DeepSeek’s Independence Is Over — and the Chip War Just Got Its Most Dangerous Player
Direct Answer: What is happening with DeepSeek’s funding and why does it matter?
DeepSeek — the Hangzhou-based AI lab that shocked Silicon Valley in January 2025 by training a frontier-class model for $5.6 million — is raising outside capital for the first time in its history. Reported on April 22, 2026 by The Information, Bloomberg, and confirmed by multiple sources, the funding round targets a valuation exceeding $20 billion, more than double the $10 billion floor reported just seven days earlier. Tencent Holdings is proposing to acquire up to a 20% stake; Alibaba is also in advanced talks. The valuation surge reflects extraordinary investor demand — investment firm partners booked emergency flights to Hangzhou from Beijing, Shanghai, and Shenzhen when the round was confirmed. Simultaneously, DeepSeek is executing a complete migration from Nvidia’s CUDA framework to Huawei’s Ascend 910C chips — a full rewrite of its technology stack driven by US export controls. Nvidia CEO Jensen Huang said on April 22 that this outcome would be “horrible.” The sovereignty implication is direct: if DeepSeek successfully trains frontier AI on Huawei silicon, the US export control strategy that assumed China could not build AI without Nvidia chips will have failed.
“It would be a horrible outcome if DeepSeek optimised its models to run primarily on Huawei chips rather than American hardware.” — Jensen Huang, CEO, Nvidia, April 22, 2026
The Vucense 2026 China AI Sovereignty Index
How China’s leading AI labs compare on capital independence, chip sovereignty, and strategic exposure to US export controls — after DeepSeek’s funding pivot.
| Company | Backing | Nvidia Dependency | Domestic Chip Strategy | Open Source | US Sanction Risk | Sovereign Score |
|---|---|---|---|---|---|---|
| DeepSeek (pre-round) | High-Flyer hedge fund only | High (CUDA-trained, now migrating) | Full Huawei Ascend migration underway | ✅ Open weights | High (military firm allegation) | 61/100 |
| DeepSeek (post-round) | Tencent + Alibaba + High-Flyer | Migrating away from Nvidia | Huawei Ascend 910C | ✅ Open weights | High | 54/100 |
| MiniMax | Public (HK, $30B+ valuation) | Mixed | Partial domestic | Partial | Moderate | 48/100 |
| Zhipu AI (GLM-5.1) | Public (HK, $50B+ valuation) | Low (GLM-5.1 trained on Huawei Ascend) | Full Huawei Ascend | ✅ MIT license | Moderate | 71/100 |
| Moonshot AI (Kimi) | VC-backed ($18B target) | High | Limited | ❌ | Low–Moderate | 39/100 |
| ByteDance AI | ByteDance (private) | High | Partially diversifying | ❌ | High | 31/100 |
Sovereign Score methodology (China AI context): weighted across capital independence from state (25%), chip sovereignty from US export controls (35%), open-source access (20%), US sanction exposure (20%). Lower US sanction risk and higher domestic chip capability = higher score in this context.
Analysis: Why DeepSeek Broke Its Founding Principle
DeepSeek’s decision to accept outside capital is not a routine fundraise. It is a strategic pivot driven by three compounding pressures that founder Liang Wenfeng could no longer solve with High-Flyer’s hedge fund profits alone.
Pressure 1: The talent war. DeepSeek’s core threat is not capital — High-Flyer achieved a 56.6% average annual return in 2025, generating an estimated $5.7 billion in profit and injecting over $700 million annually into DeepSeek. The threat is human capital. ByteDance, Alibaba, Tencent, and the Chinese arms of global AI labs are offering DeepSeek’s researchers salaries two to three times higher than what the lab pays. Without market-rate equity and compensation structures — which require external capital to establish a proper valuation — DeepSeek cannot retain the team that built R1 and V3. The funding round’s primary purpose, according to multiple sources cited in BigGo Finance’s analysis of 36Kr reporting, is to give employees market-based option pricing. The money is secondary. The retention mechanism is primary.
Pressure 2: The inference cost cliff. DeepSeek trained its V3 model for a reported $5.6 million — far below the billions Western labs spend — and made it available for free. That was possible because training is a one-time cost. Inference is not. Every query to DeepSeek’s chatbot, every API call, every model deployment in production — these ongoing costs compound at scale. DeepSeek’s chatbot went viral globally in January 2025. Its user base exploded. And unlike OpenAI or Anthropic, it generated essentially zero revenue from this scale because its models are free and open-source. High-Flyer’s profits could fund research. They cannot fund the inference infrastructure for a globally viral AI service indefinitely. External capital solves this.
