US chips. Tender documents and patents, filed even in June, still point to the use of Nvidia’s A100 processors, subject to US export restrictions from September 2022.

    Nvidia downplayed these reports, suggesting that “recycling small quantities of old, used products does not raise any national security concerns”. At the same time, pressure is mounting for national solutions. Sunny Cheung of the Jamestown Foundation, analysing data from the PLA’s bidding network, has seen a marked increase in the number of contractors declaring to use only Chinese hardware, such as Huawei’s Ascend AI chips, in 2025.

    Beijing is playing the game on two fronts. On the one hand, it is aggressively implementing its best domestic AI models into combat systems. On the other, it is trying to cut the umbilical cord linking it to Western semiconductors. The race for autonomous weapons, in which the US is also participating by deploying thousands of its own drones, has entered a decisive phase.

  • Huawei and Chinese AI censorship: new model to block ‘sensitive’ content

    Huawei and Chinese AI censorship: new model to block ‘sensitive’ content

    Chinese technology giant Huawei, in collaboration with Zhejiang University, has developed a new ‘safe’ version of the DeepSeek advanced language model. The project, dubbed DeepSeek-R1-Safe, aims to almost completely eliminate responses to topics deemed politically sensitive by Beijing.

    This is another step in adapting AI generative technology to the restrictive requirements of Chinese censorship.

    According to Chinese regulations, all publicly available AI models must comply with ‘socialist values’. In practice, this means strict content control and preventing the generation of information on topics such as the internal politics of the Chinese Communist Party or other issues deemed sensitive.

    Solutions such as Ernie Bot from Baidu already regularly refuse to respond to such requests.

    The model created by Huawei goes a step further. It uses 1,000 of its own Ascend AI chips to train it.

    The company claims that DeepSeek-R1-Safe has achieved a “near 100%” success rate in blocking “harmful content”, including hate speech, illegal activities and just political topics.

    However, Huawei admits that the effectiveness of the system drops dramatically – to just 40% – when problematic queries are hidden in more complex scenarios or encrypted.

    Despite this, the company claims that the model’s overall ‘security’ capability is 83%, which is expected to outperform competing solutions, such as Alibaba’s Qwen-2, by 8 to 15 percentage points. Significantly, these modifications were said to result in only a slight decrease of less than 1% in the model’s overall performance compared to its original, open-source version.

    The original DeepSeek models, created by the startup of the same name, caused quite a stir in the tech industry due to their sophistication, becoming one of the main Chinese competitors for Western AI technologies. Huawei’s creation of a ‘secure’ version shows how the Chinese market is adapting global innovations to local, political realities.

  • China’s DeepSeek heats up the debate about the cost of AI. Reveals surprisingly low cost of training

    China’s DeepSeek heats up the debate about the cost of AI. Reveals surprisingly low cost of training

    Chinese startup DeepSeek is shedding new light on the economics of the race for supremacy in artificial intelligence. In a publication in the prestigious journal Nature, the company revealed that the cost of training its R1 model, focused on reasoning, was just $294,000.

    This amount is a fraction of the estimates quoted by the US giants, where spending in excess of $100 million on flagship training is unofficially rumoured.

    This information reignites the debate about China’s real position in the global AI competition and challenges the narrative that building foundational models is reserved exclusively for players with almost unlimited budgets.

    According to the article’s authors, a cluster of 512 Nvidia H800 GPUs was used to train the R1 for 80 hours. The choice of hardware is no accident. The H800 chips are a version designed by Nvidia specifically for the Chinese market after the US blocked the export of the more powerful A100 and H100 units in 2022.

    However, the issue of DeepSeek’s access to advanced chips raised questions. US officials suggested that the company may have already sourced more powerful H100 chips after the restrictions were imposed.

    In a document accompanying the publication, DeepSeek acknowledged that in the early, preparatory stages of development it used legally owned A100 chips, but that the main R1 training had already taken place on H800 units.

    The second important thread is the technique of so-called model distillation, a process in which one AI system learns from data generated by another, more advanced model. This is a method that can significantly reduce cost and development time, but is seen as controversial when done without the consent of the original model creator.

