Tencent’s October bond subscription agreements illustrate the continued availability of long-duration capital to leading Chinese technology groups. The company announced dollar and offshore-renminbi tranches, with issuance expected later in the month. The proceeds are for general corporate purposes, so the transaction cannot be counted as funding for a particular AI campus. Meanwhile, the repricing of credit risk among some US AI infrastructure borrowers is a reminder that access to capital and attractive project returns are not the same thing.
12 OCTOBER · RESEARCH CONTEXT
12 October financing watch
Tencent’s proposed note issuance illustrates capital-market access, but the amount, completed proceeds and any AI-specific allocation must be verified before attributing financing to infrastructure projects. China’s 9 October policy also emphasises avoiding blind investment.
Contracts, collateral and strategic capital are different financing mechanisms
Recent industry research describes developers financing physical buildings against bank collateral and build-operate-transfer-style arrangements in which a hyperscaler provides land or shell while a partner finances mechanical/electrical infrastructure against a long service agreement. These structures help explain how AI campuses can expand without the entire cost appearing as hyperscaler-owned real estate. They are industry examples, not evidence that VNET or GDS uses each structure at a particular project.
Separate these mechanisms from policy lending eligibility, approved credit, disbursed debt, REIT/ABS recycling and strategic equity. Reported DeepSeek funding involving Tencent and CATL remains attributed reporting pending confirmation; it must not be described as closed capital or an operating compute-energy relationship.
Cash-economics test: compare financing tenor and cost with contracted service revenue, energisation, customer move-in, pricing resets, maintenance capex and debt service.
The state-coordinated financing system behind China’s AI build-out — and why it looks different from America. Chips, data centres and electricity dominate the AI infrastructure debate. But China is also building something less visible: a financing architecture that increasingly links national policy, state banks, commercial finance, infrastructure operators and the customers consuming compute.
3 October 2026China AI EvolutionPrimary-source led~18 min readPolicy → capital → compute
THE USUAL AI CAPITAL STORY
Companies raise and spend capital.
Hyperscalers, developers, utilities, private credit and capital markets finance different parts of the build-out. Government policy matters, but capital allocation remains predominantly market-led and institutionally fragmented.
Company → project → customer
THE CHINESE ANGLE
The infrastructure and the capital can be coordinated nationally.
China’s policy system is increasingly connecting compute with electricity and communications — while policy finance, state-linked banks, commercial lenders, leasing, industrial capital and local support are being organised around the same build-out.
National strategy → financial system → infrastructure → user
The global AI infrastructure conversation usually starts with hardware and ends with electricity. Who can obtain accelerators? Who can build enough data-centre capacity? Where will the power come from?
There is another constraint underneath all three: capital has to arrive before much of the revenue does. Land, substations, buildings, cooling, fibre, servers and accelerators have to be financed while customers move in and workloads ramp. At AI scale, the question is no longer simply whether demand exists. It is who carries the capital, for how long, at what cost and with what recourse.
China’s answer is becoming unusually important because the financing system is beginning to resemble the physical system we examined in Research Feature 01. The country is not only trying to coordinate compute geographically. It is increasingly coordinating the institutions that can fund the grid, network, data centre, equipment and user around it.
Our central finding: China is not simply financing more AI projects. It is beginning to build a capital architecture around its national compute architecture. That does not mean Beijing pays for everything. It means public policy can influence how capital reaches multiple layers of the system — alongside commercial and private capital.
The key difference is coordination — not “state versus private”
The comparison with the United States needs care. The US government influences AI infrastructure through semiconductor policy, tax incentives, energy policy, procurement and permitting. American hyperscalers also possess enormous cash flows and access to deep public and private capital markets.
China’s emerging distinction is therefore not that government exists on one side and markets on the other. It is the degree to which national infrastructure planning and the financial system can be deliberately connected.
