dc-thermal
Menu菜单
← All writing← 全部文章

Warmer Water Rewrites the Plant: What the CDU and Chiller Data Say更高的水温改写冷站：CDU 和冷水机组的数据说了什么

Willow LeiIndustry insight行业洞察

In my last essay I argued that the bigger shift reaches beyond the rack into the central plant. Since then I have read the specification sheets of the main coolant distribution units (CDUs) and data-center chillers, line by line, rating point by rating point. The data support the argument, but they also show where it gets hard.

My answer in one sentence: warm water squeezes the CDU and rewards the chiller, and together they push the plant from a chiller plant toward a dry-cooler plant with chillers as trim.

The starting point: 41 °C to the CDU, 45 °C to the rack

NVIDIA now designs its MGX racks for 45 °C warm-water inlet. In its example, the plant sends 41 °C facility water to the CDUs, which supply 45 °C coolant to the racks. NVIDIA's own reason is the plant: 45 °C "enables data centers in many climates to use ambient air and closed loop dry coolers for cooling, reducing the need for compressors", and it says the saved power allows up to 10% more NVL72 racks in the same power budget. Its schematic still shows a chiller, bypassed in that free-cooling scenario.

For an HVAC engineer, 41 → 45 °C means one thing first: the CDU has about 4 K to work with.

The CDU: warm water squeezes it

A CDU is a heat exchanger with pumps. Its rating depends on the approach — the difference between the facility water it receives and the coolant it delivers. In my comparison table the liquid-to-liquid units state approaches from 2 K (a vendor minimum, "as low as") to 10 K. A unit rated above 4 K has to be reselected at the 41/45 °C duty point.

Two vendors show the cost. Delta rates its GoCool-1500 at 1,500 kW at 6 K, 1,250 kW at 5 K and 1,016 kW at 4 K (with 25% propylene glycol on the rack side and roughly 1,500 L/min of coolant flow). Schneider's MCDU-70 brochure gives 5,040 kW at 32.2 °C facility water with 45/65 °C coolant, and 2,500 kW at 41 °C facility water with 45 °C coolant. The brochure does not state the coolant return temperature at the 2,500 kW point, so these figures cannot isolate the effect of facility-water temperature — but its "2.5 MW" headline is the warm-water point.

So my first judgement: CDUs must be selected at the warm-water duty point, not at the catalogue headline. Flow becomes a sizing variable — Motivair calls 1.5 L/min per kW "the industry target" — and pumps, pipes and heat-exchanger area grow with it. At the same time, announced single-unit capacities have grown from 2.3 MW (Vertiv, November 2024) to between 2.5 MW and "up to 3.5 MW" in 2026 (Schneider's WCDU, with first shipments planned for selected regions in October 2026), each at its own rating basis. Above roughly 2.3 MW, by my rule, a CDU leaves the row and becomes room-level equipment.

The chiller: warm water rewards it

On the chiller side the same temperature is good news. Several data-center chiller lines now publish warm-water rating points alongside the conventional 7/12 °C (supply/return): Vertiv quotes 18.9/27.8 °C for the Liebert AFC and 20/26 °C for the Liebert OFC (both marked preliminary), Airedale 20/30 °C for the TurboChill DCS at 35 °C outdoor air, and Carrier 20/30 °C for the AquaEdge 19MV4 with a 49/55 °C condenser loop.

The gain is visible in one STULZ brochure. The same unit gives 753 kW at 12/18 °C water (supply/return) and 35 °C outdoor air, but 1,000 kW at 20/26 °C and 33 °C outdoor air, and its EER rises from 3.91 to 4.64. The outdoor air also differs a little, so not all of that third is water temperature, but the direction is clear.

And one more number: among the chiller lines I compared, none states a supply temperature of 41 °C. The highest is Vertiv's 40 °C, and Vertiv calls that product a "Trim Cooler". At NVIDIA's example temperatures, suitably sized dry coolers can carry the liquid-cooling load without compressors when the outdoor conditions allow.