    DeepSeek has faced allegations that it deliberately ‘distilled’ OpenAI models. The company has addressed these suggestions in a recent publication. It admitted that the training data for one of their models contained a “significant number” of responses generated by OpenAI systems.

    However, DeepSeek representatives maintain that this was not a deliberate act, but merely an accidental side effect of indexing publicly available web content.

    The disclosure of low costs, coupled with explanations of the hardware and training methods used, is a strategic move by DeepSeek. The startup is demonstrating that it is able to create competitive solutions much more cheaply, which could in future influence the dynamics of the entire AI market, which has so far been dominated by a narrative of gigantic costs and barriers to entry.

  • Huawei stumble delays DeepSeek and exposes weaknesses in China’s AI strategy

    Huawei stumble delays DeepSeek and exposes weaknesses in China’s AI strategy

    Chinese AI startup DeepSeek has had to delay the release of its highly anticipated R2 model.

    The reason appeared to be technical problems that prevented effective training on Ascend chips from Huawei. The incident is forcing the company to revert to Nvidia technology, a clear blow to China’s ambitions to become independent of US components in the key artificial intelligence sector.

    The Financial Times reports that persistent performance problems with Ascend chips have forced DeepSeek to change its strategy.

    The company opted for a hybrid solution: Nvidia’s powerful chips will be used for the extremely computationally demanding process of training the model, while Huawei’s chips will be used for less taxing inference (inference), i.e. the generation of a response by an already trained model.

    This division of tasks, while pragmatic, highlights that Chinese alternatives still lag behind US leaders in the most advanced applications. The R2 was originally scheduled for release in May.

    The DeepSeek case illustrates the broader challenge facing Chinese technology companies. On the one hand, Beijing is pushing for the use of domestic solutions, such as Ascend chips, in an effort to build technological sovereignty and raise concerns about the security of US technologies.

    On the other hand, US sanctions, although recently relaxed in the case of Nvidia’s H20 chips, still restrict access to the most efficient components from the US. Companies find themselves trapped between political directives and the technological reality that US chips remain the most efficient tool.

    Despite the political support for Huawei, it is Nvidia’s H20 chips that remain the de facto standard in the Chinese AI market. It was on them that DeepSeek trained its previous groundbreaking R1 model, as did giants such as ByteDance, Tencent and Alibaba.

    Problems with implementing a domestic alternative put DeepSeek at a competitive disadvantage at a time when rivals are launching new innovations. The delayed launch of R2 is expected in the coming weeks, but the incident calls into question the real pace of China’s hardware AI revolution.

  • DeepSeek under the magnifying glass in Europe. Germany wants Chinese AI app removed from shops

    DeepSeek under the magnifying glass in Europe. Germany wants Chinese AI app removed from shops

    The rise of Chinese artificial intelligence models in Western markets is increasingly facing a barrier – not technological, but regulatory. The latest example comes from Germany, where the country’s data protection commissioner has demanded that Apple and Google remove the DeepSeek app from their shops over fears of breaches of European data protection law.

    It is another country after Italy, the Netherlands and Belgium that is questioning the security and transparency of a Chinese AI application. DeepSeek, a startup with ambitions to rival OpenAI and Anthropic, gained notoriety earlier in the year with claims of creating a competitive AI model at a much lower cost. In Europe, however, it is mainly faced with questions about where user data goes.

    The German supervisory authority found that DeepSeek stores personal data – including the content of queries or uploaded files – on servers in China, and the company was unable to demonstrate that it guarantees a level of data protection equal to Europe’s. And this is a key condition under the RODO. Also working in the background is the concern about the broad powers of the Chinese authorities to access data – even those processed by commercial entities.

    While the decision does not yet mean a formal ban – Apple and Google are to assess the situation themselves and take action – the pressure is mounting. Italian authorities have already ordered the app to be blocked. The Netherlands has banned its use on government devices and Belgium has recommended that officials avoid it. The UK government, on the other hand, remains with a ‘threat monitoring’ stance for now, as does Spain’s consumer protection agency.

    On a geopolitical level, the case is part of a broader trend: a growing distrust of Chinese technology providers. US lawmakers are planning to introduce legislation banning the use of AI models developed in China by federal agencies. In the background, there are also reports of DeepSeek’s links to Chinese military and intelligence operations.

