China Is Building AI Infrastructure Where Nobody Else Is Looking — and That Is the Point

The Lin-gang facility off Shanghai is not primarily an engineering story. It is a strategic story about who is moving from AI ambition to AI infrastructure at the speed the next decade requires.

Sign up for my Substack daily AI newsletter here.

See my AI Training course portfolio for corporate Business Leaders here.

Article link: https://open.substack.com/pub/johanosteyn/p/china-is-building-ai-infrastructure?r=73gqa&utm_campaign=post&utm_medium=web&showWelcomeOnShare=true

Follow me on LinkedIn: https://www.linkedin.com/in/johanosteyn/

In May 2026, a data centre entered commercial operation in the waters off Shanghai’s eastern coast. It sits 6.2 miles offshore and 33 feet below the sea surface. It holds 192 server racks across four levels, processing AI workloads including big data annotation and the development of domestic large language models. It uses seawater as a natural cooling medium through a copper-pipe heat exchange system, reducing the cooling share of total electricity consumption from approximately one-third — the industry norm — to approximately one-tenth. It draws more than 95 per cent of its electricity from offshore wind farms through subsea cables. It eliminates freshwater use for cooling entirely. It cuts land use by more than 90 per cent compared with an equivalent land-based facility. It maintains a power usage effectiveness of approximately 1.15, against an industry average of 1.5 to 1.6. China Telecom and local service providers have already connected to the platform.

The facility entered operation at an initial load of 2.3 megawatts and is scaling toward its planned ceiling of 24 megawatts — a capacity designed to support roughly 2,000 deep-learning and GPU servers. Total investment: 226 million dollars, approved by China’s State Council.

The engineering is impressive. The strategic signal is more important.

CONTEXT AND BACKGROUND

The physical resource constraints of AI infrastructure are not a future concern. They are a current crisis in formation, and the numbers that describe them are significant enough to demand board-level attention. The International Energy Agency projects that global data centre electricity consumption will more than double to approximately 945 terawatt-hours by 2030, with AI as the primary driver of that growth. That trajectory represents a doubling of the digital economy’s electricity bill in less than four years, concentrated in the computing infrastructure required to train, run, and deploy AI systems at scale.

The resource constraints that trajectory imposes are not only about electricity. Conventional data centres consume enormous quantities of freshwater for cooling — in many regions they compete directly with agriculture, urban supply, and ecosystem needs. They require significant land in areas where land is expensive and contested. They impose grid demand in regions where baseload capacity is already strained. And they produce heat that must be managed, a requirement that itself consumes a significant fraction of the facility’s total energy budget. Tsinghua University Professor Li Zhen has noted that conventional facilities typically spend approximately one-third of their electricity on cooling systems — at the projected scale of 2030 AI infrastructure, that cooling cost alone represents hundreds of terawatt-hours of electricity annually.

Microsoft recognised these constraints early enough to test a response. Project Natick deployed a data centre 117 feet below the sea off Scotland’s Orkney Islands in 2018, ran it for two years, and reported the concept was feasible, energy efficient, and produced lower hardware failure rates than equivalent land-based facilities — possibly because the stable, sealed, nitrogen-filled environment reduced the corrosive effects of humidity and oxygen on server components. The project then remained a research demonstration. No commercial deployment followed. The Western technology industry observed, documented, and moved on.

China took the same concept, invested 226 million dollars of state-approved capital, connected it to offshore wind, connected it to China Telecom, and deployed it as commercial operating infrastructure — currently scaling from its 2.3 megawatt launch configuration toward its 24 megawatt ceiling, with expansion plans already indicated beyond this first facility. The difference between those two outcomes is not an engineering difference. It is a governance and industrial policy difference — and it has implications that extend well beyond one facility on the ocean floor off Shanghai.

INSIGHT AND ANALYSIS

The strategic significance of the Lin-gang facility is best understood by working through what it solves simultaneously. Land competition between data centres, housing, agriculture, and conservation is already producing regulatory conflict in multiple jurisdictions. In the United States, communities in Michigan and Virginia have mobilised to block data centre developments. In Europe, municipalities are imposing moratoria on new facilities. In South Africa, the question of land use for data centre infrastructure is beginning to intersect with land reform and agricultural policy in ways that are not yet resolved. An undersea facility uses no land in any of these contested senses.

Freshwater scarcity is a more acute constraint in more places. Conventional cooling towers evaporate vast quantities of water — in water-stressed regions, this has produced direct regulatory conflict. Arizona and Nevada have imposed restrictions on data centre water use. The Middle East and parts of Africa face water constraints severe enough that conventional data centre cooling models are simply not viable at scale without creating resource conflicts that no government can afford to ignore. The Lin-gang facility eliminates freshwater use for cooling entirely by substituting seawater — a substitution that is available to any country with a coastline.

