How China funds its AI labs

Cumulative disclosed AI capital across 22 China-tagged labs sits at $9.7 billion, roughly 5% of what OpenAI alone has raised. The number is real and it is also misleading. The actual Chinese AI funding system runs on four overlapping sources and a cap-table evolution that ran from BAT-led private rounds, through state-backed financing, to two Hong Kong IPOs listed within 24 hours of each other.

The Nextomoro Atlas tracks 22 China-category labs. Their combined disclosed funding is $9.7 billion. The companion essay on the 2024-2026 funding vintages spent most of its time in a US-centric capital market where individual rounds clear $20 billion. By that frame, China looks like a rounding error.

The frame is wrong. China's AI funding system was never going to look like the US system, and reading it through US conventions misses what's actually happening. The disclosed $9.7 billion is the visible tip of a deeper and more complex capital base that includes BAT internal AI investment (largely unreported), state-affiliated capital (newly named in 2025-2026), and a Hong Kong public market that quietly opened a frontier-AI listing window in early January 2026. This essay walks through the four capital sources, the year-by-year evolution from venture rounds to state-backed financing, the role of internal capital at Alibaba, Tencent, Baidu, ByteDance, and Huawei, and the DeepSeek anomaly that demonstrates how little outside money the Chinese frontier actually requires.

The four sources

Chinese AI labs raise from four overlapping pools. They do not look alike.

Independent venture capital is the smallest of the four and has been thinning out for years. The names that mattered in the pre-AI cycle (IDG Capital, Hongshan, Source Code Capital, 5Y Capital, Qiming Venture Partners, Hillhouse) still write checks but rarely lead the largest rounds. IDG Capital led the Moonshot Series F in March 2026, the most prominent example in the dataset of a China private round still being led by a conventional venture firm. Most of the rest of the recent rounds are either led by a strategic, led by a state vehicle, or have no disclosed lead at all.

BAT strategic capital, meaning Baidu, Alibaba, Tencent (and increasingly Meituan, Xiaomi, and ByteDance), is the most active source of disclosed lead capital across the dataset. Alibaba led the Moonshot Series A in February 2024 ($1 billion), the MiniMax Series B in March 2024 ($600 million), and the Z.AI December 2024 round ($411 million). Alibaba Cloud led 01.AI's Series A in November 2023 ($200 million). Tencent led MiniMax's Series A in June 2023 ($250 million), the Moonshot Series B in August 2024 (co-led with Gaorong Capital, $300 million), and StepFun's Series A in 2023. The same investors keep appearing on the same cap tables, which means the Chinese frontier has effectively three institutional buyers writing the largest checks.

State-affiliated capital is the source that arrived most recently and is changing the picture fastest. The dataset records two named state-led China rounds, both in the second half of 2025 and early 2026, both led by Shanghai State-owned Capital Investment: a $300 million round into MiniMax in July 2025 (described in the dossier as "a $300 million round led by Shanghai State-owned Capital Investment" with no specific series letter) and the StepFun Series B+ in January 2026 at approximately RMB 5 billion ($717 million USD). Two rounds may sound modest but the trajectory matters. The state-led check did not exist in the disclosed China data before mid-2025.

The Hong Kong public market is the fourth source and the one whose mechanics shifted most recently. In January 2026, two China-frontier labs listed on the Hong Kong Stock Exchange within 24 hours of each other: Z.AI on January 8 (HKEX 2513, raising approximately $558 million at HK$116.20 per share) and MiniMax on January 9 (raising approximately $619 million at a roughly $11.5 billion debut market capitalisation). The mainland-to-Hong Kong listing path has existed for years (Kuaishou listed in February 2021 at $159 billion, SenseTime in December 2021 at $16.5 billion) but those were pre-frontier-AI listings of internet and computer-vision companies. The 2026 listings are the first frontier-LLM IPOs in the dataset, and they happened in Hong Kong rather than New York for reasons that are no longer subtle.

These four sources are not exclusive. A single Chinese lab will typically raise from a mix: Alibaba and Tencent in early rounds, IDG or Hongshan as a co-lead in later venture rounds, state-affiliated capital in the most recent round, and an HKEX listing as exit. The sequencing matters and tells you what kind of company each lab has become.

