At a trade show in Kuala Lumpur, most of the audience were electrical engineers and plumbing and HVAC engineers. What they were discussing and what Jensen Huang was saying on a Berlin stage at the same moment belong to almost separate worlds. Over there, chips need to be faster, product cycles shorter, data centers thrown up in fourteen days if possible. Over here, substation approvals take half a year, reclaimed water treatment plants need daily water-quality testing, and a few-hundred-megawatt campus has to borrow people and pipelines from the oil and gas industry just to find a test load.

The most advanced virtual intelligence is being held back by the physics of the most traditional heavy industry. GPUs can turn over a generation in a year; grid planning runs on ten-year cycles. A bad line of code rolls back in seconds, but a pipe poured in the wrong place or a mis-ordered transformer can leave a campus worth more than $100 million idle for six months.

Contract Signed for Liquid Cooling, Switched to Air Cooling Mid-Build

Tan Phei Chieh, Deputy Director of Customer Solutions at DayOne, described one case on stage. A client had signed a contract for an all-liquid-cooled data hall, and some of the piping and equipment had already been welded into place. Halfway through construction, the client changed its server procurement plan and told the site: make it all air cooling.

Tearing it out and starting over was impossible; nobody could absorb the schedule hit or the penalties. The engineering team had to put on hard hats and rework the finished structure in place, bolting on large numbers of terminal fans and rewriting the control code for the cooling towers and chillers.

That kind of awkwardness is not rare in the AI era. Clients tend to stake out ground early, locking in power quotas of tens of megawatts, yet right up to the day the shell is topped out they still haven't settled on which generation of chip they will deploy, or whether it will be air cooling or liquid cooling. The pressure all lands on the design side.

Some design teams hedge. Even when building for air cooling up front, they pre-install the piping connections; or they simply build air cooling and liquid cooling as two independent loops. Nobody wants to pay the extra cost, but next to demolishing and rebuilding, it is the cheap path.

The Details That Get Stuck in Physical Space

Beyond system design, an even stranger set of problems comes down to space.

Dimensions are confidential. On some top-end GPU racks, the position of the liquid cooling connections is a core trade secret, and construction drawings arrive partially obscured. If a prefabricated pipe comes out a few centimeters too long, it won't connect and forcing it creates stress damage; too short and it won't reach at all.

Turning radius is another hidden barrier. Some of these massive AI racks weigh hundreds of kilograms and need special industrial steering equipment to move, with a full three meters of turning radius reserved. If the data hall doors, a corridor corner, or the loading dock come up even half a meter short, an expensive rack ends up stuck at the doorway.

Noise can even push costs up. As data halls get bigger, the external cooling fans get bigger with them. Experts from BAC (Baltimore Aircoil) noted that to keep exhaust noise low enough to avoid complaints from nearby residents, the only option is larger fans turning slowly. Bigger fans mean the footprint of the heat rejection equipment balloons, and construction costs climb along with it.

None of this shows up on the drawings; it surfaces on site. None of it involves any frontier technology, but every item on the list can set the schedule back.

Past 300 Megawatts, You Can't Even Borrow a Load Bank

The most nerve-wracking stage is commissioning in the field.

In the past, once a data hall was built, engineers would wheel in a few load banks, essentially oversized industrial heating coils, to simulate server heat and power draw, run the whole electrical system and cooling through a test, and call it done. But Tracey Jumper pointed out that once a single phase of an AI campus pushes past 300MW, that approach breaks down entirely: there simply aren't enough high-capacity test banks on the market to rent.

Teams end up playing their own systems against each other. Cooling plant A generates the test load for plant B, then switches over to plant C, loading the project's existing systems against one another to verify chiller capacity and backup capability. To build out the sprawling temporary pipework this requires, they have even brought in the crews that weld long-distance pipelines for the oil and gas industry.

Sensor calibration is another hurdle. With four or five thousand sensors and instruments in a data hall, running all of them through the traditional repeat calibration procedure could take longer than the entire project schedule. Teams have had to bring AI scripts and automation onto the site, compressing engineering shift planning and test progress into an hour-by-hour sprint.

That also explains why Jensen Huang is pushing so hard on factory-prefabricated approaches like the DSX POD: every extra connection made on site is one more chance for something to go wrong. But however good prefabrication gets, it only solves the rack-level slice. The substation outside the fence, the water pipes buried underground, the turning radius in the data hall corridor, robot dogs and large models can't do that work for you.

Why 480V Has to Go

If all of the above is civil works and periphery, the closed-door electrical forum held that same afternoon cut straight into the heart of the AI factory. Zhang Di, president of Sungrow's AIDC power business unit, made the case that the power supply architecture has to be torn down and rebuilt.

