China's AI Data Centre Boom Hits Reality Check: ROI Now Key
Goldman Sachs warns that China's AI data centre sector must now shift from expanding capacity to proving profitable returns amid rising financing costs.
China's aggressive expansion of artificial intelligence (AI) data centres is entering a critical new phase. While the rush to build massive infrastructure has been relentless, a new research report from Goldman Sachs warns that the sector is now facing a stern "Return-on-Investment" (ROI) test as financing costs climb and investors demand capital efficiency.
For a long time, the primary goal for Chinese data centre operators was securing megawatts (MW) of capacity and signing new contracts. However, the narrative is shifting. According to the report, the industry is moving away from a "growth at all costs" mindset toward a period where businesses must prove they can generate returns that actually exceed the cost of the capital they borrowed.
The Shift from Capacity to Capital Efficiency
The demand for AI infrastructure in China has been stronger than many anticipated, which has bolstered the medium-term growth outlook for operators. But as these orders turn into long-term contracted capacities, the financial scrutiny is intensifying. Goldman Sachs notes that the next stage of investing in this sector is no longer about how much space or power a facility has, but rather how much profit it can actually squeeze out of that investment.
This transition is happening at a challenging time, as rising financing costs make it more expensive to fund the colossal amounts of capital required to build and maintain AI-ready facilities.
Three Business Models Under the Microscope
To understand where the money is—and where it isn't—Goldman Sachs analyzed three distinct business models currently operating within China's data centre landscape:
1. Traditional Powered-Shell Facilities
These are the basic infrastructure providers. While they offer stable revenue streams, the report suggests their returns are limited unless the operators can find very favorable financing or sell the assets at a premium. Under base-case assumptions, the estimated gross return on invested capital (ROIC) for this model is around 5 per cent.
2. Integrated AI Data Centres
These operators provide "GPU-as-a-service," offering graphics processing unit computing power under contracts typically lasting three to five years. This model has the potential for higher returns, though it is risky. Profitability depends on how much they pay for GPUs, the rental prices they can charge, and the looming threat of hardware becoming obsolete. The report suggests that configurations using servers equivalent to the Nvidia B300 could potentially support an annual ROIC of 15 per cent.
3. AI Token Factories
This is the most aggressive model, where companies monetize computing capacity through AI inference and the production of "tokens." This approach offers the highest potential rewards:
- Asset-light token factories: Modeled ROIC of approximately 21 per cent.
- Infrastructure-owning token factories: Modeled ROIC as high as 43 per cent.
The Path Forward: Efficiency and Expansion
As the sector matures, Goldman Sachs suggests that operators cannot rely on growth alone. To improve capital efficiency and survive the ROI test, companies may need to explore several strategic pivots:
- Asset Recycling: Selling off older or less efficient assets to reinvest in higher-yield technology.
- Lower-Cost Debt: Finding more affordable ways to finance their massive overheads.
- Overseas Expansion: Looking beyond domestic borders to find new revenue streams.
Ultimately, the long-term value of China's AI data centre businesses will no longer be measured by the size of their server farms, but by their ability to monetise computing power effectively while managing the mounting costs of debt.