What it means for the plant

My judgement for the next three years: in new AI halls built for 45 °C liquid cooling, the plant is sized around dry coolers and free cooling, and the chiller becomes a trim machine for hot hours, for the remaining air load and for heat recovery. Chiller capacity per MW of IT falls; the value moves to dry coolers, heat exchangers, pumps and controls. The CDU becomes the hinge of the system: its approach decides how warm the plant can run.

The condition is that the whole chain is designed for the warm point from the start — CDU rated at 41 °C water, piping sized for flow, chillers chosen as trim. Where a site is designed for 7 °C, the classic plant stays.

The industry is already reorganising. Since late 2023, six transactions in my timeline have brought CDU suppliers or thermal businesses under larger HVAC, power and water groups (including Schneider's 75% stake in Motivair and Eaton's purchase of Boyd's thermal business), and chiller makers such as Trane, Modine and Daikin are investing in factories and test labs for data-center lines. Refrigerant rules add pressure: from 1 January 2027 the EU stops new stationary chillers above 12 kW that contain fluorinated gases with a GWP of 750 or more from being placed on the market (with a safety exception), and the US EPA limits refrigerant GWP to below 700 for newly manufactured comfort-cooling chillers from 2025 and for its data-center and IT cooling subsector from 2027.

Where I could be wrong

First, chips may not keep allowing warm water. In one ASHRAE case for a dense server, moving from W27 to W45 water brought the cold plate's cooling capability down to roughly what air cooling achieves with a low cold-aisle temperature. If the next chips need colder coolant, the plant swings back toward chillers.

Second, climate. A dry cooler cannot make 41 °C water on a 45 °C day without help. In hot regions the "trim" may run for many hours, and adiabatic or evaporative assistance brings water use back into the design.

Third, chillers may catch up. A major step in chiller technology could let chillers produce 41 °C water efficiently. If that happens, the chiller could stay in the base load instead of becoming the trim, and the shift toward dry coolers would be smaller than I expect.

I will review this view at the end of 2027.

Personal view. Public sources only; numbers keep their stated rating conditions. No employer view is implied.

Sources

Sources checked 2026-10-13.

上一篇文章里，我认为更大的变化会越过机柜，延伸到中央冷站。此后我把主要的冷却液分配单元（CDU）和数据中心冷水机组的样本逐行、逐个额定工况读了一遍。数据支持这个判断，也让我看到难点在哪里。

一句话的回答：暖水挤压 CDU、奖励冷水机组，两者合起来把冷站从“冷机为主”推向“干冷器为主、冷机补充调峰”。

起点：进 CDU 41 °C，进机柜 45 °C

NVIDIA 现在按 45 °C 暖水进水设计 MGX 机柜。在它的示例中，冷站给 CDU 送 41 °C 的设施水，CDU 再给机柜供 45 °C 的冷却液。NVIDIA 给出的理由正是冷站：45 °C 运行“让很多气候区可以用室外空气和闭式干冷器散热，减少对压缩机的需求”，并称省下的电力可以在同样的电力预算里多放最多 10% 的 NVL72 机柜。它的示意图里仍然有冷水机组，只是在这个自然冷却场景中被旁通。

对暖通工程师来说，41 → 45 °C 首先意味着一件事：CDU 只剩约 4 K 的温差可用。

CDU：暖水在挤压它

CDU 是带水泵的换热器，它的额定能力取决于接近温差，也就是进来的设施水和送出的冷却液之间的温差。在我的对比表中，液-液型机组标称的接近温差从 2 K（厂商最小值，“低至”）到 10 K 不等。按 4 K 以上标定的机组，需要在 41/45 °C 工况下重新选型。

有两家厂商给出了代价。台达 GoCool-1500 在 6 K 时 1,500 kW，5 K 时 1,250 kW，4 K 时 1,016 kW（机柜侧为 25% 丙二醇，冷却液流量约 1,500 L/min）。施耐德 MCDU-70 的样本给出：设施水 32.2 °C、冷却液 45/65 °C 时 5,040 kW；设施水 41 °C、冷却液 45 °C 时 2,500 kW。样本没有给出 2,500 kW 工况的冷却液回液温度，因此不能用这两个数字单独判断设施水温的影响——但它宣传的“2.5 MW”就是暖水工况下的数字。

