Examining China’s AI Financing

Executive summary

Competition between the United States and China in artificial intelligence (AI) is one of the most important topics in global markets today. The expected Anthropic initial public offering will direct new attention to the financing of the US AI buildout, but there has been little focus on how China’s AI firms are financing their own expansion, or how constraints on compute and finance interact in China’s market.

In this report, we detail the financing channels for China’s AI firms and evaluate their scope and sustainability, including a comparative perspective with US firms.

Key takeaways

• The scale of China’s AI buildout is still only around 15-20% of the level of US investments. We expect China’s AI capex to double this year to 932 billion yuan ($139 billion) and to top 1.2 trillion yuan ($193 billion) in 2027.
• Beyond shortages of chips and access to compute, financing constraints also limit new AI capex in China. China’s AI firms face similar cash flow problems to their US counterparts, with revenues failing to cover aggressive capex plans by wide margins. Hyperscalers’ cash-generating businesses still depend upon slowing consumption from Chinese households.
• China’s AI firms rely heavily on equity financing and bank loans and have not yet turned to bond financing, unlike their US counterparts.
• However, revenue and profitability metrics for Chinese AI firms still appear far behind US peers. Hence, continued expansion likely depends upon equity market conditions. Even more aggressive state-led investments aligned with China’s industrial policy priorities are likely to remain focused on chips rather than on frontier labs.

China’s AI financing

The United States and China are by far the largest players in the global AI industry. Market attention has focused on the potential sustainability of the US data center buildout, but China’s own AI financing structures have remained under the radar. In this report, we examine capital expenditures from Chinese hyperscalers and AI companies, how they finance their expansion, and current financial conditions at these firms.

We surveyed 13 Chinese listed firms, including hyperscalers, telecom companies, AI labs and independent data center operators. Their estimated AI infrastructure capex is projected to spike by 103% this year to 932 billion yuan ($139 billion). This reflects very strong growth but remains far below the estimated US data center investment of around $800 billion this year.

China’s AI firms also post much lower revenues than their US peers and are experiencing similar shortages of operating cash flow. The three hyperscalers we surveyed—Alibaba, Tencent, and Baidu—saw combined free cash flows turn negative by 16 billion yuan in the first half of 2026.¹ Non-AI business channels can bridge some of the gap, but the slowdown in China’s economy and household consumption reduces profits and cash flows generated through channels such as e-commerce and advertising.

The key question for the future of China’s AI buildout is how far other financing channels can stretch, especially if the equity market cools down. The secondary question is what level of support the state will provide through its own guidance funds, and for how long.

A familiar story: strong capex, negative cash flows

China’s AI infrastructure capex has been limited by chip shortages. In contrast to Alibaba, which can produce its own chips and has greater access to overseas data centers, Tencent’s capex trended lower in 2025 after a short pickup at the end of 2024, due to difficulties in obtaining AI chips (Figure 1) amid conflicting signals in US export controls.

This year, domestic chip supplies from SMIC have risen, and access to NVIDIA H200 chips appears to be easing. In addition to Chinese firms accessing advanced chips through third countries, China’s AI infrastructure capex growth is expected to accelerate in 2026 from 33% growth in 2025. The total capex of Alibaba, Tencent, and Baidu reached 208 billion yuan in H1 2026, exceeding expectations from the beginning of the year. H200 shipments from August will further support H2 capex, though further US-China tensions could close off access to chips and remote access loopholes later in the year.

As detailed in Figure 2, capex guidance from China’s major hyperscalers and telecoms implies 103% y/y growth in AI capex this year to 932 billion yuan ($139 billion) and over 1.2 trillion yuan ($193 billion) in 2027, with 39% growth. ByteDance’s upgraded capex plan is the largest marginal contributor this year.

Spending on AI pressures hyperscalers’ free cash flow from both operating cash flows and capital expenditures. According to Tencent’s disclosures, around 51 billion yuan in AI-related prepayments were included within Q2 operating cash flow, which was similar in scale to Q2 capex of 59 billion yuan. Total free cash flow of Alibaba, Tencent, and Baidu has fallen from 170 billion yuan in 2025 to negative 16 billion yuan in H1 2026 (Figure 3).

