DeepSeek Seeks $300M in First-Ever Fundraise at $10B+ Valuation as V4 Model Prepares to Run on Huawei Chips

For more than two years, DeepSeek did something that virtually no successful AI startup has ever managed: it turned the money away. China’s most celebrated artificial intelligence research lab, backed by quantitative hedge fund High-Flyer Capital, declined funding overtures from the country’s top venture capital firms and tech giants alike, preferring to operate with the financial backing and independence of its parent company.

That era is ending. According to a report by The Information published on April 17, 2026, DeepSeek is now in active talks to raise at least $300 million in its first external funding round, at a valuation of $10 billion or more. The move represents a significant strategic shift — and it arrives at a moment when DeepSeek is preparing to launch its most ambitious model yet, one built deliberately to prove that world-class AI can run entirely on Chinese hardware.

Why Now? The Cost of Frontier AI

DeepSeek’s decision to seek outside capital is not a sign of financial distress. Rather, it reflects the brutal economics of training and serving frontier AI models at scale. While DeepSeek’s earlier models, including the breakthrough R1 reasoning model, were famously cost-efficient — the company claimed training costs a fraction of comparable U.S. models — the next generation demands a different order of infrastructure investment.

According to people familiar with the matter, the funding would primarily go toward expanding compute infrastructure, stabilizing services that have struggled with demand surges, and accelerating the commercialization roadmap. Estimates of DeepSeek’s true valuation have ranged widely — from $1 billion to above $20 billion — making the $10 billion figure a negotiated midpoint that likely reflects both the company’s revenue traction and the strategic premium investors attach to owning a stake in China’s most internationally recognized AI lab.

DeepSeek has previously turned down multiple offers, suggesting it is now raising on its own terms with significant leverage. The company is not cash-poor; it is capacity-constrained.

The V4 Model: A Trillion-Parameter Statement

The funding timing is closely tied to the imminent launch of DeepSeek V4, which company founder Liang Wenfeng confirmed in early April is expected to ship in the final two weeks of April 2026. V4 has already been delayed twice from earlier target dates, a reflection of the complexity involved in building what is reportedly a trillion-parameter model — and in building it to run on hardware that was never designed for it.

That hardware is Huawei’s Ascend 950PR, and DeepSeek’s decision to build V4 entirely around it is the most consequential element of this story. Reuters confirmed on April 4 that DeepSeek V4 will run on Huawei’s Ascend chips, making it the first frontier-class AI model explicitly developed to operate on Chinese domestic silicon.

The engineering effort required was substantial. DeepSeek’s team had to rewrite core components of its training and inference pipeline, adapt model parallelism strategies to the Ascend architecture, and validate performance across chip configurations that differ significantly from the CUDA ecosystem its engineers were trained on. Chinese chipmakers Huawei and Cambricon received early access to V4 development resources — access that was deliberately withheld from Nvidia, according to sources cited by multiple outlets. That is a pointed reversal of the usual dynamic, in which Nvidia hardware gets first access to model development.

Semiconductor Independence as Strategic Doctrine

DeepSeek’s V4 architecture is not just an engineering project — it’s a political statement. China’s AI industry has operated under the shadow of U.S. export controls since late 2022, when the Biden administration restricted the sale of advanced Nvidia GPUs to Chinese companies. The controls, maintained and expanded under subsequent administrations, have forced Chinese tech companies to stockpile older H800 and A100 chips, pay premiums for gray-market hardware, or pivot to domestic alternatives.

For most of 2024 and 2025, the domestic alternatives were not good enough for frontier model training. That has changed. Huawei’s Ascend 950PR, which began mass production in April 2026, delivers approximately 2.8 times the FP4 performance of Nvidia’s H20 — the most advanced chip Nvidia is currently permitted to sell in China. ByteDance has committed $5.6 billion in orders for the chip, and Alibaba, Tencent, and other major Chinese technology companies have placed bulk orders for hundreds of thousands of units.

DeepSeek’s decision to build V4 on Ascend 950PR is a proof-of-concept for the entire Chinese AI ecosystem. If a model trained entirely on domestic chips can match or surpass the performance of models trained on Nvidia hardware, it validates a strategic bet that dozens of Chinese companies have been forced to make by circumstance. DeepSeek is now making it by choice.

Valuation Uncertainty and the Funding Landscape

The $10 billion valuation figure is deceptively simple for a company whose financials are almost entirely opaque. DeepSeek does not publish revenue figures. Its primary monetization channels — API access, enterprise licensing, and integration with High-Flyer’s trading operations — are not publicly disclosed.

What is known is that DeepSeek’s models command enormous global usage. Its API is deployed by developers across dozens of countries, and the open-source release of R1 in January 2025 triggered what became known in financial markets as the “DeepSeek shock” — a sell-off in Nvidia and other AI infrastructure stocks that wiped hundreds of billions in market value and prompted a reassessment of the capital intensity required for frontier AI.

That global recognition translates into investor demand regardless of disclosed revenue. Estimates have ranged from $1 billion to well above $20 billion, with the $10 billion figure representing a plausible anchor point that Chinese institutional investors, sovereign wealth vehicles, and strategic corporate partners could underwrite. DeepSeek reportedly has significant flexibility in choosing its investors and structuring the round, given the inbound interest it has historically received and declined.

What V4 Must Prove

The stakes for the V4 launch are exceptionally high. DeepSeek’s reputation was built on a series of counterintuitive demonstrations: models that achieved performance competitive with OpenAI’s best, at training costs that were orders of magnitude lower. Each release has sharpened a narrative — that intelligence scales with architecture and data quality, not just raw compute spending.

V4 must sustain that narrative while operating on new hardware. If the model performs at or above the level of its contemporaries from Anthropic, OpenAI, and Google — models trained on the most advanced Nvidia clusters — it will mark the first definitive proof that domestic Chinese chips can power frontier AI. That would validate not just DeepSeek’s engineering choices but China’s broader semiconductor strategy.

If it falls short, the story becomes more complicated. A model that performs well but not best-in-class on domestic hardware would still be commercially valuable, but it would leave open the question of whether the performance gap between Huawei’s chips and Nvidia’s best is recoverable — or whether export controls have created a ceiling that domestic innovation cannot yet break through.

The $300 million being raised is, in some sense, a bet on the answer being optimistic. DeepSeek is funding its next chapter on the assumption that V4 will prove what it needs to prove — and that the evidence will unlock significantly more capital, at a significantly higher valuation, in the rounds that follow.