Hangzhou-based artificial intelligence start-up DeepSeek has restarted its second funding round, seeking close to $8 billion (£6 billion), after pausing the process in July when remarks the company made to prospective investors were leaked, according to local media reports.
The company is targeting a valuation of nearly 500 billion yuan ($74 billion, £55 billion), Chinese business magazine Caijing reported, citing unnamed sources. Bloomberg separately reported that Monolith Management, an investment firm that has backed Beijing-based Moonshot AI, the developer of the Kimi line of models, is in talks to contribute to the round.
Neither DeepSeek nor Monolith has publicly confirmed the terms, and the identities of other prospective participants have not been disclosed. The resumption of talks suggests the company has moved past the disruption caused by the earlier leak, which had prompted it to suspend discussions with investors in the middle of the summer.
A valuation climbing fast between rounds
DeepSeek closed its first funding round earlier in the summer at a valuation of roughly 350 billion yuan, meaning the new round would represent an increase of more than 40 per cent in a matter of months. That trajectory reflects both the scarcity value of a Chinese company with a credible frontier-model research programme and the broader re-rating of AI assets globally over the past two years.
At around $74 billion, DeepSeek would sit among the most valuable privately held artificial intelligence companies in the world, though still well below the valuations attached to leading American labs in recent private transactions. The figure also places it in the top tier of Chinese technology unicorns, alongside established platform companies and a small group of well-funded model developers.
The size of the round is notable in a market where several Chinese AI start-ups have raised capital in smaller, more frequent tranches. A single raise of nearly $8 billion would give DeepSeek a war chest comparable to the annual capital expenditure budgets of some mid-sized technology companies.
Monolith Management and the Moonshot connection
Monolith Management's reported involvement adds an unusual competitive dimension to the round. The firm is an existing investor in Moonshot AI, one of DeepSeek's most visible domestic rivals, and the developer behind the Kimi series of models. Should the talks conclude successfully, Monolith would hold stakes in two of China's most closely watched model labs.
That kind of cross-portfolio position is not unprecedented in venture investing, but it is closely watched in a sector where pricing, talent and computing capacity are all contested. Moonshot has positioned Kimi as a long-context assistant for consumers and enterprises, while DeepSeek has built its reputation on open-weight models and aggressive cost efficiency.
Representatives for Monolith have not commented on the reported discussions, and it remains unclear how much of the round the firm would take up or whether other institutional and strategic investors are participating.
Inner Mongolia data centre and a 1 gigawatt ambition
A significant share of the new capital is expected to fund infrastructure. DeepSeek reportedly plans to build a large data centre in Inner Mongolia as part of a broader effort to secure access to 1 gigawatt of computing power, a target that could be met partly through leasing capacity from third-party operators rather than wholly owned facilities.
Inner Mongolia has become a magnet for Chinese compute investment because of its cool climate, which reduces cooling costs, and its access to cheap electricity, much of it generated from renewable sources such as wind and solar. The region sits at the heart of a national programme that encourages data centres to be built in the country's western interior, where power is abundant, while workloads are consumed in the densely populated east.
Building at that scale requires more than land and electricity. Training and serving frontier models demand dense clusters of advanced accelerators, high-bandwidth networking and substantial engineering capacity, and the lead times on all three are long. Securing a gigawatt of computing power, whether owned or leased, would give DeepSeek a meaningful increase in capacity for both training runs and the inference workloads that generate revenue.
The company has not disclosed the planned size, timeline or cost of the Inner Mongolia facility, nor how much of the round has been earmarked for it.
A price rise that signals a strategic shift
Alongside the infrastructure plans, DeepSeek told users last Thursday that it intended a substantial increase in prices, without specifying what the new pricing structure would look like or when it would take effect. The notice marks a departure for a company that built its reputation on undercutting rivals.
DeepSeek currently charges $0.14 per million input tokens and $0.28 per million output tokens. By comparison, Moonshot's latest Kimi model is priced at $3 and $15 respectively, while Anthropic's newest model charges $10 and $50. The gap between DeepSeek and its international peers has been one of the widest in the industry.