The same story is playing out in the US as well, with free cash flows of Microsoft, Amazon, Alphabet, Meta, and Oracle dropping sharply from $191 billion last year to $14 billion in H1 this year. US hyperscalers have chosen to rely more on debt to offset the decline in cash flow, as net cash flows raised from debt of the five US hyperscalers surged from $90 billion in 2025 to $163 billion in H1 2026, mainly through long-term bonds. This has made the bond market and private credit central to the sustainability of the US AI boom.

However, the financing structure of China’s AI sector is quite different from that of the US, with loans and equity playing a larger role than the bond market.

How does China finance its AI spending?

To study how Chinese hyperscalers and other participants in the AI frenzy are financing their AI spending, we surveyed 13 companies representing four different types:

• Hyperscalers: Alibaba, Tencent, and Baidu
• Huawei and telecoms: Huawei, China Mobile, China Telecom, and China Unicom
• Frontier AI labs: Zhipu AI and MiniMax
• Independent data center operators: GDS, VNET, RangeIDC, and Aofei

The latter two groups, frontier AI labs and independent data center operators, are significantly smaller in scale than the hyperscalers and telecom firms, but they are still important. Frontier AI labs are growing at rapid rates and are major consumers of hyperscalers’ AI compute, while Zhipu AI also seeks to build its own 1-GW data center. Individual independent data center operators are much smaller than hyperscalers, but they outnumber hyperscalers and are spread widely across China.

To compare cash flows between them in Figures 4 and 5, we introduced a ratio of cash flow divided by cash and cash equivalents to ensure comparability. We have also made a few adjustments to cash flow categories, reclassifying convertible bond flows from bonds to the equity channel and carving out asset-backed securities and real estate investment trusts (ABS/REITs) recorded under equities. The financing structures of the four groups are very different, relying on varying types of cash flows.

Hyperscalers: The hyperscalers spend the most on AI infrastructure, and hence their financing structures changed the most significantly from 2025 to H1 2026. Borrowing via loans and bonds became more important. The hyperscalers are also raising cash by selling existing short-term investment assets, with other cash flow channels turning positive in H1 2026.

Equity is becoming more important for hyperscalers’ financing. While increasing dividend payouts to keep their stocks attractive to the market, hyperscalers scaled back their stock buybacks substantially in H1 to save more capital for AI spending. Equity inflows should grow into H2 as Alibaba announced in August that it plans to place HK$80 billion in new shares. This marks its first new share placement since its Hong Kong listing in 2019, with 100% of the proceeds planned for AI investments.

Huawei and telecoms: This group depends almost entirely on operating cash flows, reflecting their oligopoly positions within China’s economy. They barely have inflows from other channels and the telecom giants carry minimal debt. Huawei still has a large balance of long-term bank loans totaling 201 billion yuan, as well as 38 billion yuan of bonds issued as of 2025. However, they are paying down debt as of last year. At the same time, they do not spend as much as the hyperscalers, with the telecoms’ financial structure seeing little change from 2025 to H1 2026.

Frontier AI labs: The frontier AI labs have negative operating cash flows. They are in the premature stages of business operations, and are spending heavily on research and development, with undeveloped business models. Most of their compute demands were met by renting compute from hyperscalers, so most AI spending fell under operating cash outflows. Going forward, their AI spending will likely show up more in capex as they build their own data centers.

On the financing side, the frontier labs depend heavily on equity issuance, and marginally on loans as of 2025. We don’t have cash flow details for H1 2026 yet, but equity-related inflows are surging. Private equity and venture capital (PE/VC) investments into Zhipu AI, MiniMax, DeepSeek, and Moonshot totaled 9 billion yuan in 2025. Then in the first eight months of 2026, the PE/VC investments, IPO inflows, and private placements soared to 179 billion yuan (Figure 5). The booming stock market helped Zhipu AI to raise around 4 billion yuan in its January Hong Kong IPO, followed by a 27 billion yuan private H-share placement in July.

Independent data center operators: The independent data center operators have the most diversified sources of financing. They operate through four equally important channels: operating cash flow, loans and bonds, equity issuance, and ABS, REITs, and financial leases. Unlike hyperscalers and telecoms, independent data center operators do not have a cash-generating business to help alleviate the capex burdens of AI spending, nor do they attract much in equity financing. Consequently, they tend to use ABS, REITs and financial leases to a greater extent.