The renewable energy integration is the third constraint the facility addresses simultaneously. Connecting directly to offshore wind through subsea cables, with more than 95 per cent of electricity sourced from renewables, the Lin-gang facility demonstrates that AI infrastructure and renewable energy generation can be co-located in a way that reduces grid dependence, cuts transmission losses, and achieves a power usage effectiveness that the land-based industry treats as aspirational rather than standard. At 61 gigawatt-hours of projected annual electricity savings at full scale, the efficiency gain is not marginal. It is structural.

The gap between Microsoft’s 2018 proof of concept and China’s 2026 operating infrastructure is the gap that defines the AI infrastructure competition. It is not primarily a technology gap. The engineering required to submerge a data centre, protect it from saltwater corrosion, cool it with seawater, and connect it to renewable generation through subsea cables was proven in Scotland in 2018. The gap is one of industrial policy coordination, capital allocation speed, and the willingness to convert experimental results into commercial deployment at scale — the precise dimensions in which China’s model of state-directed infrastructure investment consistently outperforms the distributed, multi-stakeholder, regulatory-intensive approach that Western democracies apply to infrastructure decisions of comparable strategic significance.

IMPLICATIONS

For boards and executives evaluating AI infrastructure strategy, the Lin-gang facility raises questions that go beyond admiration for Chinese engineering. The first is whether the physical resource constraints of AI at scale are adequately represented in your organisation’s AI strategy and capital planning. Most AI strategy documents discuss model capabilities, data governance, and talent. Very few include a serious treatment of where the electricity comes from, how much water the infrastructure requires, what land or access rights it depends on, and what happens to those inputs as AI demand grows at the trajectory the IEA projects. Those constraints will shape AI’s cost structure, regulatory exposure, and geographic viability over the next decade — and the organisations that have modelled them will make better infrastructure decisions than those that have not.

The second question is about the competitive environment for AI infrastructure investment globally. China is not the only country developing undersea or offshore AI infrastructure strategies — SpaceX’s plans for orbital data centres represent a different approach to the same constraint set — but China is the country that has moved fastest from concept to commercial operation, and it has done so in a domain where the resource advantages are genuinely significant. The Morgan Lewis analysis of the emerging AI corridor between the US, Europe, and the Middle East focuses on data centres, energy supply, and the geopolitical dimensions of AI infrastructure investment. The Lin-gang facility adds a fourth geography to that corridor map — and it is one that operates outside the regulatory and governance frameworks that corridor was designed around.

The third question is specifically relevant for African and other coastal developing economies. Africa has more than 30,000 kilometres of coastline, significant offshore wind potential, and a growing AI infrastructure deficit relative to its digital economy ambitions. The Lin-gang facility demonstrates that undersea data centres powered by offshore wind are no longer a concept requiring further research validation. They are commercial operating infrastructure, scaling toward a proven 24 megawatt ceiling with expansion plans already in place. The question for African policymakers and business leaders is whether the continent’s AI infrastructure strategy is being designed around the possibilities that this model makes available — or whether it continues to import the land-intensive, freshwater-dependent, grid-constrained data centre model from contexts that have more of those resources than Africa does.

Questions about maintenance, saltwater corrosion management, long-term costs, and the cumulative environmental impact of heat release on local marine ecosystems remain open. These are not reasons to dismiss the model. They are reasons to evaluate it seriously, with the same rigour that any significant infrastructure investment deserves — which is precisely the evaluation that most AI infrastructure conversations, focused on chips, models, and regulatory frameworks, have not yet applied to the physical layer beneath all of them.

CLOSING TAKEAWAY

The Lin-gang facility is 33 feet below the sea surface, 6.2 miles offshore, connected to China Telecom, and scaling toward its 24 megawatt design capacity. It is not a research project. It is not a pilot. It is commercial operating infrastructure with expansion plans already indicated. The concept it represents — AI infrastructure that uses the ocean as a resource rather than competing with terrestrial communities for land, water, and grid capacity — was proven by a Western company in 2018 and commercialised by a state-directed Chinese programme in 2026.

That eight-year gap between proof of concept and commercial infrastructure is not primarily a story about engineering. It is a story about who treats the physical constraints of AI at scale as a strategic priority requiring immediate response, and who treats them as a future challenge requiring further study. The organisations and governments that close that gap in their own planning will be better positioned than those that discover the constraint has become a crisis after the infrastructure decisions of this decade have already been made.

Johan Steyn is a prominent AI thought leader, speaker, and author with a deep understanding of artificial intelligence’s impact on business and society. He is passionate about ethical AI development and its role in shaping a better future. Find out more about Johan’s work at https://www.aiforbusiness.net