字号·· | 护眼
南华早报

马来西亚、新加坡推进之际,香港为何在AI数据中心竞赛中落后Why Hong Kong is losing ground in AI data centre race as Malaysia, Singapore advance

点「原文对照」整页切到原文,或双击某段只看那段的原文。

从将军澳的大型厂房到葵涌紧凑的多层工业大厦,数据中心正蔓延至香港的每一个角落。

From large factory premises in Tseung Kwan O to compact, multilayered industrial buildings in Kwai Chung, data centres are spreading into every corner of Hong Kong.

前景一度看似光明。物业顾问公司戴德梁行在2023年将香港评为全球第四佳数据中心市场,而政府曾将香港定位为“数据中心的优越选址”。

The outlook once appeared promising. Property consultancy Cushman & Wakefield ranked the city the fourth-best data centre market globally in 2023, while the government had branded Hong Kong a “prime location for data centres”.

但现实并不那么美好。在该顾问公司今年的报告中,香港甚至未能跻身亚太区前十大市场,区域竞争对手如马来西亚柔佛州、澳大利亚悉尼和上海则脱颖而出。

But reality has been less kind. Hong Kong was not even among the top 10 Asia-Pacific markets in the consultancy’s report this year, with regional peers such as the Malaysian state of Johor, Sydney in Australia and Shanghai standing out.

随着算力需求超过本地供应,学者和业界领袖指出,美国对先进芯片的出口管制、土地稀缺和高昂电力成本,是制约香港数字基础设施的主要瓶颈。

As demand for computing power outstrips local supply, scholars and industry leaders have pointed to US export controls on advanced chips, scarce land and high electricity costs as major constraints on Hong Kong’s digital infrastructure.

他们说,与其试图在单纯规模上匹敌区域对手,香港不如深化与内地的联系,通过优先发展实用人工智能(AI)应用和模型部署,而非能耗密集的基础大语言模型(LLM)训练,来开辟自己的利基市场。

Rather than trying to match regional rivals on sheer capacity, they said, the city should deepen ties with mainland China and carve out a niche by prioritising practical artificial intelligence (AI) applications and model deployment over the energy-intensive training of foundation large language models (LLMs).

随着香港于9月推出首个五年规划,AI及更广泛科技产业的发展已成为焦点。

With Hong Kong rolling out its first five-year plan in September, the development of AI and the broader technology sector has taken centre stage.

然而,香港的数据中心扩张步履维艰,而内地和东南亚的区域竞争对手却加快了脚步。

However, Hong Kong’s data centre expansion has been sluggish, while regional peers on the mainland and in Southeast Asia have picked up speed.

创新科技及工业局引述市场数据称,香港拥有61个数据中心,与2022年推出《创新及科技发展蓝图》时几乎持平。

Citing market data, the Innovation, Technology and Industry Bureau said Hong Kong was home to 61 data centres, almost unchanged from 2022, when the city introduced its Innovation and Technology Development Blueprint.

根据经济合作与发展组织(OECD)的一个数据集,香港有9个可用的商业云区域,指共享电力供应和网络的数据中心集群。其中,只有4个,即44%,配备了先进AI芯片。

According to an Organisation for Economic Cooperation and Development (OECD) data set, Hong Kong has nine available commercial cloud zones, referring to clusters of data centres sharing power supplies and networks. Of these, only four, or 44 per cent, are equipped with advanced AI chips.

尽管同样受地理条件限制,新加坡拥有19个可用区,其中74%具备高AI算力。

Though geographically constrained like Hong Kong, Singapore has 19 available zones, with 74 per cent of them boasting high AI capability.

经济合作与发展组织(OECD)的测量方法由芬兰阿尔托大学技术政策教授、牛津大学牛津互联网研究院高级研究员Vili Lehdonvirta开发。

The OECD’s measurement methodology was developed by Vili Lehdonvirta, a professor of technology policy at Aalto University in Finland and a senior fellow at the Oxford Internet Institute at the University of Oxford.

