DeepSeek has restarted its second external fundraising round, with the Hangzhou-based artificial intelligence developer seeking close to $8bn in fresh capital — roughly £6bn — some three months after abruptly pausing the process when details of its conversations with prospective backers were leaked.

The company is said to be targeting a valuation of approximately 500bn yuan, or about $74bn, according to a Chinese business magazine that cited unnamed sources familiar with the discussions. That figure would represent a steep step up from the roughly 350bn yuan valuation attached to DeepSeek when its first round closed earlier in the summer, and would place the start-up among the most valuable privately held AI companies anywhere in the world.

At least one institutional investor is reportedly in advanced talks to participate. Monolith Management, a venture firm that has backed Beijing-based Moonshot AI, the developer behind the Kimi family of models, is said to be negotiating a contribution, according to financial media reports. Neither DeepSeek nor Monolith has commented publicly on the talks, and neither the size nor the structure of any individual commitment has been disclosed.

A fundraising process restarted

The second round was suspended in July, shortly after remarks the company made to investors circulated more widely than intended. The pause was widely read as a signal that DeepSeek wanted to reset the terms and the narrative around the raise before re-engaging with backers. Restarting the process at a higher target suggests the company's leadership concluded that demand for a stake remained strong despite the interruption.

Details of the structure are thin. It is not clear whether the round will consist entirely of primary capital, whether early backers will be permitted to sell down portions of their holdings, or how much of the total has already been informally committed. Chinese AI financing rounds of this size are frequently assembled from a mixture of state-linked funds, private venture capital, strategic investors from the hardware and cloud sectors, and, in some cases, offshore vehicles.

A valuation that has climbed rapidly

The trajectory of DeepSeek's valuation illustrates how quickly sentiment has shifted around Chinese frontier-model developers. The first round, concluded earlier in the summer at roughly 350bn yuan, was itself regarded as an extraordinary mark for a company that was founded only in 2023 and that had, until early 2025, been viewed mainly as a research curiosity rather than a commercial challenger.

If the new round clears at 500bn yuan, it would imply an increase of more than 40% in a matter of months — a pace of appreciation that would be unusual even for the most hyped Silicon Valley laboratories. Some of that uplift likely reflects the broader re-rating of Chinese AI assets over the past year, as domestic model developers secured larger rounds and as investors grew more confident that open-weight models could sustain viable businesses.

The Inner Mongolia data centre

A significant portion of the proceeds is expected to be directed towards a large data centre project in Inner Mongolia, the northern Chinese region that has become a magnet for compute-intensive infrastructure because of its cheap electricity, cool climate and proximity to renewable generation. The project forms part of a broader effort by DeepSeek to secure access to roughly 1 gigawatt of computing power, a target that may be met through a combination of self-built facilities and leased capacity from third-party operators.

One gigawatt is a difficult figure to grasp in the abstract. For context, a gigawatt of IT load represents a cluster of tens of thousands of high-end accelerators and the power, cooling and networking to keep them running — the sort of campus that hyperscale cloud providers typically build over multiple years and multiple phases. Securing that quantity of capacity would represent a step change in DeepSeek's ability to train and serve models, moving it from a company that has made a virtue of doing more with less into one that commands resources comparable to major technology groups.

Leasing capacity from third parties would allow DeepSeek to expand without carrying the full capital burden of construction, though it typically comes at a higher marginal cost and with less control over scheduling and utilisation. Building in Inner Mongolia offers lower operating costs but slower timelines and greater exposure to power allocation decisions made at a regional level.

A coming price increase

Alongside the infrastructure plans, DeepSeek told users last Thursday that it intended a substantial rise in prices, without specifying the scale of the increase or the date on which new pricing would take effect. The notice marked a notable departure for a company that built its reputation on undercutting rivals by wide margins.

DeepSeek currently charges $0.14 per million input tokens and $0.28 per million output tokens. By comparison, Moonshot's latest Kimi K3 is priced at $3 and $15 respectively, while Anthropic's latest Fable 5 sits at $10 and $50. Even a large percentage increase would leave DeepSeek at the low end of the market, but the direction of travel is significant: the company appears to be signalling that the era of almost give-away inference is ending.

