The Two Economies of Artificial Intelligence

One sell-off, several different financial risks

The companies supplying AI computing do not necessarily own the same assets, finance them in the same way or control the demand that makes them valuable. A modern facility combines buildings, electrical systems, cooling, communications and expensive accelerators. These assets have different economic lives. Buildings and electrical infrastructure may remain useful through several hardware generations; accelerators can depreciate rapidly.

A property developer may own the facility. A specialist cloud provider may lease it and borrow to purchase GPUs. A model developer or hyperscaler may contract for capacity. Banks, bondholders and private credit funds may finance separate layers. This division accelerates expansion, but it creates dependencies: refinancing conditions, customer concentration, equipment depreciation and long-term lease obligations can combine into acute financial stress.

A financing model can fail without the underlying demand for computing disappearing. Customers may still need the capacity even as lenders seek repayment, equipment is sold or a facility needs a replacement operator. The debt-market pressure around CoreWeave illustrates why investors have begun scrutinising such arrangements. Reuters reported in mid-October that yields on some CoreWeave debt had risen towards 15%. Source ↗

But this is not the financial model of every American AI company. Google, Amazon, Microsoft and Meta operate deeply integrated businesses with major resources, proprietary technology and direct customer relationships. Even within America, an integrated hyperscaler and a highly leveraged GPU intermediary should not be valued as equivalent risks.

The same principle must apply to China. Geography alone is not a substitute for examining assets, liabilities, contracts and capital requirements.

China is building more than data centres

China’s East Data–West Computing programme seeks to connect demand in eastern economic centres with computing facilities in regions offering land and energy resources. Ulanqab, in Inner Mongolia, is one location where this is taking physical form. The country’s wider Six Networks programme places computing alongside electricity, communications, water and other infrastructure as a coordinated national development priority.

The economic significance lies in synchronising systems that individual developers cannot control. A data centre can be built before its grid connection is ready; it can receive electricity before adequate fibre is installed; or it can be technically commissioned before customers deploy their hardware. Every mismatch delays the conversion of invested capital into productive computing.

China’s coordination efforts—including arrangements linking electricity grid operators, telecommunications companies and computing enterprises—seek to address those dependencies. The related national compute-market initiative also aims to make resources easier to discover, connect, schedule and pay for across different operators. These are ambitions with operational components, not proof of a single fully fungible national pool. Read our separate investigation of the national compute market ↗

Xi’s Ulanqab visit: priority, but also discipline

On 16 October, Xi visited the Chahar High-Tech Development Zone. Official reporting records his emphasis on efficient operation of completed projects, timely delivery of subsequent developments, coordination between computing and electricity (算电协同) and water (算水协同), and the conversion of computing advantages into wider economic development. Source ↗

His presence is a meaningful signal of political priority. But his emphasis on productive operation is equally important. A completed facility is not necessarily a productive one. It needs power, equipment, connectivity and paying workloads. The visit signals strategic attention, not a guarantee of profitability or a commitment to protect listed shareholders.

Electricity is also a question of time

A facility unable to secure electricity for years can become commercially unattractive even if its eventual power price is competitive. Financing costs accumulate, customer requirements change and hardware procurement plans age. The proposed Microsoft–Nscale development at Loughton in Essex illustrates this constraint: Microsoft withdrew in October amid reports of substantial grid-connection delays. The precise reason for its decision has not been publicly established. Source ↗

China faces its own difficulties. Wind and solar generation vary, high-density computing requires reliable supply, and transmission, storage and backup power are expensive. Ulanqab’s developing grid infrastructure illustrates the need to coordinate capacity with power delivery; construction still in progress does not automatically establish a missed deadline.

A rigorous international comparison must measure approval-to-grid, grid-to-commissioning and commissioning-to-paying-workload times for comparable projects. The available evidence establishes a different coordination model—not that Chinese projects are universally faster or cheaper.

Cheaper AI is changing the demand equation

DeepSeek, Qwen, Kimi and other Chinese model families are challenging assumptions about the computing required to deliver useful AI. Their importance is not simply benchmark performance. More efficient models can reduce the cost of each task, making new applications viable across businesses, consumer services and industry.

That creates an apparent contradiction: if AI becomes cheaper, does infrastructure demand decline or expand? Both are possible. Suppose model improvements reduce compute per task by 80%, while lower prices encourage ten times as many tasks. Total computation would double. But if the price per task falls faster, infrastructure revenue might still disappoint. Conversely, efficiency may improve faster than adoption, leaving capacity idle.

China’s population and industrial base provide a large potential market, not a guaranteed paying customer base. The outcome depends on usage, price and compute intensity. Chinese models can also run on American clouds, so Chinese model adoption does not automatically produce Chinese data-centre demand.

Why the domestic technology stack matters

China’s restrictions on advanced semiconductor access remain significant. But they have encouraged work across domestic accelerators, compilers, software frameworks, model architectures and computing systems. Alibaba’s Qwen models, T-Head chip programme and Alibaba Cloud illustrate one form of integration; Huawei’s Ascend and enterprise systems illustrate another.

These firms do not manufacture every component or own every facility. The potential advantage is influence over the full cost of useful computation. A less capable chip can be commercially useful for particular workloads when software, models and networking are optimised around it. That is a system-level economic question, not merely a contest between headline GPU specifications.

America has integrated examples too—Google’s TPUs and Gemini, AWS and Trainium, Nvidia’s hardware-software platform. China’s distinctive undertaking is to develop a more extensive domestic alternative while coordinating the physical infrastructure beneath it.