香港已不再位列全球十大成熟数据中心市场。图片来源:Getty ImagesHong Kong is no longer on the list of the world’s top 10 established data centre markets. Photo: Getty Images

Lehdonvirta表示,这些数据可能无法反映真实的算力水平——理论上,一个可用区可容纳多个图形处理单元(GPU,常用于AI工作负载的先进芯片)——但这是其团队能设计出的最佳测量方式,因为私营运营商倾向于对其部署的GPU数量保密。

Lehdonvirta said that while the figures probably did not reflect the true computing capability – in theory, one available zone could house multiple graphics processing units (GPUs), advanced chips commonly used for AI workloads – it was the best measurement his team could devise, as private operators tended to keep the number of GPUs they had deployed secret.

就香港而言,Lehdonvirta称情况可能更糟,因为受美国对其最先进GPU的出口限制,该市拥有的高算力芯片可能是旧版本。

In Hong Kong’s case, Lehdonvirta said the situation could be even worse, as the highly capable chips in the city might be older versions because of US export restrictions on its most advanced GPUs.

他说:“香港的公司要进口最新的英伟达芯片本就困难,而我们知道新加坡拥有大量非常先进的英伟达芯片。”

“It would have been difficult for the companies [in Hong Kong] to import the latest Nvidia chips, whereas we know that in Singapore there are lots of very advanced Nvidia chips,” he said.

“因此,即便是那44%,从这个意义上说,恐怕也是对香港地位的高估。”香港面临美国对尖端半导体出口许可的“拒绝推定”,工业与安全局要求对流向其认定构成国家安全关切的目的地(包括中国、俄罗斯和朝鲜)的先进AI芯片发放许可证。

“So even that 44 per cent, in that sense, might be an overstatement of Hong Kong’s position, unfortunately.”Hong Kong faces a presumption of denial for US export permission covering leading-edge semiconductors, with the Bureau of Industry and Security requiring licences for advanced AI chips bound for destinations it deems to pose national security concerns, including China, Russia and North Korea.

面对这类限制,Lehdonvirta表示,香港不妨与中国内地建立更紧密的技术联系,以获取更先进的芯片。

Facing such restrictions, Hong Kong would be better off forging closer technological ties with mainland China to secure more advanced chips, Lehdonvirta said.

他说:“如果你仍与内地保持壁垒——你现在承担了被视为内地一部分的代价(无法获得最好的美国芯片),却也得不到来自内地的芯片红利,那你就是两头都吃亏。”

“If you still maintain a barrier with mainland China – you now get the costs of being considered part of mainland China [for not getting the best US chips] – you also do not get the benefit of the chips from mainland China, then you are just losing on both accounts,” he said.

Lehdonvirta 补充说,北京在该市实施《国家安全法》也使法律制度变得更加复杂,并让外国投资者处于“观望”状态。

Lehdonvirta added that Beijing’s imposition of the national security law on the city had also complicated the legal regime and put foreign investors in a “wait and see” position.

他表示,即使围绕法律环境的不确定性消退,香港也错失了一个机会窗口,而新加坡等地区同行已经抓住了这个机会。

Even if uncertainties surrounding the legal environment subsided, Hong Kong had missed a window of opportunity that regional peers such as Singapore had seized, he said.

“对于海外公司来说,一旦它们在那里建立了 AI 中心,就很可能会继续沿着路径依赖发展。当下一代芯片推出时,它们会把芯片放在同一个地方。”他补充道,另一方面,内地公司在香港建设数据中心的意愿也可能不高,因为土地和电力价格高昂,而中国的网络安全立法使得香港设施难以接入内地市场。

“For overseas companies, once they establish an AI centre there, they are likely to continue with a pathway dependency. When the next generation of chips comes, they put it in the same place.”On the other hand, mainland companies might also be lukewarm about building data centres in Hong Kong, because of high land and power prices, while China’s cybersecurity legislation had made it difficult for Hong Kong facilities to tap into the mainland market, he added.