The 2024 BAT consolidation

For the 18 months from late 2023 through early 2025, almost every named lead investor on a major Chinese frontier round was Alibaba or Tencent. The pattern is too clean to be accidental.

Alibaba led the Moonshot Series A in February 2024 at $2.5 billion post-money. Alibaba led the MiniMax Series B in March 2024 at $2.5 billion. Alibaba Cloud had already led 01.AI's Series A in November 2023 at $1 billion. Alibaba led the Z.AI follow-on in December 2024 at $3 billion (co-led with Tencent). Tencent led the MiniMax Series A in June 2023. Tencent led the Moonshot Series B in August 2024 at $3.3 billion (co-led with Gaorong). Tencent anchored the StepFun Series A in 2023. The Moonshot Series C in December 2024 was described in the dossier as a follow-on round oversubscribed by Alibaba, Tencent, and Wang Huiwen (the Meituan co-founder).

Two companies, three founders' pockets, six of the seven major China-frontier rounds in 18 months. Alibaba alone leads roughly $1.7 billion of disclosed Chinese AI funding across the dataset. Tencent leads roughly $850 million. Together the two account for roughly half of all disclosed China-frontier capital in 2023-2024.

The structural read is simple. The Chinese venture base did not have the standalone capital pool to underwrite frontier-LLM training runs, but the BAT incumbents did, and they treated the new generation of independent labs as both competitive insurance and infrastructure suppliers. Moonshot, MiniMax, Z.AI, and 01.AI all built models that BAT could deploy in their own products, while also serving the broader Chinese developer ecosystem. The cap tables priced that strategic relationship explicitly.

A 2024 Chinese frontier round did not look like a 2024 US frontier round. The US version was a Sequoia or Andreessen syndicate writing $50-200 million into a venture-capital cap table. The China version was Alibaba or Tencent writing $300 million-$1 billion into a strategic-capital cap table whose other investors were typically Chinese growth funds with limited follow-on capacity. The labs raised less per round but raised from sources whose strategic horizon was longer.

The 2025 state-capital arrival

Mid-2025 marks a clean break in the data.

Before July 2025, no disclosed Chinese AI round in the dataset is led by a named state vehicle. After July 2025, the largest two non-IPO rounds in China are both led by Shanghai State-owned Capital Investment (SOIC). The MiniMax round in July 2025 is the first instance. The StepFun Series B+ in January 2026, at approximately $717 million, is the second and larger. Both founders had previously raised from BAT and conventional Chinese venture firms. Both moved up the capital stack to a state vehicle for their next round.

The pattern is recent enough that it could be coincidence. Read alongside the broader policy backdrop, it is more likely a signal. China's "Next Generation AI Development Plan" dates to 2017 and has produced municipal AI plans in Beijing, Shanghai, Hangzhou, and Shenzhen. The Atlas tracks $150 billion in disclosed Chinese government AI funding across these initiatives, which is roughly 15 times the cumulative private disclosed China AI funding. The state has been spending this money primarily on infrastructure (compute, fab capacity, data centres), education (university AI programs), and research labs (BAAI, Shanghai AI Laboratory, Pengcheng Laboratory). What changed in 2025 is the start of state capital flowing directly into commercial frontier-LLM companies as named lead investors.

Shanghai SOIC operates as a municipal asset manager controlled by the Shanghai government, distinct in structure and mandate from a sovereign wealth fund of the GIC or MGX kind. Its mandate includes "strategic emerging industries" and AI is now explicitly inside that mandate. Two visible rounds is not a trend, but the targeting is striking: MiniMax (one of the four major Chinese frontier-LLM independents) and StepFun (the most prominent Shanghai-based frontier lab). If the same pattern shows up two more times in 2026, the read becomes clear: state capital is now an explicit participant in the Chinese frontier cap table, and the BAT-only era is over.

