DeepSeek Restarts $74 Billion Fundraising While Announcing Major API Price Hikes
Claire Weston
DeepSeek has restarted its second funding round at a target valuation of roughly $74 billion, while announcing a broad API price increase on the same day — raising its price tag and testing customers' tolerance simultaneously, as the open-source company's monetization dilemma moves to center stage.
Where is the money coming from — and on whose terms?
The round targets roughly ¥50 billion ($7.4 bn) in new capital at a ¥500 billion valuation — up about 43% from the first round's ¥350 billion in under two months.
First-round backers included Tencent, CATL, NetEase, JD.com, China's National AI Industry Investment Fund, Monolith Management, and IDG Capital. This round is expected to close by month-end, but the process is deliberately opaque — some participating firms were not even told the deal had restarted.
This means → DeepSeek is using information asymmetry as a filter: only investors willing to move fast on limited disclosure get an allocation.
Why does Liang Wenfeng have absolute control?
Nearly all first-round capital was funneled into a limited partnership controlled personally by Liang Wenfeng (梁文锋), with a five-year lock-up and no voting rights.
The sole exception: the National AI Industry Investment Fund, which holds equity directly and retains a vote. Liang currently owns roughly 84% of equity and near-total voting control.
In plain terms = investors wrote the checks but got no steering wheel — they can neither sell nor vote for five years. That is how Liang was able to unilaterally halt the deal after a leak.
What leaked — and why does it matter?
Minutes from Liang's May 20 closed-door investor briefing surfaced on GitHub. Key admissions: China's AI lags the U.S. in compute and capital, not talent; DeepSeek is heavily dependent on Nvidia hardware; Huawei alternatives fall visibly short; the company may have acquired what he called "non-compliant" processors.
Liang said GPUs are the best use of capital. If supply allowed, he would convert all funds into Nvidia hardware; spending ¥20 billion a year on processors would be "a win for the procurement team."
This reflects an awkward reality: the founder of China's strongest AI lab has stated on record that he cannot run without Nvidia chips — the very hardware at the center of U.S. export controls.
How serious is the legal risk around chip supply?
U.S. federal investigators began probing as early as 2025 whether Singapore intermediaries re-routed restricted Nvidia chips to DeepSeek.
The U.S. House Select Committee on China formally concluded that DeepSeek's models run on tens of thousands of export-controlled Nvidia chips, labeling the company a vehicle for "subverting" U.S. export controls.
This means → DeepSeek's chip sourcing is no longer just a procurement issue — it is a geopolitical target. The higher the valuation climbs, the more damaging a supply-chain cutoff becomes.
What is the path away from Nvidia dependence?
Reuters reports that DeepSeek is quietly hiring chip designers to develop in-house inference chips.
The company also plans to expand headcount and build data centers, including a large facility in Inner Mongolia expected to cost several billion dollars.
Yet Liang himself conceded that even ¥50 billion is not enough to train a model matching the largest U.S. systems — he estimates that would require roughly 50,000 Nvidia GB300s or 200,000 Huawei Ascend 950s, excluding pre-training experiments.
Price hikes and open source — how does the math work?
DeepSeek had dragged ByteDance, Tencent, and even major U.S. AI providers into a price war with API rates as low as roughly $0.14 per million tokens. But demand consistently exceeds infrastructure capacity, and rock-bottom pricing holds little appeal for pre-IPO investors.
The catch: DeepSeek's model weights are open-source. Any data center in the world can self-host the same model. Push prices too high and customers simply run it themselves.
In plain terms = this is the classic open-source trap — you gave the code to the world, then tried to charge for the hosted version, while customers can bypass you at any time. As DeepSeek prepares a Shanghai STAR Market IPO, it must find the price point that boosts revenue without driving users to self-deployment.
Content is for reference only, not financial advice.