2024 年,戴德梁行将香港从全球十大成熟数据中心市场名单中剔除,理由是成本高昂且土地供应稀缺。

In 2024, Cushman & Wakefield dropped Hong Kong from its list of the world’s top 10 established data centre markets, citing high costs and scarce land availability.

在最近的中期选举后民主党控制了美国国会两院,Lehdonvirta 估计,全球最大经济体的数据中心投资将放缓,随着投资者胃口依然旺盛,这将为其他目的地创造机会。

With the Democrats controlling both chambers of the US Congress after the recent midterm elections, Lehdonvirta estimated that data centre investment would slow in the world’s largest economy, creating opportunities for other destinations as investor appetite persisted.

不过,他警告说,香港需要提供“具有竞争力的条件”,因为许多地区同行也将争夺这些投资。

However, he cautioned that Hong Kong would need a “competitive offering” as many regional peers would also be competing for that investment.

沙岭数据中心集群选址。图片来源:Elson LiThe site of the Sandy Ridge data facility cluster. Photo: Elson Li

在阿尔托大学与 Lehdonvirta 共事的博士后研究员李坤涵博士援引专有数据估计,香港目前的数据中心产能为 750 兆瓦。

Dr Li Kunhan, a postdoctoral fellow working with Lehdonvirta at Aalto University, estimated that Hong Kong currently had a data centre capacity of 750 megawatts, citing proprietary data.

相比之下,新加坡已拥有 1.4 吉瓦的产能——约为香港的两倍——并计划进一步扩大三分之一。

In contrast, Singapore already boasts 1.4 gigawatts of capacity – around double Hong Kong’s – while planning to expand it by a further one-third.

据李坤涵介绍,本地公司运营着香港约 40% 的产能,其余由外资公司运营,其中大多数来自美国。

According to Li, local companies run about 40 per cent of Hong Kong’s capacity, while foreign firms account for the rest, with most of them coming from the United States.

“目前内地运营商占比不足十分之一,但他们是最大新项目的幕后推手,包括沙岭集群,”他说,指的是一个占地11万平方米、毗邻内地的设施,预计在2032年建成时将提供180亿亿次浮点运算(EFLOPS)的算力,是目前全市水平的36倍。

“Mainland Chinese operators hold under a tenth today, but they are behind the biggest new projects, including the Sandy Ridge cluster,” he said, referring to a 110,000 square metre facility bordering the mainland that is expected to provide 180 exa-floating-point operations per second (EFLOPS) of computing power, 36 times the current citywide level, when it is completed in 2032.

算力既可用瓦特为单位的电力供应来衡量,代表输入容量;也可用FLOPS(每秒浮点运算次数)的输出容量来衡量。

Computing power can be measured by electricity supply in watts, which represents input capacity. It can also be measured by output capacity in FLOPS.

由于私营运营商倾向于对这类信息保密,许多经济体的实际输出容量仍不得而知。

The actual output capacity in many economies remains unknown because private operators tend to keep such information confidential.

李观察到香港对数据设施的需求不断增长,但表示土地供应制约了该行业的快速扩张。

Li observed growing demand for data facilities in Hong Kong, but said land availability had hindered the sector’s rapid expansion.

他说:“过去一年空置率下降、租金上涨,因为云计算和AI需求消化了新增供应。不过,大多数现有建筑的设计电力密度远低于AI所需。”

“Vacancy has fallen and rents have risen over the past year as cloud and AI demand absorbed new supply. Most existing buildings, though, were designed for much lower power densities than AI requires,” he said.

尽管面临逆风,一些运营商仍渴望在香港建设更多数据设施,以抓住本地AI使用增长带来的商机。

Despite the headwinds, some operators are keen to build more data facilities in Hong Kong to seize business opportunities created by growing local AI use.