The Hong Kong window

In early January 2026, two Chinese AI labs IPO'd on the Hong Kong Stock Exchange within 24 hours of each other. Z.AI listed on January 8 at $558 million raised. MiniMax listed on January 9 at $619 million raised at an $11.5 billion debut market capitalisation. Both transactions priced into the HKEX rather than the NYSE or Nasdaq, both raised mid-nine-figure amounts at single-digit-billion valuations, and both gave existing investors (Alibaba, Tencent, prior venture rounds) their first liquidity event since the labs were founded.

The Hong Kong listing path is mechanically similar to what Kuaishou and SenseTime did in 2021, but the strategic context has changed completely. The 2021 IPO wave happened when US-Chinese capital flows were still substantively open and a Chinese tech company could plausibly choose between Hong Kong and New York. The 2026 IPOs happened in a market where US listing for a Chinese AI company has become functionally impossible: PCAOB audit requirements, CFIUS review, and explicit regulatory pressure on US institutional investors holding Chinese AI equities have closed the Nasdaq path. Hong Kong is the only viable public market that gives the lab access to international capital while maintaining the company's mainland operating structure.

The valuations also tell the story. SenseTime listed at $16.5 billion in December 2021, a full computer-vision unicorn at IPO and (at the time) the largest pure-AI public listing globally. Kuaishou listed at $159 billion in February 2021, but Kuaishou is primarily a short-video platform whose AI work is internal. The 2026 IPOs of Z.AI ($558 million raised at a substantially smaller market cap than SenseTime's debut) and MiniMax ($11.5 billion debut market cap) are the first frontier-LLM IPOs in the dataset, and they priced at multiples that look modest by US AI standards. A US-frontier lab at MiniMax's stage of development would not be IPO'ing at $11.5 billion.

The Hong Kong listing is doing two things simultaneously. It provides exit liquidity for the BAT and venture investors who put money in during 2023-2024. It also opens a public-market access channel for international institutional investors (sovereign wealth funds in Singapore and the Middle East, European pension funds, Hong Kong-based hedge funds) who want frontier-AI exposure but cannot or will not buy the US frontier names at current valuations. A Z.AI or MiniMax position offers Chinese frontier AI exposure at what looks, by US comparison, like a value multiple.

The window may stay open. If it does, expect Moonshot, StepFun, and possibly DeepSeek to follow the same path within 12-18 months. The infrastructure is now in place.

The subsidiary AI labs

The disclosed $9.7 billion in China-tagged AI capital is the wrong number to focus on, because most of the actual Chinese AI investment runs through the internal balance sheets of BAT and other public-company subsidiaries.

The Atlas records seven major China labs as subsidiaries of public companies: Alibaba Qwen / DAMO (Alibaba), Tencent Hunyuan (Tencent), Baidu's AI division (Baidu), ByteDance Seed (ByteDance), Huawei Noah's Ark Lab (Huawei), Xiaomi AI Lab (Xiaomi), and Ant Group AI (Alibaba affiliate). Each one shows $0 in disclosed funding because each one is funded internally by a parent company whose AI investment is buried inside research-and-development opex on a public 10-K-equivalent filing. The disclosed-round dataset does not capture this.

Public reporting suggests the actual numbers are significant. Alibaba committed RMB 380 billion (approximately $53 billion USD) to AI and cloud infrastructure over a three-year period announced in February 2025. Tencent's 2024 capex jumped substantially, with much of the increase attributed to AI compute. ByteDance was reported to be spending more on AI infrastructure than any single Chinese company in 2024. Baidu's AI investment is harder to disaggregate from its broader R&D, but is known to be in the multi-billion-dollar annual range. Huawei's HiSilicon and Ascend chip programs, plus the Noah's Ark research lab, represent substantial multi-year AI investment that runs through Huawei's general balance sheet.

On even conservative assumptions about the share of these announced figures actually deployed against AI training and inference (with the remainder going to general cloud capex), the total internal Chinese AI investment for 2024-2026 sits comfortably above $100 billion. That number is more than ten times the disclosed independent-lab funding tracked in the Atlas. The Chinese AI capital base is large. It is also mostly invisible in venture-style datasets, because most of it sits as internal capital allocation at public companies and never appears as an external capital raise.