The most plausible reading is that DeepSeek is trying to align price with the true cost of serving a rapidly expanding user base. Cheap tokens attract volume, and volume consumes accelerators. With a gigawatt of new capacity to pay for, and with investors expecting a path to sustainable economics, higher prices on at least some tiers look increasingly necessary.

The economics of cheap inference

DeepSeek's rise was built on an uncomfortable proposition for incumbents: that frontier-level performance could be achieved at a fraction of the assumed cost. Its models were trained with novel efficiency techniques — heavily optimised mixture-of-experts architectures, aggressive use of lower-precision arithmetic, and infrastructure engineering that squeezed more throughput from constrained hardware — and released with open weights that anyone could download and run.

That combination upended assumptions in Western capitals, where hundreds of billions of dollars were being committed to data centre build-outs on the premise that scale of compute was the decisive advantage. The release of DeepSeek's reasoning model in January 2025 triggered a sharp sell-off in semiconductor stocks and a prolonged argument among researchers and investors about how much of the prevailing capital expenditure was genuinely necessary.

Efficiency, however, is not the same as free. Serving hundreds of millions of requests places different demands on infrastructure than training a single large model, and inference costs scale with usage rather than with the size of the training run. As DeepSeek's consumer application climbed download charts in early 2025, the company repeatedly had to ration capacity, pause new sign-ups and slow response times. The gigawatt plan should be understood in that light: it is an attempt to convert research prestige into reliable, always-on service.

From hedge fund spin-out to global challenger

DeepSeek was founded in 2023 by Liang Wenfeng, a quantitative investor who built his fortune running the High-Flyer hedge fund and who funded the AI venture largely from his own resources in its early phase. That unusual structure gave the company something most start-ups lack: patience. It could hire researchers, buy compute and publish results without the constant pressure of a venture timetable.

Its first models drew attention for strong benchmark scores relative to their size. A mid-2024 release widely credited with igniting a domestic price war in China established the pattern that would define the company: competitive or superior performance at prices that forced rivals to cut their own rates. Subsequent generations cemented its reputation, and the release of open-weight reasoning models turned it into a global reference point almost overnight.

DeepSeek is also a product of its geography. Hangzhou has become one of China's densest concentrations of AI and robotics talent, home to a cluster of companies spanning e-commerce, embodied intelligence and gaming. That ecosystem supplies both engineers and suppliers, and it gives local firms an advantage in recruiting against rivals in Beijing and Shenzhen.

Compute constraints and export controls

The single largest uncertainty hanging over any Chinese frontier-model developer is access to advanced semiconductors. Successive rounds of export restrictions have limited the availability of the most capable accelerators, pushing companies towards domestic alternatives, older stockpiles and heavy engineering workarounds. DeepSeek's efficiency gains were partly a response to those constraints, born of necessity rather than pure research ambition.

That makes the Inner Mongolia project strategically as well as financially important. Domestic data centre capacity, paired with domestically produced chips, would reduce exposure to supply-chain shocks and to the whims of overseas policy. It would also, however, tie the company's fortunes more closely to the performance and yields of China's own semiconductor industry, which remains some distance behind the leading edge.

What the round would unlock

The proceeds from a completed round would give DeepSeek room to accelerate construction, lock in power agreements and pre-purchase equipment before further tightening of supply. It would also provide a war chest for hiring, as competition for experienced researchers intensifies across Chinese AI laboratories that have themselves raised substantial sums.

Should the round close near the reported terms, DeepSeek would join a small group of privately held AI companies valued in the tens of billions of dollars, alongside the largest American laboratories and a handful of Chinese peers. The reported interest from Monolith suggests that investors already exposed to rival Chinese model developers see value in holding positions across several of them rather than betting on a single winner.

For now, the raise remains a set of negotiations rather than a completed transaction, with the final valuation, the roster of participants and the allocation of proceeds all subject to change. What is already visible is the strategy: a company that once distinguished itself by spending little is preparing to spend a great deal, on power, on land and on silicon, while simultaneously asking its users to pay more for the tokens it produces.

Source: Silicon UK News