An Interview with the Cooling Architect Watching His Data Center Melt in Real Time

We meet Kelvin Koh in the monitoring room of what he describes as "Singapore's most thermally ambitious data center." Through reinforced glass, 847 racks of AI processors disappear into artificial twilight, their status lights blinking like distant stars. Each rack generates more heat than a commercial pizza oven.

It's 11 AM on a Tuesday in December 2025. Outside: 34°C, 87% humidity. Inside: the air conditioning is screaming at what K² calls "emergency capacity," and the monitoring screens show inlet temperatures creeping toward 28°C—one degree above ASHRAE's recommended maximum.¹

K² is 41, with the specific exhaustion of someone who hasn't slept properly in weeks. He's wearing a polo shirt with the facility's logo and carrying what he calls his "panic tablet"—a ruggedized device showing real-time temperature data from 3,000 sensors.

Before we can start, he checks it twice.

Your LinkedIn says you're a "Lead Cooling Architect." What does that actually mean?

Kelvin: It means I spend my days trying to cool 47 megawatts of heat in a country that's basically a year-round sauna. I came from tropical fish farming. Commercial operations, breeding facilities where you're managing water temperature for species that die if you're off by two degrees.

He laughs, but it's not a happy sound.

Thought data centers would be easier. Turns out keeping discus fish alive is simpler than keeping AI processors cool when the ambient air is hot enough to incubate eggs.

The job is technically simple: keep 847 racks below 27°C inlet temperature while outside it's 33°C and humid enough to grow mushrooms. We use direct liquid cooling for the chips, immersion cooling for some racks, about 40 megawatts of traditional CRAC units as backup. On paper it works beautifully.

He glances at his tablet, frowns.

In reality we're in permanent crisis mode. Singapore has 72 data centers and every single one is built in the wrong climate.² But the data has to be here. Data sovereignty laws, latency requirements, regional hub economics. So we just engineer harder and pretend thermodynamics is a suggestion.

The World Economic Forum projects climate costs could hit data centers with $81 billion annually by 2035. Are you seeing that math already?

Kelvin: The math is playing out on my tablet right now. Our cooling costs are 43% of operational budget. Traditional data centers in cold climates? Maybe 30%. We're spending an extra $2.1 million monthly just fighting physics.

And it's getting worse. Last month we had three days at 36°C. Our systems were designed for 33°C peaks. Those three degrees don't sound like much, right? But cooling efficiency drops exponentially. We were running at 140% capacity, burning through backup chillers, and still watching inlet temps climb. Had to throttle processing on 90 racks.

Clients weren't happy. Contract penalties weren't happy. My cardiologist wasn't happy.

Another tablet check. His jaw tightens.

The AI chips are the real problem. Ten times the heat of previous generation.³ Liquid cooling helps but you're still moving that heat somewhere. And where does it go? Into cooling towers that evaporate 47,000 liters of water daily in a country that imports water from Malaysia.

The whole thing is thermodynamically insane but here we are, keeping the internet running one degree at a time.

You mentioned throttling racks. How often does that happen?

Kelvin: More than we admit publicly. Official line is "planned maintenance." Reality is we're doing triage. Medical triage but for servers.

Last August we had a cooling tower fail during a heatwave. Forty-five minutes of panic. Started shutting down non-critical systems. Development environments, test clusters, backup processes. Then secondary systems. Then we're looking at production databases and someone's asking "which client can afford downtime" and I'm thinking—

He stops mid-sentence, stares at the monitoring wall where several temperature readings have turned yellow.

Sorry. Where were we?

The cooling tower failure. You were making choices about which clients to shut down.

Kelvin: Right. The choices. We got it back online before anything critical went down but that's the future, isn't it? More failures, harder choices, higher temperatures.

“

Verisk Maplecroft says two-thirds of data center hubs face material heat risk by 2040.⁴ We're already there. We're the canary in the coal mine except the coal mine is also on fire and also we're supposed to keep the canary alive because the internet depends on it.

The engineers at the Sustainable Tropical Data Centre Testbed are working on solutions. Better immersion cooling, heat recovery systems, AI-optimized airflow.⁵ But that's a research facility. This is production. We can't experiment. We have SLAs. Uptime guarantees. If we go down, entire supply chains stop. Financial systems freeze. Emergency services lose connectivity.

So we just keep cooling. Keep spending. Keep watching the temperatures climb and hoping the next heatwave isn't the one that—

His tablet alarm goes off. He silences it without looking.

What was that?

Kelvin: Inlet temp on Row 7 hit 27.5°C. Still within tolerance but trending wrong. Happens maybe fifteen times a day now? Used to be once a week. Then once a day. Now it's constant. We're always one compressor failure away from catastrophe.

The thing nobody talks about is how the heat affects everything else. Power transmission efficiency drops in extreme heat.⁶ So we're getting less reliable electricity exactly when we need more of it. The backup generators? Rated for 35°C maximum ambient. We hit that last month. They ran but not at full capacity.