The implication for reading the disclosed data is straightforward. The independent China-frontier labs (Moonshot, MiniMax, Z.AI, StepFun, 01.AI, DeepSeek) are the visible tip of a much larger capital deployment. They matter strategically, especially for cases where the technology has to be portable across the Chinese ecosystem and cannot live inside a single hyperscaler. But the headline question of "how much capital is being deployed against AI in China" is answered inside the line items of Alibaba, Tencent, ByteDance, and Huawei's annual reports, far more than inside the disclosed-rounds dataset this essay relies on.

The DeepSeek anomaly

DeepSeek is the single most important lab in the dataset for thinking about Chinese AI funding mechanics, because it is the lab that demonstrates how little disclosed external capital a Chinese frontier company actually requires.

DeepSeek was founded in May 2023 by Liang Wenfeng, who also founded and runs the Hangzhou-based quantitative hedge fund High-Flyer. From founding through April 2026, DeepSeek raised exactly zero disclosed external rounds. The entire training operation, including the V2 and V3 model releases, was funded by High-Flyer's internal capital. The April 2026 round (described in the dossier as "in process" and not yet closed) is reported at a $10 billion-plus valuation and represents DeepSeek's first external fundraising. The lead investors are not yet disclosed.

The DeepSeek case undermines several assumptions that the US-frontier lens imposes on Chinese AI. The first is that frontier-LLM training requires multi-billion-dollar venture rounds. DeepSeek's V3 was reported to have cost approximately $5.6 million in compute to train, against a contemporaneous OpenAI training run that reportedly cost orders of magnitude more. The second is that a frontier lab requires institutional venture investors. DeepSeek had none until April 2026 and shipped competitive frontier models without them. The third is that Chinese frontier labs are structurally dependent on BAT distribution and BAT capital. DeepSeek had neither and shipped models that gained immediate global traction in early 2025, including becoming the most-downloaded AI app in multiple markets in late January 2025.

The DeepSeek result has changed how the rest of the Chinese frontier is being underwritten. Investors looking at Moonshot's $18 billion Series F valuation or MiniMax's $11.5 billion IPO debut are implicitly assuming those companies can compress their cost structure toward the DeepSeek baseline. The same investors looking at OpenAI's $852 billion valuation are assuming scale and revenue trajectory justify the multiple regardless of cost compression. Two different theses on two different cost regimes, in two different capital markets.

The efficiency premium

Across the China-tagged labs, the efficiency-to-capital ratio is the most striking feature when compared to the US frontier.

The Atlas records DeepSeek, Moonshot, Z.AI, MiniMax, StepFun, and Alibaba Qwen all in the active competitive set on Artificial Analysis Intelligence Index and LMArena rankings. These labs collectively raised less than $10 billion in disclosed external capital. The US labs at the comparable benchmark tier (OpenAI, Anthropic, xAI, Google DeepMind, Meta) collectively raised more than $300 billion. The ratio is roughly 30-to-1 in capital deployed for roughly comparable benchmark performance.

The efficiency premium has multiple sources. Lower talent costs (a senior Chinese AI researcher in Hangzhou earns a fraction of a Bay Area equivalent). Lower compute costs (domestic GPU procurement, electricity prices, government-subsidised data centre capacity). Architectural choices that prioritise inference cost (Moonshot's long-context Mixture-of-Experts approach, DeepSeek's V3 MoE design, MiniMax's linear-attention work). Constraint as a feature (US export controls on top-tier GPUs forced Chinese labs to optimise for older H800 / H20 silicon and earlier-generation domestic chips, which produced engineering practices that travel back even when better hardware becomes available).

The efficiency premium has limits. There are categories of frontier work (largest-scale pre-training runs, multimodal models with extensive video data, agent training at scale) where the US labs have a clear capital-and-compute advantage. The Chinese labs' efficiency gains compress the gap but do not close it. What the gains do mean is that the disclosed-funding gap of 30-to-1 substantially overstates the capability gap, and that the strategic question is no longer "can China close the AI funding gap" but "does China need to close it to maintain competitive parity."

Two notes on subsidiary AI labs and academic research

A complete picture of Chinese AI funding has to include two categories that the disclosed-rounds data underweights.