In August, the National Development and Reform Commission described a new coordination mechanism linking the computing-power network, new power grid and next-generation communications network: “2+3+N” — two grid companies, three telecom operators and N compute-related enterprises. The meeting included State Grid, China Southern Power Grid, China Telecom, China Unicom and China Mobile, alongside compute companies and officials from the NDRC, MIIT and National Energy Administration. NDRC primary source ↗
EVIDENCED · NATIONAL COORDINATIONThe official description is explicit: the aim is to increase overall coordination and jointly accelerate major projects. A separate NDRC explanation says the “Six Networks” are intended to operate as a deeply coupled, mutually enabling whole rather than six independent systems. NDRC explanation ↗
That matters because compute is now being treated as infrastructure that must interact with other infrastructure. Once the physical architecture is coordinated, the obvious next question is whether capital can be coordinated around it too.
September changed the financing picture
On 29 September, the People’s Bank of China expanded the eligible areas for Pledged Supplementary Lending (PSL) to include the Six Networks: the water network, new power grid, computing-power network, next-generation communications network, urban underground pipeline network and logistics network. The one-year PSL rate was cut by 0.25 percentage points, from 1.75% to 1.5%, with the stated aim of better encouraging policy banks to support the real economy and national strategy.
What PSL does — and does not meanPSL is central-bank funding to policy banks. Inclusion of the computing-power network therefore creates a policy-finance channel around strategic infrastructure. It does not mean a private data-centre operator such as VNET or GDS automatically borrows at 1.5%, nor does it establish that any particular project has received PSL-backed funding.
The same PBOC package increased the technology-innovation and equipment-upgrade relending quota by RMB200bn to RMB1.4tn and raised the support ratio to 100%. The important point is not that every yuan flows into AI. It is that compute now sits inside a broader national infrastructure and technology-finance framework. Official-government report on the PBOC measures ↗
THE CAPITAL STACK
China can attack the financing problem at several layers at once.
01 · POLICY LIQUIDITYPBOC + policy-bank channelsPSL, relending and other structural tools can influence the availability and cost of funding upstream.
02 · STATE / INFRASTRUCTURE CAPITALPolicy banks · SOEs · local platformsCapital can be directed toward strategic networks, grid access, telecom infrastructure and large projects.
03 · COMMERCIAL FINANCEProject loans · syndicates · leasingBanks and lessors can finance facilities, equipment and long-lived infrastructure against project cash flows.
04 · STRATEGIC CAPITALIndustrial partners · JVs · customersEnergy, equipment or customer partners can potentially carry specific pieces of project capital rather than leaving everything at operator level.
05 · DEMAND FINANCECompute vouchers · contract / Token creditFinance can also help the customer buy compute, improving the route from installed capacity to paid utilisation.
06 · CAPITAL RECYCLINGREITs · ABS · refinancingMature cash-generating infrastructure can be refinanced or sold into capital vehicles, releasing money for the next build cycle.
Bank of China shows what this looks like when finance follows the network
Bank of China’s 12 September announcement is one of the clearest pieces of evidence because the bank describes the change in almost architectural terms. As China moves from relatively separate compute centres toward an interconnected national network, it says financial services need to upgrade with it.
Its “One Network, Three Financing” (一网三融) initiative was launched with China Telecom, China Mobile, China Unicom, PICC, China Orient Asset Management and CDB Financial Leasing, with support from MIIT and the PBOC. During the 15th Five-Year Plan period, BOC says the ecosystem will provide at least RMB300bn of comprehensive financial support, support at least 100 major intelligent-compute-centre projects and serve more than 3,000 compute-using enterprises.
More revealing than the headline number is what the money can touch. BOC describes lifecycle finance covering infrastructure construction, equipment procurement, operating improvement and asset revitalisation. It lists project loans, medium- and long-term compute-specific loans, technology syndicates and innovation bonds; financial leasing can support chips, servers, optical communications and liquid cooling. Bank of China primary source ↗
OUR ANALYSISThis is the important shift. Finance is not being attached only to a data-centre building. The bank is describing instruments around the full compute-production chain: the site, equipment, connectivity, operation, user and eventual recycling of the asset.