其中之一是总部设在本地的AI企业商汤科技,该公司已在内地建设并运营大型数据设施。

One of them is the locally headquartered AI firm SenseTime, which has built and run large data facilities on the mainland.

商汤科技与政府拥有的香港科技园公司合作,计划在2030年全面投运时推出一个主要由国产GPU提供算力的数据中心,目标提供40 EFLOPS算力。

In collaboration with the government-owned Hong Kong Science and Technology Parks Corporation, the firm is planning to launch a data centre mainly powered by Chinese-made GPUs when it is fully operational in 2030, aiming to provide 40 EFLOPS of computing power.

商汤科技香港及澳门业务董事总经理冯乐思表示:“如今,人人都在消耗Token。仅就香港本地的不同行业而言,从教育、金融到物流,都消耗大量Token,而目前的供应根本不够。”

“Nowadays, everyone consumes tokens. Even just for different industries within Hong Kong itself, ranging from education and finance to logistics, they consume a lot of tokens and the current supply is simply not enough,” said Lewis Fung, SenseTime’s managing director of Hong Kong and Macau business.

在人工智能领域,Token 是微小的数据单位——例如词片段、像素或声音片段——模型将其拆解并处理,以读取提示词并生成回复。作为 AI 服务的基本货币,Token 的大量消耗需要数据中心提供海量的算力吞吐量。

In AI, tokens are the tiny units of data – such as word fragments, pixels or sound segments – that models break down and process to read prompts and generate responses. Acting as the basic currency of AI services, heavy token consumption requires massive computing throughput from data centres.

冯骥才表示,本地专业服务机构以及在港设立据点的非本地公司,将有动力将数据存储在香港,以避免跨境数据流动带来的合规负担,从而带来对数据设施的旺盛需求。

Fung said local professional service firms and non-local companies establishing a presence in the city would have incentives to store their data in Hong Kong to avoid compliance burdens arising from cross-border data flows, leading to keen demand for data facilities.

冯刘易斯(Lewis Fung)是商汤科技香港及澳门业务董事总经理。图片来源:Jonathan WongLewis Fung is SenseTime’s managing director of Hong Kong and Macau business. Photo: Jonathan Wong

商汤科技在上海拥有并运营一座数据中心,目前算力规模超过 40 EFLOPS。据冯骥才透露,该设施十分受欢迎,公司正在规划二期项目。

SenseTime owns and operates a data centre in Shanghai currently with a computing capacity of more than 40 EFLOPS. According to Fung, the facility has been so popular that the company is planning a second phase.

尽管面临美国出口管制,上海设施通过以国产芯片为主、辅以较不先进的美国 GPU,实现了良好的效率,商汤科技计划在香港复制这一成功经验。

Despite facing US export restrictions, the Shanghai facility has achieved good efficiency by anchoring Chinese-made chips supported by less advanced US GPUs, and SenseTime plans to replicate such a success in Hong Kong.

他说:“我们的主要限制意味着国产芯片将是主力,但如果混合使用能获得最佳性能和功耗表现,我们就会采取这条路线。技术迭代迅速,我们正在逐日评估各种方案。”

“Our primary limitation means domestic chips will be the main force, but if mixing them yields the best performance and power savings, that is the route we will take. The technology is iterating rapidly, and we are assessing the options day by day,” he said.

不过,冯骥才补充道,香港企业从内地获取稳定芯片供应也将面临困难,因为内地需求同样旺盛,商汤科技必须更加努力地与国产供应商开展合作。

However, Hong Kong companies would also have a hard time securing stable supplies from across the border, as demand on the mainland was also high, and SenseTime would have to work harder on collaborating with Chinese suppliers, Fung added.

针对香港市区土地受限的挑战,公司正关注北部都会区大型项目下的新田科技城地块,首批土地预计明年交付。

Acknowledging the challenges posed by land constraints in urban Hong Kong, the company is eyeing plots in the San Tin Technopole under the government’s Northern Metropolis megaproject, with the first batch of land expected to be delivered next year.