    For Western platforms – such as Apple and Google – the situation is another test of their ability to operate in compliance with local regulations while balancing relationships with global partners. The DeepSeek decision could set a precedent for similar cases in the future, especially as Chinese AI models increasingly try to win users outside Asia.

    For users and companies in Europe, the key question remains not so much the performance of AI, but trust in the ecosystem in which it operates. And in this case – data, once transferred to China, may not come back under European jurisdiction.

  • DeepSeek halts work on R2 over Nvidia chip shortage

    DeepSeek halts work on R2 over Nvidia chip shortage

    When Chinese startup DeepSeek unveiled its R1 language model, many experts rubbed their eyes in amazement. The model rivalled the latest OpenAI and Anthropic designs, and cost significantly less to train. However, success proved difficult to replicate – according to findings by The Information, production of the successor, the R2 model, was halted due to a shortage of Nvidia GPUs.

    DeepSeek has built its success on a huge scale – the R1 model was trained on 50,000 Hopper GPUs, including 10,000 H100, 10,000 H800 and 3,000 H20 chips. The latter – specially prepared for export to China – are particularly difficult to obtain today. Since the US imposed further export restrictions, Chinese companies have found it difficult to access even truncated versions of Nvidia’s GPUs. DeepSeek has already used up most of its available resources, serving the demand of local companies and government agencies.

    The situation is not only affecting the plans for the R2 model, but also the current performance of R1. Users are reporting drops in the quality of the model’s performance, which may indicate system overload. The company is trapped: without new GPUs, it is unable to develop the model, and declining performance is discouraging potential customers.

    While Chinese manufacturers such as Huawei offer alternative AI accelerators, their performance still lags behind Nvidia’s chips. To make matters worse, they are not compatible with the popular CUDA ecosystem, further complicating model and infrastructure migration.

    For DeepSeek, this is a serious problem. The company had a chance to become the local equivalent of OpenAI, but without continued access to advanced hardware, it may lose momentum. In practice, this confirms a broader problem for the Chinese AI ecosystem – limitations in access to semiconductor technology translate into difficulties in scaling models and services.

    In the context of the global arms race in AI, the delay of the R2 model demonstrates the importance of supply chains and the dominance of a few hardware suppliers. Even the best-designed model doesn’t stand a chance without the right computing facilities.

  • DeepSeek surprises again. Chinese startup raises the bar in the AI race

    DeepSeek surprises again. Chinese startup raises the bar in the AI race

    Chinese startup DeepSeek has quietly released an update to its R1 inference model, designated R1-0528, on the Hugging Face platform. Although the company has yet to make an official announcement or provide detailed specifications, the new version of the model was immediately noticed – and for good reason.

    In the latest LiveCodeBench ranking, which assesses AI models in terms of the quality of code generated, the DeepSeek R1-0528 ranked just behind OpenAI’s o4 mini and o3, and ahead of the Grok 3 mini from Elon Musk’s xAI and Alibaba’s Qwen 3. This is a strong indication that the Chinese startup is not only catching up with the US leaders, but can realistically compete with them.

    DeepSeek’s success so far has been about more than just the quality of its models – the startup has shown that advanced AI can be created more cheaply and quickly, without infrastructure on the scale of Google or OpenAI. The launch of the R1 model in January this year caused a noticeable stir in the technology markets, with some listed AI companies outside China scoring declines. The model challenged previous dogma: that effective AI requires unlimited computing power and billions of dollars of investment.

    Since then, Chinese companies such as Tencent and Alibaba have responded with their own models, signalling an intensification of the local AI arms race. Global players OpenAI and Google, on the other hand, have responded with price cuts and the introduction of lighter models, also indicating increasing pressure from smaller, more agile competitors.

    All of this is playing out in the shadow of anticipation for the launch of the R2, the next generation DeepSeek model, which according to earlier leaks was due to debut in May. The delay may suggest that the company is aiming higher and is aiming not just to catch up with the leaders, but to overtake them.

    Conclusions? DeepSeek is emerging as a symbol of a new era in AI – more decentralised, less capital-intensive and therefore more threatening to the existing hegemons. If R2 proves to be a breakthrough, the balance of power in the industry could be permanently altered.