Follow the money through a 100 MW project
A simplified project makes the mechanism easier to see. Imagine an operator has a large customer commitment for 100 MW of AI-ready capacity. The operator still has to fund land and construction, grid connection, cooling, networking and possibly some equipment before all of the customer revenue arrives.
THE CAPITAL FLYWHEEL · SIMPLIFIED
The same yuan does not have to remain trapped in one operator forever.
1 · DEMANDCustomer contract / reservation
→
2 · BUILDProject debt + operator equity
→
3 · EQUIPLeasing + vendor / industrial capital
→
4 · OPERATEMove-in + compute consumption
→
5 · CASH FLOWContracted revenue becomes visible
→
6 · RECYCLEREIT / ABS / refinance / asset sale
→
7 · BUILD AGAINReleased capital funds new capacity
This is a framework, not a claim that every Chinese project follows the same path. The analytical question is how much capital can migrate to the layer best able to carry it, rather than forcing one listed company to fund the entire chain on its own balance sheet.
The lending model is moving from physical collateral toward compute cash flow
The financing evidence becomes more interesting when banks start underwriting the economics of compute itself. Earlier Chinese examples already included long-tenor fixed-asset loans, project finance, syndicated loans, equipment finance and green-compute lending. The latest Beijing experiment moves another step downstream.
Beijing Economic-Technological Development Area has launched the first batch of “Token loans” (词元贷). Six banks provided nearly RMB2bn of credit to AI-industry companies. The programme says credit assessment can use Token production and consumption and the value of compute-service contracts, alongside technology, team and business prospects, rather than relying only on property collateral and conventional profit statements. Beijing municipal source ↗
Land / building collateral→Project cash flow→Compute-service contracts→Measured Token activity
That progression matters. If compute becomes increasingly measurable, schedulable and contractable, the financial system can potentially underwrite more of the economic activity generated by the infrastructure rather than only the physical assets underneath it.
It is still early. A local pilot does not establish a national credit standard, and high Token consumption does not guarantee profitable customers. But it demonstrates the direction in which underwriting can evolve.
China is financing demand as well as supply
A data centre can be fully built and still be economically disappointing if customers do not consume enough compute. That makes demand-side finance unusually relevant.
Beijing’s revised compute innovation vouchers allow qualifying technology SMEs to cover up to 25% of the actual value of an eligible compute-service contract, subject to annual caps. The supply side therefore receives finance to build capacity while the demand side can receive support to consume it. Beijing policy text ↗
WHY THIS MATTERSSupply finance can accelerate capacity. Demand finance can accelerate utilisation. If both work, the interval between construction, customer consumption, cash flow and refinancing can shorten. Whether that produces attractive returns is a separate test.
Three company windows: the same system, different financing problems
The national architecture becomes more useful when we ask what it could mean for companies occupying different layers of the stack.
VNET · PHYSICAL AIDC
Who carries the build-out?
VNET’s demand evidence is strong: its 2026 research profile includes large wholesale orders and a substantial delivery pipeline. The financing question is whether every new megawatt must ultimately sit on VNET’s corporate balance sheet, or whether project debt, equipment finance, energy partners, JVs and asset recycling can carry more of the capital.
GDS is useful as a comparison because institutional and project capital can separate ownership of mature infrastructure from the operating platform. The key question is not simply leverage: it is how efficiently capital can be recycled while the operator keeps building.
KC sits closer to the compute and cloud layer. Equipment leasing, project credit and customer advances can change how infrastructure is funded; Token lending and compute vouchers show how the financing architecture can also reach the businesses buying AI compute.
