On the floor at Data Centre World Asia: the whole industry hinges on power, and behind the crowds sit three fundamental shifts

This time last year, Computing Power Heart was also at Marina Bay Sands in Singapore.

Returning to the same venue and the same show this year, the most immediate physical sensation comes down to one word: packed.

Aisles were jammed to a standstill, booths were mobbed, and even catching an industry talk meant showing up early to fight for a seat.

The official numbers match what we felt on the ground. Attendance at this year's Data Centre World Asia jumped 25% from last year, with more than 600 exhibitors. Last year's floor plan was nowhere near this scale; this year the show filled all four levels of the Sands Expo and Convention Centre. From B2 in the basement to Level 1, Level 3, and Level 5 at the top, every floor was crammed.

Caption: The morning of day one, the corridor outside the exhibition halls at Sands Expo

If you were not there, this piece aims to answer one question for you: where does this industry actually stand right now?

The answer from the giants and buyers on the floor was remarkably uniform. The entire industry hinges on power.

More telling than the bigger crowds, though, are the three fundamental shifts behind them.

The Topic Shift: From "Where to Build" to "How to Deliver"

At last year's show, everyone was talking grand strategy, comparing Batam in Indonesia, Johor in Malaysia, and Jakarta, weighing where to secure land and where to break ground.

Walk in this year and the wind has changed completely.

Nobody is pitching visions of the future anymore. Everyone is pinned on three things: is there enough power, can the heat be pulled out, and how many months until delivery.

In today's computing power world, you can only start building a qualified AI data center once you have locked in a huge block of electricity. And only once you can show compliant high-density power and a physical facility delivery plan will NVIDIA approve your order and ship you its latest B300 chips.

Without power, every grand plan fails to even get past NVIDIA's door.

Caption: A talk on GPU power density and cooling. The aisles were standing room only, with the audience holding up phones in unison to photograph the slides.

That explains the crush around the joint Delta and NVIDIA booth, several rows deep. On display was a power foundation built specifically for the Vera Rubin platform: 800V DC power distribution, MGX server racks, and a 90kW DC-DC power shelf.

In plain terms, servers now have appetites so large that rack power draw has climbed to figures nobody would have dared imagine before.

Caption: The Delta × NVIDIA joint booth, showing 800V DC power and full MGX racks for the Vera Rubin platform

Power scarcity and the fear of outages have produced an unusual sight.

Guangxi Yuchai hauled a diesel generator set taller than a person straight into the Sands exhibition hall and parked it right beside the aisle. Two years ago, if someone told you the biggest draw at a high-tech data center show would be a heavy-duty diesel generator, you would have assumed it was a joke. Today, backup power has become customers' most urgent source of reassurance.

Caption: Guangxi Yuchai's diesel generator set in the flesh. Power anxiety now comes down to the specifics of backup supply.

Once the power is in, heat becomes the next deadline, which is why liquid cooling was effectively standard issue on the floor:

Rittal split its CDUs, the units that circulate liquid to cool servers, into row-level and rack-level versions and placed them front and center. Machines from GRC and SK enmove were submerged in specialized dielectric fluid, powered up and running at full speed on site. Supermicro took a full wall to push its complete packaged liquid-cooled racks.

Caption: Rittal split its CDUs into row-level and rack-level versions, placed in the booth's prime spot

Caption: Immersion-cooled systems from GRC and SK enmove, powered up and running

If cooling is a contest of engineering depth, the blunter and more brutal fight on the floor was over speed.

Building halls and pulling power the old way is simply too slow, and the big players who can afford the chips cannot wait. That anxiety set off a floor-wide scale-model arms race. Last year, Fourier was the only exhibitor at the entire show to bring a miniature diorama of a full modular data center. This year, there were at least a dozen or two containerized facility models and panoramic dioramas competing for attention, everyone racing to prove their prefabricated delivery chops.

Huawei put a scale model of its FusionPower9000 power module on a table, with prefabrication and fast delivery as the core pitch. LG's diorama went further, simulating the entire process of a crane stacking container modules into an AI factory. One exhibitor simply painted a line across its booth header: "Faster to Market, Faster to Revenue." The sooner you go live, the sooner the money comes back. One sentence, and the nature of this business is laid bare.

Caption: Fourier's booth, with a diorama of a modular data center cluster

Fourier, the first to bring a diorama last year, is betting precisely on that time arithmetic.

The company spent its early years honing hard engineering in specialized settings with extremely demanding thermal requirements, and has since carried that over to AI data centers and their startling heat loads. From cold plate and immersion cooling to high-density modular designs for 10MW-class clusters, what they do is simple: help customers buy time.

They have turned civil-works facility projects that once took years into standardized building blocks that can be delivered and racked in as little as six months. On top of pushing PUE down to 1.15, they can get water and power flowing faster than anyone. On the strength of that fast-delivery playbook, their modular facilities now run in more than 20 countries, from the deserts of Dubai to Texas. In Taipei this year they also showed, alongside Intel, a 20-foot modular facility you can walk right into.

When computing power turns over this fast and depreciation is counted by the day, every month earlier a customer takes delivery is another month of revenue from that compute.

The Player Shift: Four Floors, One Complete Food Chain

The second change on the floor was how sharply the exhibitors divided up the work.

The show filled four floors, and walking from the bottom up you are not seeing a random collection of stalls but a layered industrial pyramid.

Level B2 is the pyramid's base, all critical component suppliers. Grundfos drew a crowd around its industrial cooling pumps, and the valves, cables, piping, and sensors here are the basic parts of the whole AI world.