The state research labs (BAAI in Beijing, Shanghai AI Laboratory, Pengcheng Laboratory in Shenzhen, Tsinghua's KEG group) operate on grant cycles funded by central government, municipal government, and university budgets. The Atlas records them with $0 in disclosed funding because their financing does not appear as venture rounds. Their actual annual operating budgets are estimated to total in the low single-digit billions of dollars. They produce a meaningful share of Chinese open-source AI research output, including the BGE embedding family (BAAI), InternLM (Shanghai AI Lab), and a long tail of academic models. Pengcheng Laboratory's PCL-PanGu efforts and Tsinghua's GLM family of models came out of this academic-state-research ecosystem.

The smaller specialty labs (OpenBMB, ModelBest, RWKV, Megvii's research group) operate on a mix of academic grants, BAT strategic investment, and small private rounds that often go undisclosed. These labs are influential in specific technical communities (RWKV's recurrent attention work, OpenBMB's MiniCPM small-models work) but do not appear in the disclosed-funding dataset at meaningful scale. Their impact on the Chinese AI ecosystem is real and is mostly invisible in capital data.

Both categories underline the same point: Chinese AI capital does not flow primarily through venture rounds. It flows through state research budgets, public-company internal allocation, BAT strategic investment, and (most recently) state-vehicle direct equity. The venture pool is a small fraction of the total.

Why this matters

The structural read on Chinese AI funding is that it is a system designed to be insulated from US capital markets. Each of the four sources operates inside a capital pool that is mostly or entirely Chinese: BAT and other domestic public companies, mainland state vehicles, mainland-to-Hong Kong listing, and a domestic venture base whose largest funds are now structurally constrained from US LP capital. The 2026 cap tables tell you the decoupling is functionally complete. There is no US lead investor in any 2025 or 2026 Chinese AI round in the dataset. That fact is unusual in any historical comparison: as recently as 2021, US funds (Sequoia China legacy, Tiger Global, Coatue) were the largest single category of late-stage Chinese tech investor.

The decoupling has been built deliberately on both sides. From the US side, regulatory pressure (Treasury sanctions, CFIUS reviews, executive-order restrictions on outbound investment to Chinese AI companies) has made US institutional capital functionally unavailable to Chinese AI labs. From the Chinese side, the listing reorientation toward Hong Kong, the introduction of state-vehicle direct equity, and the BAT consolidation have built a capital stack that does not require US participation.

The result is that the Chinese AI funding system is now an alternative architecture, not a delayed copy of the US system. It will produce different companies. Its labs will be cheaper to operate, more strategically aligned with their state and corporate investors, more dependent on Chinese-language data and use cases, and more likely to exit through Hong Kong or domestic Chinese channels than through global capital markets. None of this is hypothetical: the 2024-2026 dataset is the first cohort of frontier labs raised entirely inside this alternative system, and it is shipping competitive products.

What to watch

Five concrete signals over the next twelve months that would change the picture.

1. Whether state-vehicle leads become the modal China AI round. Two Shanghai SOIC-led rounds in twelve months is a pattern but not a regime change. If three or four more major China rounds in 2026 are led by named state vehicles (Beijing Equity Investment, Shenzhen Capital Group, China Internet Investment Fund, the National Big Fund's AI-extension vehicles), the disclosed-China cap table will look fundamentally different by year-end.

2. The next Hong Kong frontier-AI IPO. The Z.AI and MiniMax listings tested the window. Moonshot is the most plausible next candidate (the Series F at $18 billion already prices it for a public-market debut). StepFun could file faster given its state-backed cap table. A third HKEX frontier-AI listing within twelve months locks the window open. A failed listing or a quiet pulling of an S-1 closes it.

3. The DeepSeek April 2026 round closing. DeepSeek's first external fundraise will reveal what valuation a self-funded frontier lab can command, and which capital pool ends up underwriting it. If it closes at $10 billion-plus with a state-vehicle or BAT lead, the round will validate the efficiency-premium thesis in capital terms. If it closes at a substantially higher number with international anchor investors (sovereign wealth funds in the Gulf or Singapore), it will signal that DeepSeek's strategic value is being read globally.

