Chinese AI Funding: DeepSeek, Moonshot and Manus Raise Billions

Three Chinese AI deals in one week

Within a few days in early October, three of China's best-known AI companies took on large new rounds of funding. The three companies run very different businesses, but the investors behind them overlap heavily. DeepSeek, the lab whose cheap open models upset Silicon Valley in early 2025, is reportedly close to raising at least 80 billion yuan, about $12 billion, which is well above its original target.22 People familiar with the talks say the round could reach 100 billion yuan ($14.9 billion), and that DeepSeek wants a valuation of about 500 billion yuan, or roughly $75 billion.21 Moonshot AI, which makes the Kimi models, closed its own round over the summer. It raised $3.5 billion at a $35 billion valuation after setting out to raise only $1 billion to $2 billion.11 Then, on October 8, Manus's parent company Butterfly Effect announced more than $500 million in new funding. It was the company's first raise since Beijing forced Meta to undo its roughly $2 billion acquisition.1

Saying "Chinese AI is back" is accurate, but it skips the more useful point. Each company makes money differently, and each founder has handled capital, control and the state in a different way. Taken together, the three deals show what Chinese AI businesses look like in 2026.

DeepSeek: a research lab that took outside money on its own terms

For years DeepSeek didn't take outside investors. Founder Liang Wenfeng could fund the lab through his quant hedge fund, High-Flyer.23 That changed in 2026. Reuters reported that the first external round brought in about 50 billion yuan from investors including Tencent and CATL, and that Liang personally put in 20 billion yuan.23 The reason given was compute. The industry has moved from cheap chatbots toward AI agents, which need far more computing power.23

The terms show how Liang approached the raise. According to The Information, as reported by Forbes, most investors in the first round didn't buy shares in DeepSeek directly. Their money went into a limited partnership controlled by Liang, which came with a five-year lock-up and no voting rights.25 The one exception was China's National Artificial Intelligence Industry Investment Fund. It invested directly, got voting rights and has no lock-up.25 The current round follows a similar pattern of selective access. DeepSeek is reportedly turning away private funds raised from individual investors and limiting the round mostly to government and corporate money.21

The business underneath is growing quickly but still loses money. DeepSeek's revenue was reported to be running at close to $500 million a year by mid-July.24 The Information reported that revenue has grown about tenfold since 2025, while the company lost about 715 million yuan between January and July.30 Liang told investors privately in July that reaching artificial general intelligence matters more to him than maximizing profit.21 Investors seem to accept that, at least for now. DeepSeek has hired CITIC Securities to prepare a possible listing on Shanghai's STAR Market, and Bloomberg says it is aiming for an IPO in early 2027.21

The reporting is less consistent on price and timing. Reuters put the first round's post-money valuation at $52 billion to $59 billion.23 A filing tied to the same round implied about 350 billion yuan.21 The second round was supposed to close by the end of August but ran late, partly because some investors pushed back on price.21 Talks were also paused briefly after Liang was reported to have told investors that Chinese AI still trailed far behind the U.S. and depended heavily on Nvidia chips.24 Most likely, reported demand is real but sensitive to price. Investors want exposure to China's national AI champion, though not at any valuation.

Moonshot: open models that bring in revenue

Moonshot shows the quickest route from open models to revenue. Its open-weight Kimi models drive both paid subscriptions and API usage.13 TechCrunch reported in May that Kimi K2.6 was the second-most-used model on the distribution platform OpenRouter.13 Developer adoption reaches into U.S. companies too. Kimi K3 has served as the base for American products, including Cognition's Devin, and Moonshot's use by DoorDash and Cursor maker Anysphere has drawn criticism from a U.S. House committee.12

The valuation rose fast. Moonshot was valued at about $4.3 billion at the end of 2025 and $10 billion early in 2026.13 A Meituan-led round in May then lifted it to $20 billion.13 Two months later, the July round put it at $35 billion.11 Accounts of its revenue differ in timing. Bloomberg said annual recurring revenue (ARR) passed $200 million in April.20 Other coverage has ARR rising from $100 million in March to more than $300 million by June.15 Either way, the price is steep. At $35 billion against about $300 million in ARR, Moonshot is valued at more than 100 times its recurring revenue.15