Caption: Level B2: Grundfos's cooling pump booth, drawing traffic on par with the full-system vendors

Level 1 is the pyramid's middle: vendors of integrated power, water, and thermal management. Siemens, Carrier, Mitsubishi Heavy Industries, Vertiv, and Schneider were all packed onto this floor. They do not sell a single part but whole-system solutions, from the grid connection at the perimeter down to cooling the chip.

But after a lap of this floor, one thing becomes clear: suppliers of components and of power and cooling solutions are everywhere, while real customers who actually hold ample power resources and are building or preparing to build AI data centers are a small minority. Hundreds of solution providers are elbowing each other for that tiny pool of power-rich buyers.

Caption: Level 1: Siemens's booth, pitching integrated data center fire protection and power distribution

Level 5 is the apex: the computing power and full-rack giants. GIGABYTE's display board read "NVIDIA Vera Rubin NVL72" in plain sight, and NVIDIA's GPU compute trays sat locked in glass cases, with onlookers packed in to photograph them. Players on this floor sit closest to the customer's wallet, because what downstream buyers want most directly is a compute rack that can run models immediately.

Caption: Level 5: NVIDIA GPU trays in the flesh, with onlookers several rows deep

Caption: GIGABYTE's display board reading "NVIDIA Vera Rubin NVL72"

In the cloud compute section on Level 5 stand the cycle's biggest winners.

Take Bitdeer. Many people still think of it as a mining rig company, but walk into its booth and the signs at the entrance read "AI Cloud" and "NVIDIA Preferred Partner." The company now sits firmly at the top of the industry pyramid: on one side, using its strong energy position to buy up top-tier machines like the B300 and build its own AI data centers at scale; on the other, closing a highly effective monetization loop, selling some facilities and full compute systems as packages to large customers while slicing up the rest to sell tokens and cloud compute at the front end.

Last year, the players at the apex were not present in such complete formation.

This year, from components at the base, to power distribution and liquid cooling in the middle, to the top-tier giants building AI data centers, stockpiling chips, and distributing tokens and compute, a complete but highly concentrated AI infrastructure supply chain was laid out in concrete form for the first time.

Caption: Bitdeer's booth: from mining rigs to AI cloud, with "NVIDIA Preferred Partner" hung at the entrance

The Buyer Shift: From Browsing to Arriving With a Budget

The third and most tangible change is in the people writing the checks.

Last year's show had plenty of curious onlookers just trying to figure out what an AI data center even is. This year the halls were full of Southeast Asian faces, with English in Indonesian, Malaysian, Thai, and Vietnamese accents bouncing between booths.

They crowded around engineers with blunt, practical questions: how many weeks to get a full set delivered, what the minimum order is, whether local installation is supported, and whether there is a local technical team on call when something fails.

Several exhibitors said privately that the brochures and giveaways they had prepared for two days were cleaned out by the afternoon of day one. The people who came were not window shopping. Almost all of them had budgets in hand and specific engineering requirements, looking to lock in schedules.

Caption: SuperX's "modular AI data center" booth, with a line at the inquiry desk from morning to night

Special Roundup: China's Presence on the Floor

One last detail many people care about: Chinese companies were impossible to overlook at this show. Computing Power Heart confirmed each Chinese exhibitor on site and organized them into a list, arranged by the pyramid's three layers for easy reference (partial list, in no particular order):

Apex, computing power and full systems: Huawei, xFusion, ZTE, YMTC, openEuler, DapuStor, KAYTUS (part of Inspur Information), Naddod, Yunbo Online, TuringData, China Mobile International, Zelu Security, and Haiqing Zhiyuan.

Middle, power and thermal management: Guangxi Yuchai, Shenling Environment, Midea Building Technologies, KSTAR, MEGMEET, CIMC Digital Power, Haiwooo, ENFORT, EVE Energy, Shuangdeng Group, and Trina Solar.

Base, components and materials: CHNT, DAQO, Sieyuan Electric, KINGFIT, Shanghai Yuhu Stainless Steel, Taixin Zhileng, YOEC, Anmei, and Kejingyuan.

The Taiwanese contingent was no smaller: Delta, GIGABYTE, and MiTAC, among others, each holding key positions in power, full systems, and servers.

From diesel generator sets to GPU racks, nearly every link in this supply chain had a Chinese booth. If we missed any exhibitors, let us know in the comments and Computing Power Heart will stop by each booth at the next show.

The Industry Has Moved From Storytelling to Delivery

After a full lap of the show, it all comes down to one line: this industry has left the storytelling phase behind for good and is now in a hard fight over delivery.

That the conversation has moved off macro site selection means projects are past the paper stage. That four floors laid out a tightly linked upstream and downstream chain means the industrial supply base is fully mature. That customers are arriving with budgets and pressing on schedules means real money has started moving across the table.

On one rack panel in the hall sits a signature Jensen Huang left this year: "Spark of AI, June 3, 2026."

Caption: Jensen Huang's signature on a rack panel: "Spark of AI, June 3, 2026"

For all the AI euphoria, the brightest sparks still land on these very plain hardware realities.

Whether the power can be brought in, whether the heat can be pushed out, and whether the custom racks can be racked and energized on schedule.

According to show organizers speaking privately, many exhibitors were already scrambling to lock in prime booths for next year before this year's show had even ended. Move too slowly and the good spots are gone. Next year's show, they say, will only be bigger than this one.

The logic behind it is unshakeable. As long as qualified AI data centers are still being built against the clock, B300 supply stays this tight, and power remains scarce, this AI fire will not cool. It will only burn hotter.

This time next year, those four floors at Sands will likely be even more packed.