From "breaking through computing power" to "ecosystem landing": the two ace cards behind DeepSeek's 10 billion financing - Huawei and XUNCE (03317)

Two years later, the "Price Butcher" of the AI world has finally extended an olive branch to the capital market. According to "The Information" and several other media outlets, DeepSeek is launching its first external financing since its establishment, with a target valuation of no less than $10 billion and plans to raise at least $300 million. The shock of this news lies not only in its massive valuation, but also in the fact that it breaks the long-standing myth of self-sufficiency that DeepSeek has upheld. At a time when peers are burning money and raising billions in parameters, DeepSeek had previously rejected the advances of many capital giants with the backing of its parent company, Illusionary Quantitative. This shift towards capital marks a key step for the company, known for its technological geek culture, towards commercialization and ecological development. In this potential financing game, two key words are gradually emerging: computing power and tokens. The former determines whether DeepSeek can break free from the NVIDIA ecosystem, and the latter determines whether it can evolve from a "technical powerhouse" to a "business king". What DeepSeek truly needs is not just ordinary financial investors, but a breakthrough in computing power and the construction of a token ecosystem. Models alone cannot make a king; they must rely on data, scenarios, and actual implementation. Who will be the ecological partner? Perhaps this is the key puzzle that will determine whether DeepSeek can complete this transformation. 1. The truth about financing: Farewell to the "ideal state", hunger for computing power, and talent defense battle DeepSeek's decision to seek financing at this time is not coincidental. First, the V4 trillion parameter challenge. The delayed V4 of DeepSeek is expected to debut at the end of April. The total number of parameters has increased to around 1 trillion (MoE architecture, with only 370 billion parameters activated per token), the context window has been expanded to 1 million tokens, and it now supports native multimodality (text, images, videos) for the first time. Internal benchmarks show that SWE-bench is over 80% and HumanEval is at 90%. Second, the cost of computing power is no longer covered by quantitative funds. If the competition for large models two years ago was about algorithmic ingenuity, the current competition is a naked war of computing power consumption. The daily token call volume in China has exceeded 14 trillion in 2026, an increase of over 1000 times in two years. The training and deployment of V4 requires tens of thousands of computing cards, with investment scale jumping to the billion-dollar level. Third, talent defense battle. After the success of the R1 model in early 2025, DeepSeek experienced a loss of core talent. Key contributors such as Luo Fuli joined Xiaomi, and Guo Daya joined ByteDance. In the capital frenzy of the Chinese AI battlefield, even with no equity binding and market-beating salaries, it is difficult to retain top talent even with an "ideal" environment. Initiating financing and setting up equity incentives are necessary moves for DeepSeek to achieve commercial transformation and stabilize its existing team. 2. The trump card for breaking through computing power: Huawei Ascend and "de-NVIDIA-ization" The core reason for the delay in V4 is not the model itself, but a major hardware migration. According to Reuters, V4 will run on Huawei's latest Ascend chips. Engineers at DeepSeek spent a lot of time rewriting core code to transition from NVIDIA's CUDA ecosystem to Huawei's CANN architecture. If V4 can run competitively on Huawei chips, it will be the world's first cutting-edge AI model not dependent on NVIDIA. NVIDIA CEO Jensen Huang stated in a recent interview that the new model based on the Huawei platform "will be a bad result for the United States". Implicitly, once an AI model is optimized to run best on Chinese hardware, the moat of American chips will no longer be solid. V4 is planned to be released in two versions: a full version with over a trillion parameters, optimized for advanced reasoning and complex code tasks, aimed at Huawei Ascend chips; a lightweight version with about 200 billion parameters, aimed at general dialogue and API services, capable of running on other domestic chips. In terms of open source, V4 is planned to be released under the Apache 2.0 license. Various signs indicate that V4 is no longer confined to the laboratory but is preparing for large-scale deployment. 3. The trump card for landing the ecosystem: XUNCE's scenario tokens While Minimax and KNOWLEDGE ATLAS generate revenue by selling tokens and calling APIs, DeepSeek faces a more fundamental dilemma: its tokens are sold too cheaply. The price butcher was once its market-sweeping weapon, but it also led it into a quagmire of high concurrency and low margins. In addition to breaking through computing power, DeepSeek also needs to find ecosystem partners who can define the value of data and tokens. To understand this, we must first return to the core metric of the AI industry: Tokens. Tokens are evolving from simple technical measurement units into core assets. In this new value system, the industry has formed a clear three-tiered structure: Bottom layer (computing power tokens): NVIDIA, Huawei, and major cloud providers, selling computing power itself. The value of this layer is determined by chip performance and cluster size, competing on capital density. Middle layer (model tokens): DeepSeek, OpenAI, Minimax, and other large model vendors, selling intelligent capabilities. The value of this layer is determined by model effectiveness and reasoning costs, competing on algorithm efficiency - DeepSeek has already excelled in this area. Top layer (scenario tokens): This is the core of the current value reassessment. If generic model tokens solve the "intelligence" problem - giving answers to given questions; scenario tokens solve the "efficacy" problem - whether tokens truly produce measurable business results in specific business processes? A common pain point in the industry is that generic large models do not always fit well in vertical industries, leading to a waste of tokens in ineffective trial and error and "illusions". Companies pay huge bills but may not necessarily see real productivity improvements. In other words, the value of model tokens is depreciating, while the value of scenario tokens is appreciating. Also stemming from the financial industry, XUNCE Technology, is known as the "first stock of tokens". Its core competence lies in transforming heterogeneous real-time data from finance, energy, manufacturing, Siasun Robot & Automation, bio-medical, and other fields into high-quality scenario tokens that large models can directly consume through data cleaning, governance, and feature engineering. The barriers mainly lie in three aspects: data pipelines as infrastructure; pricing rights of scenario tokens; and the closed loop from data to decisions. If DeepSeek wants to establish a first-mover advantage in the token ecosystem, it will need to find strategic partners with high barriers to scenario data and scenario token value anchoring. And XUNCE, or another key partner in this direction. In conclusion: who is leading the next round of AI competition? DeepSeek's financing this time, on the surface, is due to a lack of money, but at a deeper level, it is due to a lack of "computing power" and "ecological position". The landscape of general large models has been basically settled, and the diminishing marginal effects of burning money are decreasing; while at the application layer, the golden decade has just begun. This reveals a new trend in the AI industry: the competition in Chinese AI is shifting from the "battle of a hundred models" to the "Token ecosystem battle". And this time, players with data and scenario tokens are truly holding the power of pricing.