Moonshot is also raising quickly ahead of a listing. It is approaching investors about a final round at a $50 billion pre-money valuation before a possible Hong Kong IPO.11 It has sent investors an IPO proposal and is preparing to unwind its offshore red-chip structure, under which a Chinese company is held through an overseas entity.18 The strategy seems clear. Moonshot is raising at high valuations while it has momentum from its latest model, then plans to list before that momentum fades. The state AI fund that backs DeepSeek co-led this round.11

Manus: building apps on top of models, and caught between Washington and Beijing

Manus is the application-layer company of the three. It doesn't build foundation models. It sells general-purpose agents that carry out multi-step tasks.10 It launched in China in early 2025 and moved to Singapore after Benchmark invested.1 Meta then agreed to buy it, and China's National Development and Reform Commission blocked the deal by prohibiting foreign investment in the project.1 Its original investors eventually bought back Meta's shares.10

The new round was co-led by Boyu Capital and IDG Capital, with Tencent, HSG and ZhenFund also investing.1 Butterfly Effect didn't disclose a valuation. Bloomberg had reported the round would put it at $4 billion, double what Meta agreed to pay.1 Manus also has real revenue. The Information reported that its annualized revenue run rate reached $400 million to $500 million in June, up from about $100 million in December, though those figures are not audited sales.10

The product approach here matters for startups. The obvious risk for any app built on someone else's models is that better models make it redundant. Manus is responding with its own infrastructure. Manus 2.0 runs on an in-house execution system called Cascade. Manus has also launched Cue, a personal-agent app in which each agent gets its own email address, phone number and mobile wallet.1 SiliconANGLE reports the new framework cuts token use by 23.2% on some tasks and speeds up processing by more than 28%.2 Investor commentary points the same way: investors are paying for capabilities that are hard to copy, such as execution infrastructure, distribution and customer relationships, not simply access to models.5

Some details in the coverage conflict. Reports disagree on whether Manus resumed independent operations in August or September.38 One outlet also named a CEO who differs from the founder, Xiao Hong, identified elsewhere.46 Analysts largely agree on what comes next. Han Lin of The Asia Group says Manus now needs to prove scale, profitability and regulatory alignment.1

The common thread: state money and strategic investors

Across all three deals, the money comes from a small group of Chinese backers. Tencent appears in DeepSeek's and Manus's rounds and is an existing Moonshot investor.22112 CATL and Geely Auto have committed to DeepSeek's current round.21 Meituan's venture arm led Moonshot's May round.13 The state AI fund holds positions in both DeepSeek and Moonshot.11

That has consequences for strategy. Reuters notes that geopolitics largely limits DeepSeek's fundraising to China, and that export controls shape its hardware decisions.23 That shows up in the product work. DeepSeek recently partnered with Huawei to build programming tools for Huawei's Ascend chips.26 The Manus case makes the limits even clearer. Selling a Chinese-founded AI company to a U.S. buyer is now effectively blocked, so the realistic exits are Chinese capital and listings in Hong Kong or Shanghai. One analysis says Manus's investors are betting on later rounds or a regional IPO, not a U.S. listing.4

The overall size still needs perspective. Even a $15 billion DeepSeek round is small next to the $122 billion OpenAI raised in March and Anthropic's $65 billion round in May.23 Chinese companies can't spend as much as U.S. labs, so they compete on cost and openness and depend more on state-linked investors.

Lessons for founders

Three lessons come out of these deals. First, open models can produce revenue. Moonshot and DeepSeek both turned free open weights into growing API and subscription income. Second, founders with leverage can raise money without giving up control. DeepSeek's partnership structure is an extreme example, but it worked because investors badly wanted in. Third, app-layer companies need technology the model makers don't control. Manus is betting on its own execution system and new consumer products such as Cue.

The main risk is valuation. Investors are paying more than 100 times revenue for these companies, and DeepSeek is still losing money. The planned IPOs in Hong Kong and Shanghai will show whether public investors will pay the same prices. These rounds show that Chinese AI can attract very large sums. Whether the businesses justify those valuations is still unproven.