Beijing And Its AI Billions

Another month, another round of billion-dollar investments in artificial intelligence.

Chipmaker Nvidia is giving Sam Altman’s OpenAI $100 billion to get its data centres going — ostensibly for the high-profile $500 billion Stargate Project.

Meta chief Mark Zuckerberg wants to spend $600 billion on building American data centres and AI infrastructure by 2028. He says he’s okay with “misspending a couple of hundred billions” instead of missing the AI train.

American tech firms have clearly dug in their heels on AI — “bubble” be damned. And their Chinese counterparts are now starting to catch up too.

This month, despite Beijing’s larger pushback against chips from Nvidia, Chinese internet giant Alibaba announced a partnership with the chipmaker for its AI development tools. Alibaba also pledged to invest $53 billion in AI-related infrastructure over the next three years.

That number is relatively tiny compared to the hefty billions announced by America Inc just this month, but analysts say Chinese companies have accelerated their AI investments over the past couple of years, and specifically in the last quarter.

Those investments run parallel to Beijing’s larger vision for AI infrastructure in the country — one that involves building data centres on rice fields, deserts and even underwater. Beijing’s plan — which one supplier described as “the Stargate of China” to the Financial Times this week — is an ambitious one, and driven now by a dearth of advanced chips and a glut of computing power.

Proposed first in 2020, the plan called ‘Eastern data, Western computing’ set out to develop a nationwide data centre and computing system under which data from China’s populous eastern cities was to be transferred to the more resource-rich, but rural, western provinces. As part of the project, Beijing set out to build eight “computing hubs” across the country containing ten “data centre clusters”.

Construction on many of these is still ongoing. The FT described one of these hubs in the eastern city of Wuhu as a “Data Island” where some of China’s biggest state-funded firms — Huawei, China Telecom, China Unicom and China Mobile — have built four new AI data centres to cater to bustling cities like Shanghai and Nanjing. The data centres in the east were meant initially to handle and process data as industries and businesses in the region quickly digitised. But as Beijing is also tweaking its plan as the AI industry evolves, these data centres now run a far more important task — AI inference, the process that allows large language models (LLMs) to use their existing data to make predictions or decisions in response to a user query.

Meanwhile, data centres in remote Western regions were initially designed to carry out more resource-hungry tasks like providing data storage and backup. They are now also tasked with the all-important training of LLMs that make breakthroughs like DeepSeek possible.

The plot twist

Back in 2020, the idea behind creating a network between the west and the east was — as China Briefing explained back then — to “correct an imbalance in supply and demand”. A vast majority of data centres at the time were located in the east to meet the huge supply of data, but they became very costly and difficult to maintain as they needed immense computing power and, hence, massive amounts of energy. Meanwhile, the vast and thinly populated western regions were left largely untapped even as they became powerhouses of renewable power. So, on paper, balancing out the two was practical.

All that practicality went out the window, however, once Beijing officially launched the project in 2022. Local provincial governments, who are often eager to get in line with the authoritarian CCP’s policies, began investing blindly in data centres. And that trend exploded with the launch of ChatGPT later that year. Between 2023 and 2024, over 500 new data centre projects were announced in the country, according to MIT Technology Review.

Companies, often with little to no experience in AI, were jumping into the industry in partnership with local governments. MIT noted how one firm manufactured MSG, the ingredient known to give Chinese food its distinct flavour. Another was a textile firm. These businesses were really just in it to boost their revenue, and because they clearly had no idea about the technology, they weren’t prepared for what came next.

Until late 2024, companies could make money from data centres by renting them out to companies that needed computing capacity for AI training. But as inference became the more coveted feature, AI companies no longer prioritised training LLMs. Now they needed data centres powered by advanced chips and closer to tech hubs to minimise transmission delays, and ensure access to highly skilled staff. So all these data centres in the Western regions — made over many months and with many millions worth of processors — were now left idle. As of July this year, only about 30% of China’s intelligent computing capacity was being used, one state-linked report found.