冯骥才还指出,供电设备供应存在瓶颈,变压器——将高压输电线上的电能转换供商业使用的设备——短缺,阻碍了向更大规模数据中心扩展供电能力。

Fung also noted a bottleneck in electricity equipment supply, with a shortage of transformers – devices that convert power on high-voltage transmission lines for commercial use – hindering the expansion of power supply to bigger data centres.

他呼吁本地供应商与内地大型制造商合作,提前锁定变压器供应,并呼吁政府提供政策支持,确保来自内地的电力供应稳定,并给予优惠电价。

He called on local providers to partner with major mainland manufacturers to secure transformers ahead of time, and the government to provide policy support by securing a stable power supply from the mainland and offering concessionary electricity prices.

针对《南华早报》的询问,创新科技及工业局表示,香港发展成为区域数据中心枢纽拥有独特优势,包括强大的光纤连接、稳定的电力供应、较低的自然灾害风险,以及由城市商业活动产生的海量数据流。

In response to the South China Morning Post’s queries, the Innovation, Technology and Industry Bureau said Hong Kong possessed distinctive advantages in developing into a regional data centre hub, including its strong fibre cable connectivity, a stable power supply, low natural disaster risks and an abundance of data flows generated by the city’s business activities.

该局补充称,政府正准备委托一项关于算力的新顾问研究,重点关注算力供给效率、电力供应稳定性和低碳转型。

It added that the government was preparing to commission a new consultancy study on computing power, focusing on computing power supply efficiency, electricity supply stability and the low-carbon transition.

创新科技界立法会议员邱达根表示,面对高昂的电价,香港应将自己定位为开发人工智能模型应用场景的枢纽,而不是盲目寻求在本地从头训练基础模型。

Innovation and technology sector lawmaker Duncan Chiu said that, faced with high electricity prices, Hong Kong should position itself as a hub for developing application scenarios for AI models, rather than blindly seeking to train fundamental models from scratch locally.

他说:“可以肯定的是,香港不是运营数据中心成本最低的地方,因此在香港训练基础模型是否有益仍属未知。”

“It is certain that Hong Kong is not the cheapest place to run data centres, so it remains unknown whether it is beneficial to train fundamental models in Hong Kong,” he said.

“但在一个基础模型之下,有许多应用层可以在香港开发,香港企业和科研项目对人工智能应用有强烈需求。”邱达根还警告不要过快扩张人工智能数据中心。他说,步调应主要与市场需求保持一致,因为设备折旧很快。

“But under one fundamental model, there are many application layers that could be developed in Hong Kong, and there is strong demand [for AI applications] from Hong Kong enterprises and scientific research projects.”Chiu also cautioned against expanding AI data centres too rapidly. Instead, the pace should largely be in line with market demand, as equipment depreciated quickly, he said.

Lehdonvirta、李氏和冯氏也表达了相同观点,认为香港可以通过在其传统优势领域——包括金融、贸易、医疗和物流——探索应用场景,在全球市场中找到独特定位。

That sentiment was echoed by Lehdonvirta, Li and Fung, who said Hong Kong could find its unique position in the global market by exploring application scenarios in its traditional strengths, including finance, trade, healthcare and logistics.

“虽然在我看来,它可能不是许多国际投资者的首选,但对于中国大陆的AI公司来说,它是一个优质的离岸选址,可提供更好的国际数据连接,”阿尔托大学数字经济安全实验室的李某表示。“它也可以专注于推理,以及在其已具优势的领域应用AI。”

“Although it might not be a top choice for a lot of international investors in my opinion, it is a prime location for mainland Chinese AI companies as an offshore location for better international data connectivity,” Li, of Aalto University’s Digital Economic Security Lab, said. “It can also focus on inference and on applying AI in sectors where it is already strong.”

阅读媒体原文 ↗