"‘AI Closed-Loop’ Goes Viral During the Holiday! An Overview of the North American Data Center Supply Chain"

```

The essence of artificial intelligence competitions is a race over physical infrastructure. Behind every smooth AI interaction on the screen are tens of thousands of servers running at high speed in data centers, all supported by a multi-trillion-dollar physical industry—data centers—expanding at a staggering pace.

According to Bank of America (BofA) estimates, global capital expenditure on data centers broke $400 billion in 2024, and is expected to reach $506 billion in 2025, of which IT equipment spending is $418 billion, and infrastructure spending is $88 billion. Driven by AI demand, this market is projected to expand at an astonishing CAGR of 23% between 2024 and 2028, eventually becoming a market worth over $900 billion by 2028.

So, amidst this unprecedented construction boom, where does the real value chain lie? Who will be the greatest beneficiaries?

This article will delve into the panorama of the data center market ignited by AI, reveal the core logic of technological changes, systematically break down its complex supply chain, fully present a panoramic view of the North American data center supply chain, and identify the true "shovel sellers" in this gold rush.

I. The $500 Billion Market Panorama

The growth of the data center market is no longer driven by traditional self-built and self-used enterprise data centers. Since 2017, when the total capacity of cloud and colocation providers first surpassed enterprise data centers, almost all new capacity has come from two types of players: hyperscale cloud service providers represented by Amazon AWS and Microsoft Azure, and colocation companies that provide leasing services for them or other clients.

Globally, the Americas account for more than half of the world's power capacity, with Northern Virginia on the US East Coast, holding nearly 15% of global hyperscale data center capacity, becoming the undisputed largest single aggregation site in the world. Next is Beijing, China, with about 7%.

The continuous influx of capital is fueled by the clear and considerable return model of data centers as high-value infrastructure assets. Taking a typical new wholesale colocation project as an example, its unit investment economics are as follows:

Initial investment: To build a data center with 1 MW capacity, land and power access upfront costs about $2 million, while the "powered shell"—including construction, MEP (mechanical and electrical), and cooling—costs about $11 million. Total investment per MW is about $13 million.
Revenue and profitability: Each MW can generate $2-3 million per year in rental income. After deducting power (average US industrial electricity price is $0.08/kWh), manpower (about 2 FTEs per MW), property tax (about 1% of property value), and other operating costs, EBITDA margins usually reach a solid 40–50%.
Investment return: Based on a typical 20-year holding period and combining project financing (assuming 46% loan-to-value, 6% debt interest rate, 10% equity cost), the project's internal rate of return (IRR) can reach 11.0%. This is extremely attractive for infrastructure investors seeking long-term, stable cash flow.

This high-certainty business model forms the financial cornerstone for the expansion of the entire data center industry.

II. Arrival of the Technological Singularity: the “Density Revolution” from Chips to Racks

The starting point of all infrastructure changes inside the data center comes from AI chips. Its core evolution can be precisely summarized as a “density revolution”.

The root of the revolution is the exponential surge in single-chip power consumption. From NVIDIA's first-generation Volta architecture to today's Blackwell, the power of a single GPU has increased fourfold in just a few years. The physical law behind this is simple and brutal: Integrating more transistors and running at higher clock speeds necessarily leads to a linear increase in power draw.

The direct chain reaction is a sharp rise in server rack power density. In AI training clusters, network latency is the fatal bottleneck. To minimize data transmission delay between GPUs, the solution is to pack as many GPUs as possible into a single server rack, communicating via high-speed interconnects (such as NVLink). The inevitable result of this architectural optimization is an explosive increase in rack power density. In 2021, average data center rack density was less than 10 kW; today, a standard NVIDIA Hopper (H200) rack draws 35kW, while the latest Blackwell (B200) can reach as high as 120kW. According to NVIDIA's roadmap, the Rubin Ultra platform planned for late 2027 will have a single-rack power draw of an unprecedented 600kW. AMD's MI350, future MI400, and Intel's Gaudi series are following a similar trend.

Meanwhile, the vast majority of global existing data center infrastructure lags seriously behind. According to a 2024 survey by Uptime Institute, only 5% of existing global data centers have average rack density >30kW. This means that 95% of centers can’t even support NVIDIA’s previous generation Hopper chips, let alone higher-power Blackwell. So, deploying AI computing power depends on a massive upgrade/retrofitting of existing data centers and huge volumes of new builds.

It's worth emphasizing this isn't just about GPUs. Google (TPU), Microsoft (Maia 100), and Amazon (Trainium) have also announced their latest ASIC chips must use liquid cooling to pursue maximum performance. This highlights that the cooling challenge brought by high-density computing is likely an irreversible trend across the AI hardware industry.

III. Reshaping Infrastructure: A Revolution Over "Water and Power"

The "density revolution" sparked by chips is driving disruptive changes from the bottom up in data center infrastructure, with the main battlefields being: cooling systems ("water") and power supply systems ("electricity").

(1) Battlefield One: Cooling—Transition from Air to Liquid Cooling

Traditional data centers have long relied on air for cooling. But even the most optimized air-cooling system is physically limited to around 60–70kW per rack. For AI racks with 100kW or more, air cooling is powerless. Liquid cooling—a technology with a brief stint in the mainframe era—has returned to the mainstream.