For all of China’s AI firms, the ranking of different financing channels in order of overall scale and importance would be: operating cash flows > equity issuance > loans and bonds > ABS, REITs, and financial leases. But marginal new sources of financing probably depend far more upon equity issuance and new loans and bond issuance at this point.

China’s AI firms face familiar problems

The operating cash flows generated from the cloud, data centers, and AI business itself cannot yet sustain the capex planned for the sector. Both revenue and profitability remain low compared to other business lines for hyperscalers, and relative to US firms. Admittedly, the hyperscalers and telecom giants can cross-subsidize from their other business lines, effectively providing an ongoing cash infusion. However, their other business lines, especially for the hyperscalers, are closely linked to China’s domestic consumption, which is also facing significant headwinds.

Comparing AI revenues

Annual recurring revenues (ARR) of frontier labs and hyperscalers’ AI model businesses have surged this year, with Zhipu’s Model as a Service (MaaS) ARR soaring by 20x from $74 million in January to $1.6 billion in August. ARR of MiniMax also grew by 4x from February to August. Using ARR as an aggregate revenue metric is highly imperfect, of course, but it is used widely to capture trends in rapidly growing firms.

However, the total ARR for all of China’s AI models was still only around $10.7 billion based on the latest available data, which is around 10% of the recently reported levels of OpenAI and Anthropic at this point (Figure 7). For hyperscalers and telecoms, the scale of their cloud service revenues is also quite limited, accounting for around 13%² of total revenues.

Chinese frontier labs are supporters of open-weight models, which has limited their potential revenues. If a third-party cloud service provider decides to deploy their model and offer an API service to others, they do not have to pay the Chinese frontier lab that developed the model. This could be changing, as Moonshot and Alibaba are pushing for revenue-sharing agreements with major users of their open-weight models. Moonshot is already in discussion with Microsoft, Amazon, and Google about these arrangements, and its share of revenues could be as high as 30%.

However, even if a revenue-sharing agreement can be reached, if the cooperating cloud service provider is from China, then it is just a reallocation of existing revenue across firms, rather than new revenue. If the revenue comes from overseas customers, it will further increase the frontier labs’ reliance on foreign demand and associated potential geopolitical and compliance risks. Over 50% of Moonshot’s revenues came from overseas users since the end of 2025. Kuaishou’s Kling reportedly received 75% of its ARR from overseas customers through March. MiniMax reportedly received 61% of its revenue from overseas users in H1 2026.

Problems with eventual profitability

Chinese AI firms generally face greater financial pressures than their US peers, despite improving economics at the API level. The gap reflects structurally lower pricing power in China, shaped by customers who are less willing to pay, intense competition, and the widespread availability of open-weight models. Rising infrastructure and compute costs are adding further pressure, while efforts to shift toward higher-margin enterprise-focused and API services require additional investment and often result in near-term margin trade-offs. For hyperscalers, profitability is constrained by relatively low cloud margins and costly consumer-focused AI apps, where monetization continues to lag behind rapid growth in usage.

At the API service level, China’s best frontier lab’s gross margin matches Anthropic’s. Through optimization in its models and inference infrastructure, DeepSeek achieved a reported 83% gross profit margin from its API service by July, which is close to the 80%+ level of Anthropic, even with lower API pricing. Meanwhile, other Chinese frontier labs lagged behind: Zhipu’s interim report showed a 25% gross margin from its open platform and API service in H1, while the API service margin can reportedly improve to 50-60% running on company-owned devices under ideal conditions. MoonShot reportedly has a 45% gross profit margin from its API service last December.

However, at the company-wide level, Chinese frontier labs’ gross margins are lagging behind their US counterparts across the board. DeepSeek’s gross margins were reportedly 45% by July, lower than Anthropic’s margin in the mid-60% range. OpenAI reportedly has a lower Q1 gross margin of 39% as its revenue structure just shifted to 50+% revenue from corporates in August (enterprise-focused services generally produce higher margins than consumer-focused business). Nevertheless, OpenAI still outperformed Zhipu’s 26% and MiniMax’s 18% H1 gross profit margins despite the fact that these two Chinese labs feature enterprise-oriented revenue structures.

Lower token prices are probably responsible for the weaker profits of China’s AI firms. Deflationary pressures in the economy are persistent. China’s SaaS and cloud markets have historically featured customers less willing to pay, along with intense price competition. The widespread availability of open-weight models amplifies these pressures.