Pressure 3: The Nvidia migration. The most operationally urgent pressure is the chip transition. US export controls block DeepSeek from purchasing Nvidia’s most advanced AI accelerators — the H100 and H200 GPUs that OpenAI, Anthropic, Google DeepMind, and every Western frontier lab use for training. DeepSeek trained its earlier models on chips that are now subject to retrospective enforcement scrutiny. Going forward, it has no path to Nvidia at the frontier. It is therefore executing a complete migration to Huawei’s Ascend 910C chips — rewriting operator libraries, communication libraries, and vast amounts of low-level code to replace Nvidia’s CUDA framework with Huawei’s CANN framework. This is not a chip swap. It is rebuilding the engine of a plane in flight. It costs money, it takes engineers, and it is directly responsible for the delay of DeepSeek’s V4 model.
The Sovereign Perspective
-
DeepSeek Is Becoming China’s National AI Lab by Default. Tencent’s proposal to acquire 20% of DeepSeek is not simply an investment thesis — it is a strategic positioning move by one of China’s most politically connected technology companies. A 20% Tencent stake in DeepSeek would create a structural alignment between the lab and the Chinese government’s AI ambitions that no amount of government funding could achieve more cleanly. DeepSeek’s open-source models are already deployed across Chinese enterprises, government departments, and military-adjacent institutions. Tencent’s involvement converts an independent research lab into part of China’s sovereign AI infrastructure stack.
-
The Huawei Pivot Is the Chip War’s Pivotal Moment. Jensen Huang’s “horrible outcome” comment is the most revealing public statement from a US tech CEO about China’s AI development in years. Nvidia’s business model depends on AI training happening on its GPUs. If DeepSeek proves that frontier AI can be trained on Huawei Ascend chips — and deploys that capability at scale — the export control strategy designed to prevent China from building AI capability collapses. Not because the controls were wrong, but because they accelerated China’s domestic alternative. The US banned Nvidia exports to China; China funded Huawei’s chip development; DeepSeek migrated to Huawei; and if it works, the ban achieved the opposite of its intent.
-
The “Zero Revenue Paradox” Is the Sovereignty Opportunity. DeepSeek’s open-source approach — models freely available, chatbot free to use — means its AI is accessible to every developer on earth without payment or account creation. This is the precise model that makes DeepSeek’s technology a genuine challenger to OpenAI and Anthropic’s proprietary ecosystems. For Vucense readers building sovereign AI stacks, DeepSeek V3.1 and its successors are the models most likely to provide frontier-competitive performance under completely open terms — as long as you can run them on your own hardware and are comfortable with the geopolitical context of their origin.
The Valuation Doubles in a Week: What That Signals
The speed of the valuation revision is the most statistically unusual aspect of this story. Seven days separated a $10 billion floor from a $20 billion target. That delta requires explanation.
The immediate trigger was confirmation that the round was actually happening. DeepSeek had previously declined multiple investment approaches from leading Chinese venture firms and technology groups. Liang Wenfeng’s shift from rejecting capital to actively seeking it created a compressed competitive dynamic among potential investors — the investment firms that had been tracking DeepSeek for months moved simultaneously when the window opened.
The benchmarking logic is straightforward. MiniMax went public in Hong Kong at under $10 billion and now trades above $30 billion. Zhipu AI listed at under $10 billion and now trades above $50 billion. DeepSeek has technically superior models than both — its V3 architecture’s efficiency benchmarks are publicly documented and widely validated. If Zhipu is worth $50 billion with inferior models, DeepSeek’s $20 billion target is conservative, not aggressive. The remaining variable — the one investors are pricing with caution — is monetisation. DeepSeek has essentially zero revenue from its AI products. Its value is entirely technical and strategic.
The Nvidia CEO’s April 22 comment landed on the same day as the Bloomberg and The Information reports, creating a single news cycle that simultaneously validated DeepSeek’s strategic importance and raised the geopolitical stakes. Jensen Huang does not make casual public statements about competitors. His “horrible outcome” framing tells you exactly how seriously Nvidia’s strategic planning team takes the Huawei Ascend migration scenario.
The Huawei Ascend Migration: What It Actually Involves
DeepSeek’s transition from Nvidia CUDA to Huawei CANN is the technical story that most news coverage treats as a footnote. It deserves detailed treatment because it is the hinge on which China’s AI sovereignty turns.