Of the various liquid cooling paths, the industry’s current mainstream is direct-to-chip (D2C) liquid cooling.

This tech uses a metal "cold plate" with microchannels in direct contact with heat-intensive chips like GPUs and CPUs, with coolant (water and glycol mix) efficiently carrying away heat. The system core is the Coolant Distribution Unit (CDU), which drives heat exchange/circulation between a "secondary loop" inside the server and a "primary loop" outside the data center.

The CDU market is only about $1.2 billion in 2024, but growing explosively.

There are more than 30 suppliers now. However, the extreme demand for "uptime" among data center operators makes them highly conservative, and they prefer mature and reliable suppliers. Thus, established brands with comprehensive product lines and global service networks have a natural moat. Vertiv (strengthened via the 2023 CoolTera acquisition), Schneider Electric (moved in with the 2025 Motivair acquisition), Delta Electronics, and nVent are seen as first-tier leaders.

(2) Battlefield Two: Power Supply—Revolution from AC to High-Voltage DC

Traditional data center power paths are long and inefficient: medium/high-voltage AC from the grid is stepped down, distributed via switchgear, enters UPS (uninterruptible power supply) with AC-DC-AC conversion for backup, then routed to racks via PDUs or busways, and finally into server PSUs for the last AC-DC conversion.

As AI racks head toward 100kW+, the drawbacks of low-voltage AC are clear: huge currents require thick copper cables, which are expensive, space-consuming, and restrict cooling. Thus, a shift to high-voltage direct current (HVDC) is underway.

Microsoft, Meta, etc., proposed 400V DC solutions in the Open Compute Project.
NVIDIA announced its own 800V DC plan for future megawatt-scale racks, targeted for deployment in 2027.

The key advantage: since power = voltage x current, boosting voltage by an order of magnitude lets current drop by the same, enabling much thinner, cheaper wiring. This greatly reduces expensive and heavy copper in racks. Schneider Electric estimates 400V DC can reduce copper wire weight by 52% vs. conventional 208V AC systems.

This reshapes the power architecture:

Simplified UPS: DC eradicates the need for an inverter to turn battery DC into AC within the UPS, which can theoretically reduce cost by 10–20% (though this may be offset initially by high-voltage safety gear).
PSU relocation: PSUs, which consume space inside servers, will move out to become standalone “power sidecars”, freeing space for compute units.

NVIDIA has clearly stated its 800V DC plan will be co-developed with Vertiv, Eaton, and other industry leaders, again verifying that incumbents maintain a crucial role in industry standard transitions.

IV. Supply Chain Breakdown: Who Are the “Shovel Sellers” in the Gold Rush?

The rise of AI is significantly raising per-unit construction costs of data centers. A traditional data center’s all-in cost is about $39 million per MW, while a next-gen AI-architecture data center (chip-level liquid cooling + HVDC) jumps 33% to $52 million per MW. Most of the increase is due to pricier AI servers, but infrastructure upgrades drive significant cost as well.

Across this large, intricate supply chain, “shovel sellers” in each segment are sharing in the boom:

Cooling Systems (Thermal): A market of around $10 billion (2024); Vertiv is the recognized leader. Core products: chillers, cooling towers, precision air conditioners (CRAHs), etc. Johnson Controls (via Silent-Aire acquisition), Carrier, and Trane are also prominent HVAC giants.

Power Systems (Electrical): A market of about $18 billion (2024); Schneider Electric leads. Portfolio covers UPS (install market ~$7 billion), switchgear (~$5–5.5 billion), busways, and other distribution gear (~$4.2–4.7 billion). Eaton, ABB, Siemens are key industrial electrical firms.

Backup Power: Diesel generators are the last line of maximum reliability. The generator equipment market is about $7.2 billion in 2024, with Cummins as the global leader.

IT Equipment: The largest capital segment. In 2024, the global server market is about $280 billion, with AI servers now representing half the value. Network gear is about $36 billion, mainly dominated by Cisco and Arista.

Engineering and Construction: Turning blueprints into reality requires specialist engineering and construction. Engineering design (4.5–6.5% of infrastructure cost) is a ~$4 billion market, with major players like Jacobs Solutions and Fluor. Construction is even larger (~$65–80 billion, including large material and equipment pass-throughs), with participants including Balfour Beatty, Skanska and other international giants.

Conclusion

The "AI flywheels" and dazzling generative capabilities we see on screen are not only leaps in AI tech, but also the results of a physical infrastructure arms race involving concrete, copper, and coolant.

An astonishing fact: in the coming months, global data center construction spending will, for the first time in history, exceed total commercial office building construction spending.

In this unprecedented AI-fueled gold rush, the supply chain giants who master core cooling and power supply tech—able to "cool" and "feed" this vast compute engine—will inarguably become the silent but true winners of the era.

This article is from WeChat public account “Hard AI”. For more AI insights, click here.

Risk Warning and DisclaimerThe market has risks; investment requires caution. This article does not constitute personal investment advice, nor does it take into account individual users’ specific investment objectives, financial situation, or needs. Users should consider whether any opinions, points of view, or conclusions in this article are suitable for their circumstances. Invest accordingly at your own risk. ```