That pricing strategy was made possible by an unusually efficient training approach. DeepSeek became known for training high-performance models at low cost and offering them cheaply, shaking assumptions held by Western companies that have been ploughing hundreds of billions of dollars into infrastructure for training and deploying models. Its published work on mixture-of-experts architectures and other efficiency techniques has been widely studied and, in some cases, replicated.
Even so, low prices and high usage volumes are difficult to sustain once inference becomes the dominant cost. Serving millions of users and thousands of business customers requires continuous access to accelerators, and pricing below the true cost of delivery is a strategy that typically has a shelf life. The announced increase suggests DeepSeek is moving from a land-grab phase toward something closer to commercial discipline.
From a quant hedge fund to a frontier lab
DeepSeek was founded in 2023 by Liang Wenfeng, who also founded the quantitative hedge fund High-Flyer. The lab grew out of the fund's internal AI research effort, and it inherited both a culture of engineering pragmatism and a pool of computing resources that had been assembled for financial modelling.
The company's early releases focused on coding and mathematics, areas where model performance can be measured precisely and where customers are willing to pay. It then moved into general-purpose models, publishing open weights that allowed developers to download, modify and deploy them without paying for API access. That openness helped the company build an international developer following far larger than its headcount would suggest.
In 2024, DeepSeek's low-cost model releases helped trigger a price war among Chinese cloud providers, with larger companies cutting rates for their own models in response. The pattern repeated when the company released a reasoning-focused model that attracted intense global attention and briefly pushed its consumer app to the top of download charts in several markets.
Compute constraints and the chip question
DeepSeek's efficiency-first approach has been shaped in part by circumstance. Successive rounds of export restrictions have limited access to the most advanced accelerators for Chinese companies, pushing labs toward workarounds: stockpiling chips purchased before restrictions took effect, designing around interconnect bottlenecks, and working closely with domestic chip suppliers.
Those constraints raise the stakes of the Inner Mongolia project. A facility designed around domestically produced accelerators would reduce exposure to import controls but would also require extensive software optimisation to reach competitive performance. A facility built on leased third-party capacity would offer flexibility but less control over cost and availability.
The planned investment also comes as Chinese authorities have encouraged the construction of large-scale computing clusters as public infrastructure, with local governments offering land, power and subsidies to attract operators.
The wider Chinese AI funding race
DeepSeek is one of a small group of Chinese labs competing for capital, talent and compute. Moonshot AI, Zhipu AI, MiniMax and StepFun have all raised substantial sums over the past two years, often from a mix of venture funds, internet majors and state-linked investment vehicles. Alibaba and Tencent have continued to fund both their own models and external challengers.
The funding environment has been uneven. Some start-ups have struggled to close rounds at their previous valuations, while a handful of leaders have been able to command premium terms. DeepSeek's ability to return to market months after pausing, at a higher valuation, is a signal about where it sits within that hierarchy.
The comparison with American spending remains stark. US hyperscalers have committed hundreds of billions of dollars to data centre construction, and the largest AI labs are raising capital in increments that dwarf most Chinese rounds. Chinese labs have responded by emphasising efficiency, open weights and lower inference costs, an approach that has found traction with developers worldwide even as it compresses margins.
What the capital is for
Beyond the Inner Mongolia site, the proceeds are likely to fund accelerator procurement, networking equipment and the engineering teams needed to operate clusters at scale. Training runs for frontier models consume enormous quantities of compute, and the gap between a successful training run and a stalled one is often measured in hardware availability.
Capital will also be needed for inference. Every additional user of the company's API or consumer app translates into sustained demand for processing capacity, and that demand is recurring rather than one-off. Companies that priced aggressively during a period of abundant capacity have found themselves raising prices once utilisation rose.
Recruitment is the third claim on the funds. Compensation for senior model researchers has risen sharply in China as well as in the United States, and retaining a team capable of competing at the frontier requires both cash and equity that new investors must be willing to underwrite.
DeepSeek has not commented publicly on the reported terms of the round, the identity of its investors, or the timing of any price changes to its API.
Source:Silicon UK News