From rack landlord to computing-energy infrastructure

AI is changing what data-centre operators must provide. High-density clusters demand sophisticated electrical conversion, cooling, storage, connectivity and power management. An operator may not own the chips or models, yet its engineering can determine the cost and reliability of the computing those systems deliver.

VNET is an instructive example. In August 2026 it announced a strategic cooperation agreement with CATL spanning gigawatt-scale computing-energy facilities, distributed computing-energy networks and a proposed zero-carbon token ecosystem. Subsequent Zhongwei plans described its Powerlet MicroGrid and Grid-to-Chip architecture, including cooperation involving local government and a CNNC-linked partner. CATL announcement ↗ Zhongwei announcement ↗

The ambition is to manage more of the energy system supporting high-density AI computing: conversion, storage, power quality and coordination with renewable supply. If it reduces operating costs or enables faster commissioning, this could make the infrastructure operator more valuable than a conventional property-oriented description suggests.

But these are proposed strategies, not demonstrated earnings. Their value depends on commissioning, capital requirements, achieved efficiency and who captures the savings. Customers, energy partners and equipment suppliers may receive much of the benefit.

Could computing become a flexible energy product?

China’s green electricity certificate system is already operational. Separately, researchers associated with China Southern Power Grid have proposed computing products that account for hardware characteristics, green electricity attributes, workload portability and interruptibility. Such designs could allow flexible AI tasks to run when and where electricity is cheaper or more abundant. Research paper ↗

The Shanghai Futures Exchange has also reportedly studied contracts linked to AI inference tokens. Those are a different concept: a potential financial instrument referencing AI usage rather than a composite electricity-computing product. VNET–CATL’s proposed zero-carbon token ecosystem is a third, commercially undefined concept. There is no demonstrated formal connection among the three.

Nevertheless, they raise an interesting question: could infrastructure providers eventually sell not just powered capacity but computing with specified timing, flexibility, location and electricity attributes? A customer requiring immediate uninterrupted inference may pay differently from one able to shift suitable workloads around renewable generation. The commercial market is immature, but the business model could change.

Could China experience its own AI infrastructure unwind?

Yes. But it need not resemble a crisis in American GPU-backed debt. Where customers own the accelerators, the facility operator is less directly exposed to their depreciation. Where energy firms finance generation and storage, the capital burden is shared. Infrastructure vehicles and bank financing can create different refinancing and asset-recycling options.

Those arrangements redistribute risk; they do not remove it. China has experienced overbuilding and poor data-centre utilisation. Rapid mobilisation of capital can accelerate the creation of stranded capacity when paying demand develops too slowly. A Chinese unwind might manifest as declining service prices, underused facilities, weak returns and loan losses absorbed by banks, local authorities or infrastructure investors rather than a concentrated neocloud default.

Less visible distress does not mean a smaller economic loss. Nor does a nationally important project necessarily create attractive returns for listed shareholders.

Commitments are not the same as utilisation

VNET’s second-quarter 2026 disclosures offer a useful illustration. Its wholesale capacity was approximately 96.3% committed but 73.9% utilised. Mature capacity was 92.5% utilised; newly developed ramp-up capacity was only 36.6% utilised. Company results ↗

The difference does not automatically demonstrate weak demand: newly delivered facilities often take time to fill and activate. But it shows why investors must distinguish customer commitments from operational capacity and mature revenue. Contract terms, ramp-up speed, capex, interest expense and cash conversion ultimately determine the returns.

The same discipline applies to national projects. Megawatts announced, megawatts delivered, megawatts equipped and megawatts generating revenue are not interchangeable measures.

Who actually profits from China's AI economy?

Four Chinese companies traded in American equity markets illustrate why a single AI-infrastructure valuation framework is inadequate.

Alibaba (BABA) combines models, cloud infrastructure, semiconductor design and customer relationships. Kingsoft Cloud (KC) provides cloud and AI computing services and participates in the developing scheduling and control layer. GDS (GDS) and VNET (VNET) are physical infrastructure operators with different assets, financing arrangements and strategic directions.

All participate in China's expanding AI economy, but they do not share the same exposure to hardware depreciation, power delivery, customer concentration or capital costs. An integrated platform may capture the value of better models and software; an infrastructure provider may benefit from demand without owning the accelerators, yet face bargaining pressure from its customers.

VNET’s CATL and Powerlet proposals raise the possibility that an infrastructure operator could retain additional value through energy engineering and management. That possibility needs to be demonstrated in margins and cash flow, not assumed from strategic announcements.

What would prove a genuine advantage?

The strongest test is not announced investment or theoretical processing power. It is the speed and cost of turning capital into useful computing: time to reliable power, time to commission, time to paying workloads, cost per useful AI task, utilisation, returns on invested capital and sustainable free cash flow.

China’s coordination could improve those measures. It could also facilitate overinvestment. The evidence currently supports a different system of development, not a categorical claim of superior financial performance.

Two economies, different financial consequences

The United States retains formidable advantages in semiconductors, software, frontier AI and capital markets. China is developing a different combination of domestic models, chip programmes, cloud platforms, electricity systems and coordinated infrastructure.

Both require enormous investment. Both can overbuild, misallocate capital and disappoint shareholders. But their financial vulnerabilities are not necessarily concentrated in the same places. The American neocloud boom has exposed risks in certain arrangements for financing computing equipment. China may distribute those exposures differently, while facing its own problems of utilisation and returns.

As electricity, storage, connectivity and workload flexibility become more important to the economics of computing, the companies enabling AI may also change their roles. The market may be right to question the price of the AI revolution. It may be less justified in treating every business associated with it as though it faces the same financial model.