In July, Zhang launched Sungrow's EnerNeo line of solid-state transformers in his capacity as head of the company's AIDC unit. Three months later, sitting in a Kuala Lumpur conference room, he was facing a tougher crowd: overseas engineers who run the numbers on job sites every day.

He opened with data. In July, CoreWeave published test results for the Vera Rubin NVL72 architecture: running frontier models like DeepSeek-R1, token throughput per megawatt of power was several times, in some cases close to ten times, that of the previous-generation GB200. The implication is that the contest in computing power ultimately comes down to how many tokens each kilowatt-hour can produce.

But rack power has simply gotten too high. Zhang ran the numbers at the forum: supplying a high-power AI rack at the traditional 480V means frightening levels of current. To keep things from burning out, the data hall has to be filled with copper cable and busbar as thick as your thigh, which not only eats up floor space but also incurs painful conversion losses along the way.

Middle-school physics: raise the voltage and the current drops, sharply reducing cable thickness and heat loss. That is the fundamental reason NVIDIA, Microsoft, Google and other giants are pushing 800V DC power distribution through OCP.

What SST Wants to Cut Out Is the Middle

On the road to 800V DC, the hottest technical concept is SST, the solid-state transformer. To understand it, start with how power reaches a server today.

A typical chain looks like this: AC from the high-voltage grid comes in, passes through a conventional line-frequency transformer, then a switchgear cabinet, then a large data hall UPS for battery backup, then row-level distribution cabinets and power distribution units, and only then the server's own power supply module. That long sequence is like stacking three or four adapter plugs onto a phone charger. Every hand-off burns some power, and every device is a hunk of iron with a large footprint.

SST aims to go straight there. Using modern power semiconductors such as silicon carbide (SiC), it converts medium-voltage AC coming off the grid directly into 800V DC in a very compact package. If that path works, it not only delivers higher conversion efficiency but also frees up the large areas of floor space previously taken by UPS systems and switchgear, enough for several more rows of compute racks.

Sungrow launched its EnerNeo SST line in July and has signed framework orders.

Three Hurdles: The Math, the Risk, and the Inertia

During the Q&A, an engineer from a design institute named Daniel put the microphone right in front of Zhang: if SST is so good, why is everyone still running proofs of concept, and when does it actually go into data halls at scale?

Zhang's answer was candid. Everyone is on a first round of field testing now, the next phase of testing is scheduled for Silicon Valley in the first quarter of 2027, and real industry volume probably won't arrive until around 2028. Another engineer followed with the blunt version: today's 800V DC and SST designs are simply too expensive all-in.

Getting from slideware to real volume orders means clearing three mountains.

First, the math doesn't work. A conventional iron-core transformer is mature, rugged and cheap; an SST is full of expensive high-voltage silicon carbide chips and precision control boards. Even if it saves on UPS systems and busbar, once you add the upfront price premium and the cost of replacing chips later, plenty of cost-conscious data center owners run the total and pull their hands back.

Second, the downside is too costly. An AI server that crashes can be restarted; if the power entry point goes down, tens of thousands of chips lose power instantly and losses run into the tens of millions of yuan. A dumb old transformer sits outdoors in all weather and rarely fails in thirty years. Whether precision semiconductor devices can run five or ten years continuously under high voltage and high current without failing is something nobody anywhere has enough long-run, large-scale operating data on yet.

Third, engineering inertia. For most existing data halls, ripping everything out is not an option. The industry's more realistic compromise is to keep the conventional transformer and add a conversion cabinet called a PowerRack next to the rack, stepping the power to 800V DC at that point. This patch-style transition is safer than going all-in on SST, and easier to deploy.

In other words, the direction toward 800V DC power distribution is becoming clear, but whether projects ultimately adopt SST, conventional transformer-rectifier units, or a transitional architecture like the Power Rack still depends on each project's cost, reliability requirements and retrofit conditions. What SST has to win is not just a slot in the technical standards, but the volume orders data center owners are actually willing to fund.

When Delivery Actually Happens

Put the two venues side by side and the gap is obvious. Berlin represents the Silicon Valley pursuit of extremes: the fastest chips, the shortest cycles, data centers up in fourteen days if possible. The electrical and plumbing engineers in Kuala Lumpur are dealing with the physical world as it actually is.

That makes for the strangest feature of the current AI frenzy: the most advanced virtual intelligence is being dragged to a halt by the physics of the most traditional heavy industry.

For an AI factory, the delivery moment that matters has never been when the GPUs leave the factory in their crates. It is the moment power and water are connected, commissioning is complete, and the data hall roars to life for the first time. However fast the chips run, the bricklayers and electricians outside the chips have to finish laying the road first.