VNET is especially instructive. Its Q2 2026 research profile showed 1,007 MW of wholesale capacity in service, 585 MW under construction and RMB10–12bn of 2026 capex guidance. Our earlier research also identified Chinese precedents for project loans, long-tenor fixed-asset loans, equipment finance, supply-chain finance and green-compute funding. The issue is therefore increasingly financing conversion, not simply proving that customers want capacity.
CATL adds a second question. Its acquisition of VNET shares was a secondary transaction and did not itself inject those proceeds into VNET. The companies have disclosed strategic cooperation, but no public commitment establishes that CATL will fund VNET projects, energy assets or JVs. Those remain scenarios to test — not facts to assume. See the VNET–CATL evidence boundary →
Why the American comparison matters
In the United States, AI infrastructure is also becoming financially innovative: hyperscaler balance sheets, investment-grade debt, private credit, infrastructure funds, utility investment, equipment finance and increasingly complex data-centre structures all matter. Public policy also influences semiconductor and energy investment.
What the Chinese evidence adds is a stronger top-down coordination layer. The same national policy framework can identify compute as strategic infrastructure, coordinate it with power and telecom networks, make it eligible for policy-bank liquidity support, encourage state and commercial financial institutions to create dedicated products, and support selected users of the resulting compute.
The distinction in one sentenceThe US system is predominantly capital-market-led with government intervention around it; China is increasingly placing market and commercial capital inside a nationally coordinated infrastructure framework. Neither system is purely public or purely private.
This does not automatically make China’s model cheaper, more efficient or more successful. It does mean that looking only at a listed operator’s cash balance can miss the wider set of institutions capable of carrying pieces of the build-out.
The advantage is coordination. The risk is also coordination.
A system able to align compute, grid capacity, telecom networks, banks, industrial partners and end users can potentially build faster and solve bottlenecks that would otherwise sit in separate institutional silos. That is the strategic attraction.
But abundant or policy-supported capital does not turn a poor project into a good one. AI hardware depreciates quickly. Utilisation can disappoint. Electricity and network constraints can strand nominal capacity. Subsidies can obscure weak economics. Project-level debt can still create losses even if it is not sitting directly at a listed parent.
What could work
Lower financing friction; longer-duration capital; better matching of asset life to debt; project-level risk sharing; faster grid/network coordination; recycling of mature assets; support for customer adoption.
What still has to be proved
Utilisation; returns above the true cost of capital; technology obsolescence risk; who absorbs losses; whether debt is genuinely non-recourse; whether listed shareholders capture enough of the upside.
The financing system therefore changes the question. Instead of asking only “Can China find enough money to build AI infrastructure?”, we should ask:
Can capital move repeatedly from policy and financial institutions → infrastructure → utilised compute → cash-generating assets → the next round of infrastructure, without leaving operators, banks or shareholders with inadequate returns?
CONCLUSION
The next phase may be determined as much by financial engineering as semiconductor engineering.
China’s AI infrastructure story is usually told through chips, megawatts and data centres. The financing evidence now adds another layer. Compute has entered the Six Networks infrastructure framework. The PBOC has opened a PSL-supported channel around those networks. Bank of China is assembling dedicated lifecycle finance around the national compute network. Commercial banks are experimenting with project, equipment, contract and Token-based credit. Beijing is supporting the customers buying compute as well as the infrastructure supplying it.
None of this proves attractive economics. It does show why the capital question cannot be answered by looking at company balance sheets in isolation.
WHAT THE EVIDENCE SUPPORTSChina is increasingly coordinating finance around the same compute, power and communications architecture it is coordinating physically.
WHAT IT DOES NOT PROVEThat every project receives state-backed finance, that private operators borrow at policy rates, or that coordinated capital guarantees good returns.
Research method: official and company sources establish the disclosed mechanisms. The capital-stack and flywheel diagrams are China AI Evolution analytical frameworks showing how those mechanisms could connect; they are not representations of a single government programme.