4. BAT internal AI capex disclosure. The single most consequential change in the disclosed-funding picture would be Alibaba, Tencent, Baidu, and ByteDance breaking out their AI capex from general cloud and R&D in their financial reports. The pressure to do so is increasing as institutional investors in those companies want to underwrite the AI optionality independently of the broader business. If even one of the four starts disclosing a separated AI investment line, the disclosed-China AI capital base will jump by an order of magnitude and the BAT-subsidiary labs will move into the visible-funding tier.

5. The next generation of Chinese frontier labs. The 2024-2026 cohort (Moonshot, MiniMax, Z.AI, StepFun, 01.AI, DeepSeek) is the visible China frontier. The 2027-2028 cohort has not yet been founded and will reflect what the new capital regime selects for. If the next wave of Chinese frontier labs is founded by ex-BAT researchers funded by state-vehicle seed capital, the system will have produced a new generation of state-aligned commercial labs that look quite different from the 2024 Tencent-Alibaba-backed cohort. If the next wave looks more like DeepSeek (founder-funded, strategically independent, surprising in its capital efficiency), the system will have produced something the central planners did not intend and may not control.

The honest summary

The Chinese AI funding system is structurally distinct from the US system. It produces different kinds of companies, on a different time horizon, exiting through different markets. The disclosed numbers are misleading because they ignore the BAT internal capital base, the state research-lab funding, and the fact that Chinese frontier labs are structurally cheaper to operate than US comparables. The decoupling from US capital is essentially complete: there is no meaningful US LP money in any major 2025 or 2026 Chinese AI round.

The four capital sources work together. BAT strategic capital underwrites the early rounds. State capital is now a named participant in the late-stage rounds. The Hong Kong public market provides the exit. Independent Chinese venture capital fills the gaps. Internal capital at the public-company subsidiaries dwarfs all of it and remains mostly undisclosed.

What this produces is a frontier-AI ecosystem that the US-centric venture lens systematically underrates. The 2024-2026 cohort of Chinese frontier labs raised less than $10 billion in disclosed external capital. They are competing on global benchmarks against US labs that raised more than thirty times that figure. The benchmark gap is real but is far smaller than the funding gap. That divergence, more than any single round size or cap-table detail, is the actual signal in the data.

The next two years will determine whether this configuration is fragile or durable. The fragility case rests on continued US export controls, possible Chinese regulatory tightening on AI, and the ever-present risk that a frontier-LLM training cycle gets meaningfully more expensive (multimodal video pre-training, large-scale agent training) in ways that reward capital scale and punish efficiency optimisation. The durability case rests on the four-source funding architecture being mature, the Hong Kong window being open, the DeepSeek result demonstrating that capital scale is not the only path to frontier capability, and the state's growing willingness to underwrite the gap directly.

Either outcome is consequential. The Chinese AI funding system as it exists in mid-2026 is the first observable instance of a fully decoupled, structurally distinct, frontier-scale AI capital base operating at competitive parity with the US system. Whatever it produces over the next 24 months will reshape the rest of the AI industry as much as the OpenAI-Anthropic capital arms race has reshaped it.

Sources used in this piece:

• The Nextomoro Atlas dataset (rounds.json, labs.json, manifest.json) extracted 2026-04-30, filtered for category="china" across 22 labs and 24 disclosed rounds.
• Hong Kong Stock Exchange filings for Z.AI (HKEX 2513) and MiniMax January 2026 listings.
• Public reporting on DeepSeek's funding history and High-Flyer's role from late-2024 and early-2025 coverage in The Information, Bloomberg, and Chinese-language tech press.
• Alibaba's February 2025 announcement of RMB 380 billion in three-year AI and cloud infrastructure commitment.
• Atlas dossiers for each named lab cited (linked from each lab's Nextomoro page) for round detail and extraction notes.
• The companion essay "Vintages: how AI's funding cycle decoupled from reality" for the 2024-2026 US-centric vintage frame this piece counterposes.

Last updated: April 30, 2026. All disclosed round amounts and valuations are at announced figures, converted to USD at the rate effective at announcement. The state-affiliated capital classification follows the lead-investor field; many Chinese rounds carry no disclosed lead and are not classified to a single source. Send corrections.