In typical Chinese fashion, state subsidies worsened the problem too. As of last year, the Chinese state had spent over $6 billion to fund these data centres. One person associated with the Wuhu project also told the FT how local government subsidies were potentially covering up to 30% of AI chip procurement costs. According to MIT, the allure of these subsidies was such that many contractors got into the business simply to game the system.

It wasn’t until early this year that Beijing finally began discouraging AI overinvestment, with a stern warning coming from the highest rung of the ladder — President Xi Jinping.

Meanwhile, state investment has continued for data centres used in intelligent computing (tasks like inference). They received 24.7 billion yuan ($3.4 billion) in funding last year and another 12.4 billion yuan in the first half of this year.

Banking on homegrown tech

Where China’s plan to rival Stargate gets truly ambitious is when it comes to ‘linking’ the hubs in the east and west. According to the FT report, Beijing has directed its firms China Telecom and Huawei to use their networking technology to link processors scattered across the country. It’s a method cloud service providers already use to connect multiple sites together in order to create redundancy in the event that one goes offline.

But connecting China’s old and new data centres this way is not likely to be very efficient. Different data centres can be powered by different kinds of chips from both foreign and domestic firms, like Nvidia and Huawei. A potential integration of these will be tedious and expensive. Plus, considering many data centres were constructed by companies that were not even in the tech industry, they are likely to fall short of the standards necessary to make data transfers efficient.

The technology to fix these challenges is so far underdeveloped. Still, Huawei is reportedly working on fixing that with new open-source technology that it says is designed to unify all interconnections across AI data centres. Early this month, China’s Open Data Centre Committee (ODCC), an organisation representing the country’s key AI infrastructure investors such as Alibaba and China Telecom, also hosted a three-day conference to discuss the creation of a unified network.

A breakthrough on that could also be key to Beijing’s larger plan to harness all of the country’s surplus computing power. In July, Reuters reported China was working towards building a national, state-run cloud service that would standardise the country’s public computing power. Beijing also plans to sell its excess computing power by way of this cloud service.

If successful, the plan and any innovation it involves would serve another crucial objective for Beijing — cutting China’s reliance on Western technology. That objective is already key to the project, considering Chinese AI data centres have had to depend on less powerful processors or tap into the black market to put together advanced hardware.

Early this month, though, state media reported China Unicom was using homegrown AI chips to power a $390-million data centre in the Qinghai province in the west. The state-run telecom giant had so far used 23,000 domestically made AI chips sourced from companies like Alibaba and planned to buy more from others like Biren Tech, Moore Threads, and Enflame. The report also said a newly developed Alibaba chip was a close competitor to Nvidia’s H20, signalling China was set to step up its use of homegrown chips for intelligent computing. That same day, the Financial Times reported that China’s internet regulator had banned top tech firms from buying Nvidia’s AI chips.

Amidst all this, the one area where China could gain a definitive upper hand over its Western competitors will be with energy. Data centres are known to consume humongous amounts of energy and water and China has set out lofty goals to tap into renewables to meet these demands, such as by channeling green electricity directly into data centres. In July last year, Chinese officials said green energy will provide for 80% of the power demand from newly built data centre hubs. Early this month, state-backed CCTV claimed that some newly built data centres in regions like Inner Mongolia already were already using green electricity to meet 85% of their needs. Tech giant Tencent, meanwhile, has pledged to power its data centre operations with 100% renewable electricity by 2030 even as the transition remains fraught with challenges.

The focus on managing energy costs has spurred innovation in methods to cool data centres as well. According to CCTV, at the data centres in Wuhu, servers are fully immersed in a special liquid to cool down processors. This liquid cooling technology has the potential to save the hub’s energy consumption by 30%, the state broadcaster says.

• US President Trump has signed an executive order approving a deal that would keep TikTok alive in the US and value the business at $14 billion.

• Trump’s $100,000 fee for new H-1B visas could push American firms to increase their AI use and set up offshore offices for more of their operations, specifically in India, Canada and Mexico.

• Asian nations, including Bangladesh, Sri Lanka, Hong Kong and Chile, are queuing up to join the world’s largest free-trade bloc RCEP.
  

• And Vietnam’s exports to the US could fall by close to 20% — or more than $25 billion — because of tariffs imposed by Trump in August.