所以我的第一个判断：CDU 必须按暖水工况选型，而不是按样本标题上的数字。流量成为选型变量——Motivair 把每 kW 1.5 L/min 称为“行业目标”——水泵、管路和换热面积都随之变大。与此同时，厂商公布的单台容量从 2.3 MW（维谛，2024 年 11 月）增长到 2026 年的 2.5 MW 至“最高 3.5 MW”（施耐德 WCDU，计划 2026 年 10 月在部分地区首批发货），各自的额定依据不同。按我的规则，超过约 2.3 MW 的 CDU 会离开机柜列，变成房间级设备。

冷水机组：暖水在奖励它

在冷水机组这一侧，同样的温度是好消息。部分数据中心冷水机组系列已经在传统的 7/12 °C（供/回水）之外公布暖水额定工况：维谛 Liebert AFC 为 18.9/27.8 °C，Liebert OFC 为 20/26 °C（两者都标注为初步数据）；Airedale TurboChill DCS 为 20/30 °C、室外 35 °C；开利 AquaEdge 19MV4 为 20/30 °C，配 49/55 °C 冷凝回路。

收益在一份 STULZ 样本里看得很清楚。同一台机组在 12/18 °C 水温（供/回水）、35 °C 室外时是 753 kW，在 20/26 °C、33 °C 室外时是 1,000 kW，EER 从 3.91 提高到 4.64。室外温度也略有不同，所以这三分之一的增量并不全来自水温，但方向很清楚。

还有一个数字：在我比较的冷水机组系列中，没有一个标称能供 41 °C 的水。最高的是维谛的 40 °C，而维谛把这个产品直接叫作 “Trim Cooler”（补冷机）。在 NVIDIA 示例的温度下，只要室外条件允许，配置合适的干冷器就能在不开压缩机的情况下承担液冷负荷。

这对冷站意味着什么

我对未来三年的判断：在按 45 °C 液冷新建的 AI 机房里，冷站会围绕干冷器和自然冷却来配置，冷水机组变成补充调峰的设备，负责高温时段、剩余的风冷负荷和热回收。每 MW IT 所需的冷机容量下降，价值转向干冷器、换热器、水泵和控制系统。CDU 成为整个系统的关节：它的接近温差决定了冷站能跑多高的水温。

条件是整条链从一开始就按暖水设计——CDU 按 41 °C 设施水标定，管路按流量配置，冷机按补冷选型。按 7 °C 设计的项目，传统冷站仍会保留。

行业已经在重组。2023 年底以来，我的时间线里有六笔交易把 CDU 供应商或热管理业务并入了更大的暖通、电力和水处理集团（包括施耐德取得 Motivair 75% 股权、伊顿收购 Boyd 的热管理业务）；特灵、Modine、大金等冷水机组厂商也在为数据中心产品线投资工厂和测试实验室。制冷剂法规也在施压：从 2027 年 1 月 1 日起，欧盟不再允许含 GWP 750 及以上含氟气体、额定容量 12 kW 以上的新固定式冷水机组投放市场（满足安全要求所必需的除外）；美国 EPA 要求新制造的舒适性冷水机组自 2025 年起、数据中心与 IT 冷却细分领域自 2027 年起使用 GWP 低于 700 的制冷剂。

我可能错在哪里

第一，芯片未必会一直允许暖水。在 ASHRAE 针对一款高密度服务器的案例中，从 W27 提高到 W45 水温，冷板的冷却能力降到接近冷通道温度较低时的风冷水平。如果下一代芯片需要更冷的冷却液，冷站就会重新偏向冷水机组。

第二，气候。在 45 °C 的天气里，干冷器不借助其他手段造不出 41 °C 的水。在炎热地区，“补冷”可能要运行很多小时，而绝热或蒸发辅助又会把用水问题带回设计中。

第三，冷水机组可能追上来。如果冷水机组技术有重大进步，能够高效地制备 41 °C 的水，冷水机组就可能继续承担基础负荷，而不是退为补冷，冷站向干冷器转移的幅度也会比我预期的小。

我会在 2027 年底回看这个判断。

个人观点。只使用公开来源；数字保留其标称工况。不代表任何雇主的观点。

来源

来源核对日期：2026-10-13。

All writing →全部文章 →