Nvidia’s CUDA framework is not just a driver. It is a complete software ecosystem — tens of thousands of libraries, optimisations, debugging tools, profiling systems, and third-party integrations built over fifteen years — that makes Nvidia GPUs the path of least resistance for every AI lab on earth. AI researchers write CUDA code almost by default. The entire open-source AI ecosystem — PyTorch, JAX, HuggingFace, vLLM, and every model training framework — is CUDA-optimised first.
Huawei’s CANN framework (Compute Architecture for Neural Networks) is the Ascend chip’s answer to CUDA. It is technically functional — Zhipu AI’s GLM-5.1 model (744 billion parameters, SWE-Bench Pro top scorer) was trained entirely on Huawei Ascend chips, providing the first independent validation that Ascend-based frontier training is possible. But CANN lacks CUDA’s ecosystem depth. Migrating requires rewriting operator libraries (the mathematical primitives that underlie every neural network layer), communication libraries (which coordinate gradient updates across thousands of chips during training), and vast amounts of optimisation code that was written assuming CUDA’s specific memory architecture.
DeepSeek’s V4 model was expected in February 2026. It has been repeatedly delayed. Sources confirm the delay is architectural — the team is rewriting the training stack, not struggling with algorithmic design. The training instability and subpar inter-chip communication speeds reported on Ascend 910C chips are engineering problems, not fundamental capability gaps. They are solvable. The question is timeline and cost.
If DeepSeek solves them — and the evidence from Zhipu’s GLM-5.1 suggests they are solvable — the outcome is a frontier AI lab that trains and deploys models entirely on Chinese hardware, with open-source models available globally. That is Jensen Huang’s “horrible outcome” made real.
What DeepSeek’s V3.1 and V4 Mean for Open-Source AI Users
For users and developers who rely on open-weight models, DeepSeek’s funding trajectory has direct practical implications.
DeepSeek V3.1 — the current release — is available as open weights and performs competitively with GPT-4o on reasoning and coding benchmarks. It requires substantial GPU memory to run locally (approximately 400 billion parameters), placing it in the same hardware tier as Llama 4 Maverick: accessible to well-funded small teams or via cloud API (Groq, Together AI, Fireworks), not individual desktops.
DeepSeek V4 — the delayed next release — was designed to demonstrate Ascend-native training. When it ships, its benchmark performance on Chinese hardware will be the first real-world validation of whether Huawei’s ecosystem can produce models that match or exceed Nvidia-trained equivalents at scale. This is the benchmark event that matters most for the chip war’s outcome.
For users choosing between DeepSeek and Meta’s Llama 4 as their primary open-weight model: the practical choice today remains Llama 4 Scout or Maverick (better CUDA ecosystem support, runs on more hardware configurations) or Gemma 4 (Apache 2.0, consumer hardware). DeepSeek’s open weights are technically impressive but the geopolitical context — US military designation allegations, Huawei chip provenance, potential future sanctions — makes them a more complex procurement decision for regulated industries and government entities.
Actionable Steps: What This Means for AI Developers and Enterprises
1. If you use DeepSeek’s API or models commercially, document your risk assessment now. The ongoing US legislative push to designate DeepSeek as a “Chinese military firm” — if successful — would make its API inaccessible for US companies and potentially any company with US operations. Document your current usage, evaluate alternative models for your use cases, and ensure you have a migration path that does not depend on DeepSeek availability.
2. Monitor Huawei Ascend’s ecosystem development as the chip war’s leading indicator. The question of whether Huawei Ascend can match Nvidia for frontier AI training will be answered definitively when DeepSeek V4 ships. Subscribe to updates from DeepSeek’s research blog and the Huawei Ascend developer community. V4’s benchmark performance on CUDA-equivalent tasks will be the single most important AI hardware data point of 2026.
3. For open-source AI users: maintain model portfolio diversity. If DeepSeek’s open weights are part of your AI toolkit, ensure you also have local deployments of Gemma 4 and Mistral models that carry no geopolitical exposure. Model diversity is sovereignty: no single open-weight model should be a single point of failure in your AI stack.
4. For investors tracking AI: the $20B DeepSeek valuation sets the floor for Chinese AI. MiniMax at $30B+, Zhipu at $50B+, and DeepSeek targeting $20B+ establish a clear valuation band for China’s frontier AI tier. Moonshot AI’s $18B target and ByteDance’s AI spending at $23 billion annually complete the picture. China’s AI market has a coherent valuation structure for the first time. For investors with exposure to Chinese tech through BABA or Tencent, the DeepSeek stake would represent the most technically credible AI asset either company could acquire.