Since the beginning of this year, most frontier labs and hyperscalers in China have raised model prices. DeepSeek initially bucked the trend, announcing that its API prices would remain low throughout the second half. In August, however, it also increased prices across models (V4.1 Flash prices will reportedly be lowered in September, but output token prices will be still higher than pre-August pricing) and introduced higher peak-hour prices, given greater cost pressures.

However, even after the price hikes, Chinese frontier models are generally priced at $0.04 to $0.50 per task³ according to estimates from Artificial Analysis, while only the best models offer slightly higher prices in the range of $0.50 to $1 per task (Figure 8). In comparison, top-tier US models from Claude are priced at $2 to $4 per task, while higher-end GPT models are priced at $1 to $2 per task.

Besides comparisons between firms, longitudinal comparisons within Zhipu AI and MiniMax also show declining profitability. Zhipu and MiniMax’s gross profit margins fell from 41% and 25% in 2025 to 26% and 18% in H1 2026 respectively, despite an expansion of API service ARRs. Both companies are trying to adjust their revenue structure toward business-facing and API services. MiniMax made significant progress, with enterprise-focused ARR rising from 30% of the total in August 2025 to 80% by August this year. Zhipu shifted away from its most lucrative enterprise general-purpose large model service (42% gross margin in H1 vs 25% open platform and API margin), since its potential market is smaller than API services. Zhipu also acquired an AI infrastructure specialist firm Xcore Sigma for its 1GW data center project using domestic chips to improve margins of its API services. These firms had to offer discounts, make new investments, and rent more resources in the process, driving up costs and reducing overall margins.

Most hyperscalers do not offer a detailed breakdown of the profit margins of their cloud businesses. AliCloud’s adjusted EBITA margin was 9% in 2025 and 10% in H1 2026. This is significantly lower than the 2025 cloud service operating profit margins of Amazon and Alphabet’s 35% and 24% in 2025 and 39% and 34% in H1 2026. Alibaba CEO Eddie Wu said that he believes AliCloud’s margin can improve to 20%, which will still be lower than the operating margins of major US cloud service providers.

Beyond the cloud business, hyperscalers also have a strong presence in consumer-facing AI apps, which have higher user acquisition costs. Alibaba has attributed the fall of operating cash flow in its Q1 results to both AI spending and the cost of user acquisition of its Qwen AI app. In Q2, Alibaba reclassified its business segment, separating out a new AI app segment, which generated 3 billion yuan in Q2 revenue but created a 14 billion yuan net loss in adjusted EBITA terms. It was also reported that although ByteDance’s Doubao app has over 200 million daily users, its daily revenue is under one million yuan, primarily from e-commerce commissions. In contrast, Douyin (TikTok) has around 750-800 million daily users in China, generating much higher implied daily revenues exceeding one billion yuan via advertisements. High user acquisition costs and the lack of a clear path to profitability from consumer-facing AI apps will add to hyperscalers’ financial burdens.

Gross profit margins of independent data center operators have been quite stable, since they mainly operate a rental business and do not need to spend as much on software, AI models, and R&D. GDS, VNET, RangeIDC, and Aofei, maintained gross margins of 28%, 21%, 47%, and 34% respectively as of H1 2026.

Problems with non-AI business lines

Hyperscalers can use non-AI business lines to subsidize their AI investments, but household consumption trends in China suggest this will be less sustainable. 57% of hyperscalers’ 2025 revenues are linked to domestic consumption through e-commerce, entertainment, and advertisements (Figure 9).

For instance, the adjusted EBITA of Alibaba’s e-commerce group plunged by 44% in fiscal year 2026, primarily due to the price war with JD.com and Meituan. While Huawei and the telecom giants rely upon more stable revenue bases, 15% of their revenues were also exposed to domestic consumer demand as Huawei sells phones and intelligent driving solutions, directly linked to weakening consumer electronics and electric vehicle markets.

How much new financing is available?

So far this year, equity issuance and government-led financing have been the key marginal sources of cash flow for the AI sector. At the same time, the low-interest-rate environment has helped AI firms secure cheap financing through loans and bonds. Policy-supported ABS and REITs have also been a useful supplement for independent data center operators. Financial leasing has supported small data centers, but has started to generate new risks.

