Rhodium examines the financing constraints behind China's AI expansion

The funding question behind China's AI buildout

Rhodium Group's September 17 report examines how Chinese AI companies pay for infrastructure, arguing that finance deserves attention alongside access to advanced processors. The authors estimate AI infrastructure capital spending of RMB 932 billion in 2026, about double the previous year's level, and more than RMB 1.2 trillion in 2027.

The study surveys thirteen companies across four groups: large internet platforms, Huawei and telecom operators, frontier-model laboratories, and independent data-center operators. Its central point is that these businesses do not draw on the same sources of money. Established operators can use cash from other activities, while young model companies depend more heavily on external capital.

Why rapid growth can still consume cash

Rhodium finds a mismatch between aggressive investment plans and the cash generated by AI businesses. It reports negative combined free cash flow for Alibaba, Tencent and Baidu in the first half of 2026 and argues that financing through equity and loans is more important in China than the bond-led approach seen among US peers.

This is a cash-flow question rather than a claim that revenue has stopped growing. Revenue, profit and free cash flow describe different stages of the accounts. A company can sell more cloud services while spending still more on equipment or advance payments. Those outlays can create a funding requirement before the resulting capacity earns a return.

A distinction between the lab and the data center

The report also separates a laboratory renting compute from a company buying infrastructure. Rental bills appear within operating spending, whereas owned equipment can appear as capital expenditure. Comparing only capex would therefore miss part of the cost of developing and serving models.

That accounting distinction explains why the same AI boom can look different in the financial statements of a model developer and its cloud supplier. One company's operating payment helps fund another company's infrastructure business. Adding their spending without examining the relationship can overstate the amount of independent new investment.

Different businesses raise money in different ways

The report puts private-equity and venture-capital funding for Zhipu AI, MiniMax, DeepSeek and Moonshot at RMB 9 billion in 2025. It then counts RMB 179 billion across private investment, IPO proceeds and private placements in the first eight months of 2026. The comparison spans different periods and a broader set of financing channels in 2026, so it is evidence of a financing surge rather than a like-for-like measure of annual financing growth.

Rhodium illustrates the public-market route with Zhipu's roughly RMB 4 billion Hong Kong IPO followed by a RMB 27 billion private H-share placement. It distinguishes this equity-supported laboratory model from independent data-center operators, which also use asset-backed securities, real-estate investment trusts and financial leases.

Those instruments shift the funding question toward the underlying asset and its future payments. A laboratory selling shares raises capital against expectations about its business; an infrastructure operator may also finance equipment or monetize property-linked cash flows. That difference is why the report's four business groups cannot be treated as having one common financing runway.

The report's argument, rather than a settled forecast

Rhodium concludes that further expansion remains sensitive to capital-market conditions and the cash generated by incumbent businesses. Its estimates are a research firm's scenario, not official totals for every Chinese AI investment.

The useful contribution is its breakdown of who bears which costs and how those costs are financed. It shifts the question from how many GPUs or models a company announces to whether the organization can fund the interval between building capacity and receiving enough cash from customers.