5. Evaluate whether your AI sovereignty strategy accounts for geopolitical supply chain risk. The DeepSeek story is a concrete demonstration of how AI infrastructure is subject to geopolitical disruption: US export controls blocked Nvidia; DeepSeek migrated to Huawei; Nvidia warned publicly. Your sovereign AI stack should be resilient to vendor changes driven by policy, not just business decisions. Open-weight models on hardware you control are the only architecture that eliminates this risk entirely.
FAQ: DeepSeek’s Funding Round and China’s AI Sovereignty
Q: Why has DeepSeek never raised outside capital before? DeepSeek is owned by High-Flyer Capital Management, a quantitative hedge fund founded by Liang Wenfeng. High-Flyer achieved a 56.6% average annual return in 2025, managing over $10 billion in assets and generating enough profit to fund DeepSeek’s research independently. Liang’s stated philosophy was that external investors seeking time-bound financial returns were incompatible with DeepSeek’s long-term research orientation. The decision to raise now reflects three structural changes: the inference cost of serving a globally viral product at scale exceeds what High-Flyer can subsidise, the talent war requires market-rate equity structures, and the Huawei chip migration requires capital beyond research budgets.
Q: Why does Tencent want a 20% stake specifically? A 20% stake would make Tencent DeepSeek’s largest external shareholder and create a formal governance relationship between China’s most politically connected technology company and its most technically capable AI lab. DeepSeek is reportedly resistant to ceding this level of control — the Information’s source notes DeepSeek “isn’t keen” on the 20% figure. The final stake will likely be smaller. But Tencent’s interest in the largest possible stake reflects its strategic calculation that DeepSeek’s open-source models and Huawei-native training capability make it the most valuable AI asset in China’s domestic market.
Q: What is Huawei Ascend and how does it compare to Nvidia H100? Huawei’s Ascend 910C is China’s most advanced domestically produced AI accelerator. Zhipu AI’s GLM-5.1 — a 744B parameter model that scored #1 on SWE-Bench Pro in April 2026, edging GPT-5.4 and Claude Opus 4.6 — was trained entirely on Ascend chips, providing the first independent validation that Ascend-based frontier training is viable. Technical comparisons with Nvidia’s H100/H200 show Ascend trailing on raw floating-point throughput but closing the gap on memory bandwidth efficiency. The CANN software ecosystem is significantly less mature than CUDA, which is why DeepSeek’s migration is causing V4 delays — it is an ecosystem problem, not a hardware capability problem.
Q: Is DeepSeek safe to use for enterprise AI workloads? For non-US companies without regulatory restrictions on Chinese technology: DeepSeek V3.1 is technically excellent and available as open weights under permissive terms. For US companies and regulated industries: the pending US legislative push to classify DeepSeek as a Chinese military firm, combined with its provenance from a Nvidia-export-control-adjacent environment, makes it a higher-risk procurement choice than EU or US-origin open-weight models. The risk is regulatory, not technical. US government contractors and defence-adjacent enterprises should treat DeepSeek as unavailable pending legal clarity.
Q: What happens to DeepSeek’s open-source models if the company accepts Tencent investment? DeepSeek’s commitment to open-source has been a founding principle. External investment does not automatically change the licence terms on existing released weights. However, future models — V4 and beyond — may face pressure from investor governance structures to adopt more restrictive licences that enable commercial monetisation. The precedent from Meta’s Llama 4 EU ban shows that “open” models can carry restrictive terms despite broad availability. Users who depend on DeepSeek’s open weights should monitor V4’s licence carefully when it ships.
Q: How does this affect the global AI race? DeepSeek’s funding round converts the global AI race from a two-tier competition (US frontier vs. everyone else) into a three-tier competition: US proprietary (OpenAI, Anthropic, Google), China frontier (DeepSeek, Zhipu, MiniMax), and global open-weight (Mistral, Gemma, Qwen). The Huawei chip migration’s success or failure will determine whether China’s frontier tier can sustain independent development at scale. If it succeeds, the US AI export control strategy has failed at its primary objective. If it fails — if V4 underperforms significantly on Ascend chips — US controls will have successfully constrained Chinese frontier development for at least another two to three years.
Related Articles
- Best Open-Source AI Models in April 2026 — Ranked by Sovereignty
- China’s Robot Just Beat the Human World Record — Now It’s Coming for Your Workplace
- Amazon Just Bet $33 Billion on Anthropic — and Claude Is Now AWS Infrastructure
- SpaceX Just Bet $60 Billion on Cursor — and It Changes AI Coding Forever
- America Just Banned AI Data Centers — and Your State May Be Next