China’s AI sector financing has not relied heavily upon bond issuance, in contrast to US hyperscalers. In the months ahead, it is worth watching how much more financing new equity issuance can provide, and whether hyperscalers and frontier labs will begin tapping more onshore debt facilities to fund AI capex.

Equity issuance and government support

Announced equity investment in China’s AI sector in 2026 has reached 282 billion yuan through August 21, breaking historical records and more than doubling from the lows in 2023–2025 (Figure 10). Direct investment from government guidance funds and government funding platforms has also climbed, but state funding represented a lower proportion of the total this year (25%) relative to 30% in 2024 and 2025.

The decline in the share of government investment reflects rising interest from private investors in AI, as well as the June State Council guidelines pledging strict control of the establishment of new government investment funds. In principle, county and district-level local governments are not allowed to establish new funds.

Nevertheless, the proportion of direct government investment is still high at 25%, doubling from the average 13% level in the pre-pandemic years. Direct investment from government was highest in the AI chip and server sector. 47% of equity investment into the sector from 2023 comes from government guidance funds and investment platforms, while another 14% comes from banks, which are mostly state-owned (Figure 11). This means that more than 60% of equity investment into the subcategory of AI chip and server sectors comes from state-affiliated sources. Direct government investment in the AI cloud sector is the second-largest share, at 30% of the total.

These two sectors received relatively little investment compared to large language models and AI applications, where private investment has dominated. But AI chips and servers include financing of hardware, which can serve as collateral for loans, while Beijing is also prioritizing hard tech and chokepoint technologies for strategic reasons.

The valuation of China’s frontier labs also appears substantially lower than US counterparts OpenAI and Anthropic (Figure 12). Valuations relative to revenue appear exorbitant for Moonshot and DeepSeek at present. While the estimated ratios for OpenAI and Anthropic are 34x and 21x, the valuation-to-ARR ratios of Zhipu and Moonshot are higher at 46x⁴ and 50x, while DeepSeek has a ratio as high as 163x. The reported valuation of Moonshot has soared from $4 billion by the end of 2025 to $50 billion as of August.

The AI frenzy has helped frontier labs tap 179 billion yuan in equity financing so far this year, while Alibaba also plans to raise 69 billion yuan through a Hong Kong private placement. Hyperscalers and independent data center operators also managed to generate 56 billion yuan in cash inflows through convertible bond issuance among surveyed companies in 2025.

The key question is whether Beijing can maintain performance in the tech sector within the equity market, as the “national team” of state-directed funds did with blue chips in 2024. As most tech stocks do not offer dividends, the task will be more difficult than supporting stock prices of high-dividend companies like banks and utilities. With valuations at very high levels, any fluctuation in broader equity market performance in China likely risks the AI sector’s access to future financing. A Financial Times report in early September claimed that the securities regulator (CSRC) had recently warned bankers against a wave of offerings of “low-quality” companies, indicating regulators’ concern about maintaining investor confidence in the equity market. Government support to top-tier companies can be sustained, but the rest of the AI sector might be cut off from equity financing, limiting future capex.

Loans and bonds

Among AI firms, loans are generally preferred to bonds, at least for now. For hyperscalers, loans require less public disclosure of financial information. The largest marginal contributor to China’s AI capex this year is ByteDance, which has primarily financed its spending through offshore loans. ByteDance has reportedly secured a $29.6 billion offshore syndicated loan from around 30 banks in September.

Loans are also a larger pool of available financing within China’s state-directed banking system. Banks favor data center financing because it offers effective collateral, and the PBOC is encouraging banks to lend more to the tech sector. Amid weak credit demand, banks are also increasingly offering loans at discounted rates below LPR levels.

Cash net inflows through loans totaled 639 billion yuan for the surveyed companies in 2025, while inflows from bonds (excluding convertible bonds) totaled 341 billion yuan. Most of the balance of outstanding bonds is from hyperscalers, 87% of which were long-term USD bonds. Hyperscalers had preferred offshore USD financing through bonds for flexibility, but they are steadily shifting to yuan financing in recent years as interest rates have fallen.