Our redundancy isn't redundant anymore when the whole system is stressed simultaneously. It's not one failure. It's cascading failures. Cooling tower struggles, so chillers work harder, so they draw more power, so electrical systems strain, so backup capacity shrinks, so we have less margin, so next heat spike hits harder, so we throttle more systems, so clients complain, so executives demand solutions, so we spend more money on equipment that's still fighting physics in the wrong climate and—

Another alarm. He doesn't silence it this time.

Sorry, I need to... Row 12 is yellow now. That's two rows. If we hit three simultaneous I have to activate emergency protocol which means calling the VP of Operations who's going to ask why we didn't prevent this and I'm going to say "because it's 34 degrees outside and we're trying to cool pizza ovens with air conditioning" but obviously I won't say that I'll say something about unexpected thermal loads and optimization opportunities and—

Should we pause?

Kelvin: No, it's fine. This is fine. This is just Tuesday. Every day is Tuesday now. Every day is watching temperatures climb and hoping the engineering holds and knowing that eventually it won't because you can't engineer your way out of thermodynamics you can only delay the inevitable and we're running out of delays but we keep trying because what else are we supposed to—

Third alarm. Row 18.

Three rows. I have to make the call. But first let me just... if we increase flow rate on the liquid cooling by 8% and bump the immersion tanks down by 1.5 degrees and redirect airflow from the cold aisle on Row 4 which is running cool anyway then maybe we can avoid... we can...

He's typing rapidly on the tablet, not looking at me anymore.

Kelvin?

Kelvin: Just need to... if I can balance the load across... redistribute the thermal... the problem is the chips generate... AI processing creates... we're trying to cool...

Fourth alarm. His hands are shaking slightly.

Four rows now. That's emergency protocol territory. But if I can just... the cooling towers are at maximum but the chillers could... no wait the chillers are already... maybe if we throttle the AI training jobs that would reduce heat load by maybe 12% which buys us... which gives us time to...

Fifth alarm.

Five. Okay. Five rows yellow. I need to call this in. But they're going to ask what happened and I don't... the temperature outside is only 34 which is normal for December which means this shouldn't be happening which means something else is failing which means...

Sixth alarm.

Six rows. The system is... we're losing thermal control across... if this cascades to the immersion tanks then we'll have to... but we can't shut down the immersion systems because those are the AI clusters and those are the highest priority clients and if we lose those then...

Kelvin, we should stop.

Kelvin: Stop? Can't stop. Have to... need to... the cooling... if the temperature... when the systems...

Seventh alarm. He's staring at the tablet but not seeing it.

Seven. That's cascading failure. That's the scenario we... that's when you... protocol says... but protocol assumes... doesn't account for...

His words are coming faster, fragmenting.

Heat rising... chillers maxed... power draw... backup capacity... throttle which systems... who gets shut... can't lose... have to choose... but choosing means... someone's data... someone's services... medical systems maybe... financial... emergency... can't know which... just have to...

Eighth alarm.

Eight rows. The VP will... executives will... clients will... but the heat... the temperature... Singapore wasn't... shouldn't have built... wrong climate... everyone knows... but data sovereignty... but regional hub... but economics... so we just... keep trying... keep cooling... keep fighting... thermodynamics... physics... entropy...

Ninth alarm.

Nine. That's... need to... have to...

He stands abruptly, still staring at the tablet.

The system... cooling... failure... cascading... heat... can't... losing...

Tenth alarm.

Ten rows. Emergency. Full. Have to... now... immediately... shut down... which ones... how many... clients... SLAs... penalties... but heat... temperature... 28.2°C... 28.5°C... rising... still rising...

He's moving toward the door, words barely coherent.

Call... VP... emergency... protocol... activate... shutdown... partial... which racks... priority... tier one... tier two... medical... financial... can't... choose... have to... heat... cooling... failing... system... cascade...

Eleventh alarm. He's at the door.

Eleven... infrastructure... critical... losing... thermal... control... complete... need... backup... generators... but heat... affects... everything... power... transmission... efficiency... drops... when... temperature...

He fumbles with the door handle.

Can't... stop... heat... physics... entropy... thermodynamics... Singapore... wrong... climate... built... anyway... trying... cool... impossible... fighting... losing...

Twelfth alarm.

Twelve... rows... yellow... turning... red... emergency... full... cascade... shutdown... imminent... clients... services... internet... supply... chains... everything... depends... can't... fail... but... failing... now...

The door opens. He's halfway through it, still talking, but the words are fragments, barely connected.

“

Heat... rising... systems... failing... cooling... impossible... temperature... 29°C... threshold... exceeded... cascade... complete... shutdown... necessary... choosing... which... data... whose... services... who... loses... connectivity... supply... chains... freeze... medical... financial... emergency... all... depends... this... facility... keeping... cool... but... can't... physics... wins... always... entropy... increases... heat... death... universe... starts... here... Row... 7... spreading... cascading... losing... control... thermal... complete...

He disappears down the corridor. Through the glass wall, the monitoring screens show more yellow indicators spreading across the facility map like a fever. The cooling system's white noise pitch has changed—higher, more strained. Outside, the Singapore sun climbs toward its midday peak.

Inside, the temperature keeps rising.

The interview ends here. K² did not respond to follow-up requests. The facility's public status page reported "scheduled maintenance" for the next four hours.