According to Tencent’s disclosures, the midpoint of the interest rate range of long-term yuan loans was 2.8% in 2025, lower than 5.1%⁵ for long-term USD-denominated loans and 3.3% for long-term USD bonds. That rate for short-term yuan loans fell from 2.4% in 2025 to 2.1% in H1 2026. Consequently, Tencent added 42 billion yuan in short-term yuan-denominated loans in H1 2026, with the proportion of yuan loans within the total balance growing from 64% in 2025 to 71% in H1 2026.

Bond financing costs are still significantly lower in China than in the US, even though Chinese firms are using the bond market more sparingly. Long-term yuan bond coupon rates for Tencent and Baidu were around 2.6% by H1 2026, while the average effective rate of long-term bonds of Amazon, Alphabet, Meta, Microsoft, and Oracle was around 4.4%.

To encourage the bond market to support technology firms, there were new pilot projects to issue innovation-related corporate bonds starting in March 2021. However, China’s bond market still prefers SOE issuers, which enjoy implicit guarantees and usually have more tangible assets as collateral. State issuers remain over 60% of this market since 2022 (Figure 13). In May 2025, PBOC and CSRC announced new measures encouraging financial institutions, tech companies, and PE/VC funds to issue tech innovation bonds. Financial sector issuance rose to 491 billion yuan and 22% of the total in 2025, but direct issuance by tech companies remained limited. Instead, most support from the bond market is indirect, channeled through equity investments of financial institutions, PE/VC funds, and SOE or local government investment platforms issuing tech innovation bonds.

ABS, REITs, and financial leases

ABS and REITs are facilities designed by regulators to revitalize the existing assets of independent data center operators and reduce their debt ratios. In April 2024, the China Securities Regulatory Commission (CSRC) released “16 Measures for the Capital Market to Support the High-Quality Development of Technology Enterprises,” which proposed supporting new infrastructure such as data centers by issuing tech-innovation REITs. Last March, the private firm GDS issued China’s first data center ABS. Later in August 2025, China’s first publicly offered data center REITs were issued by RangeIDC and GDS on the Shenzhen and Shanghai Stock Exchanges. Financing through this channel should continue to grow in the future with policy support.

But for smaller independent data center operators who cannot smoothly tap these financing channels, financial leases are an alternative option. According to data released by the China Academy of Information and Communications Technology, in H1 2026, 54.3% of newly added computing center projects in China involved financial leasing. In Q1, China’s computing power leasing market grew by 62% y/y to 68 billion yuan.

However, the risks accumulated in this sector are also the highest among all the channels we have discussed. For instance, the bankruptcy of Sichuan Binji Technology, which used sale-and-leaseback structures to acquire GPUs using leverage, led to direct losses for 15 licensed leasing companies in Q2. Depreciation and market demand vary significantly by GPU model. Managing risk in this area remains a challenge, and if the compute market experiences further volatility, we could see more examples of bankruptcies and defaults.

Outlook for future AI-related financing in China

Overall, China’s AI sector faces many of the same questions as the industry in the United States.  Fundamentally, the frontier labs in China remain highly unprofitable, and they need new pricing models or ongoing infusions of cash to continue operating. AI-related capex will continue growing in China in late 2026 and 2027, but the pace remains uncertain.

Compute shortages are still binding, although they could ease later this year with more domestic supply and additional imports of H200s. Beyond limits on compute, financing remains another key constraint on the AI sector’s expansion.

For the economy as a whole, AI-related investment will remain far too small in scale to offset the macroeconomic impact of slowing property or infrastructure investments. Even in 2027, the entire sector would represent only around 1.5 to 2% of China’s overall investment, and a smaller proportion of GDP.

Equity issuance remains the most important financing channel for Chinese AI firms, unlike the US market. Thus, the stabilization of China’s stock market is important: Many domestic investors are highly concerned about the rapid fall in Unitree’s shares a week after its debut. The initial public offerings of MoonShot and DeepSeek are looming on the horizon, at still uncertain valuations.

The extent of government support of the stock market as a whole remains uncertain as well. Historically, the “national team” has supported index values via ETF purchases of blue chips, not the already soaring prices of a small number of tech companies.

Chinese AI firms do have options to raise funds in the bond market. It may be difficult for frontier AI labs and small data center operators to borrow through bonds, but hyperscalers and telecom giants, with ample cash flows and data centers serving as collateral, should be able to finance additional capex through bond issuance.