$1 Trillion in Chips. One Physics Problem Nobody Has Solved.
Vera Rubin draws 2,300 watts per chip. The racks are fanless. The cooling demand has doubled.
My son Maison is eight. A few weeks ago, he watched me reading about data centres and asked me a question I have not been able to shake.
“Dad, where does the internet live?”
I told him it lives in very large buildings filled with computers. He thought about this, then asked: “What happens when you run your computer for a really long time?” He knew the answer. “It gets hot.” Now imagine a million computers running at the same time, all day, every day, forever. “That would be really, really hot.” Yes it would. “So who keeps them cold?”
I paused. Because the truth is, I have spent months studying that exact question. And the more I studied it, the more I realised that the answer is one of the most important and most overlooked investments in the entire AI buildout.
He looked at me and said: “Dad, if nobody keeps them cold, do the computers break?”
Yes, Maison. They do. And when they break, it costs roughly $9,000 per minute. For a hyperscaler running a frontier AI training cluster, the number is far higher. A single thermal failure in a rack filled with Vera Rubin GPUs could destroy millions of dollars of silicon in seconds. The cooling system is not an amenity. It is the difference between a functioning data centre and a very expensive oven.
What Vera Rubin Actually Requires
Five days ago, Jensen Huang stood on stage at GTC and said he now sees at least $1 trillion in purchase orders for Nvidia’s Blackwell and Vera Rubin platforms through 2027. Last year, that number was $500 billion. It doubled in twelve months.
But the number is not what matters most. What matters is what sits inside the rack.
Each Vera Rubin GPU draws approximately 2,300 watts, a 64 percent increase over Blackwell at 1,400 watts. A single NVL72 rack consumes up to 230 kilowatts at peak power. That is the electrical load of 175 American homes in a cabinet the size of a refrigerator.
Vera Rubin’s compute trays are entirely fanless. Airflow requirements drop 80 percent. Liquid cooling flow nearly doubles. This is not air cooling with liquid assist. This is liquid cooling, period.
Nvidia made this choice because 2,300 watts per chip cannot be removed by air. The physics are non-negotiable. You are removing the heat output of a space heater from an object the size of a dinner plate, for 72 such objects simultaneously, without a single thermal excursion.
Rubin Ultra, on the roadmap for late 2027, draws 3,600 watts per GPU. A full Kyber rack of 576 GPUs will consume 600 kilowatts. The electricity of nearly 500 American homes in a single cabinet.
At that density, the building is no longer a data centre.
It is a cooling system that happens to contain computers.
Blackwell to Vera Rubin to Rubin Ultra to Feynman. Each generation hotter. The demand for cooling does not flatten until chips stop getting hotter. Thermodynamics guarantees this curve has no visible ceiling.
The Consensus Is Wrong
Here is what the market believes: cooling is a commodity input. The hyperscalers will buy it from whichever large manufacturer offers the best price. Carrier, Trane, Schneider, Vertiv, they all make cooling equipment. The market is competitive. Margins will compress. There is nothing special here.
This view is wrong, and it is wrong for a specific reason.
Vera Rubin does not need a cooling product. It needs a cooling solution engineered from scratch for a rack geometry, chip architecture, and power density that has never existed before. It needs a thermal path designed for 72 fanless GPU packages drawing 2,300 watts each, with liquid delivered at precise temperature, pressure, and flow rate, maintaining full redundancy in an enclosure where a single failure can destroy millions of dollars of hardware in minutes.
The giants of the cooling industry make excellent standardised products. But a standardised product cannot solve a bespoke problem. This is not a problem you solve with a catalogue. This is a problem you solve with engineers.
The Bespoke Manufacturer
We own a position in a company that I believe sits at the centre of this problem. I will not name it, because we are still building the position. But I will describe the business in enough detail that a serious reader can evaluate the thesis independently.
The company is a manufacturer of highly engineered, custom-built thermal management systems. Several years ago, it acquired a specialist in high-performance cooling for hyperscale data centres, with deep expertise in both airside and direct-to-chip liquid cooling. The co-founder of that acquired business is now the CEO of the entire company. His background is structural engineering and high-performance thermal systems. The board has made its bet on where the future lies.
The data centre cooling subsidiary’s revenue more than doubled in the past year. Backlog surged 140 percent to over $1.3 billion, with a book-to-bill of 2.4. The revenue run rate is approaching $550 million, with a target of $1 billion within three to four years. The company builds equipment designed to last 25 to 30 years, compared to the industry average of 15 to 20. It sells directly to building owners who optimise for total cost of ownership, not the lowest sticker price.
And the one we own is on the very short list of companies that can engineer what Vera Rubin demands.
I will be writing about the other side of this physical layer, the companies that build the ground beneath the data centres before the first server ever arrives, in my next letter.
Cooling Is Not a Cost. It Is Revenue.
Nvidia itself frames this in terms that should change how investors think.
In its technical documentation for Vera Rubin, the company describes “parasitic energy,” the roughly 30 percent of total data centre power lost to conversion, distribution, and cooling before it reaches the GPUs. At scale, that represents billions in wasted revenue. Every watt spent on cooling instead of compute is a watt that could have generated tokens. Every token not generated is revenue not earned.
A more efficient cooling system directly increases the tokens a facility produces. The cooling manufacturer is not selling a commodity. It is selling economic capacity. When GPU rental rates are surging 15 to 20 percent year to date and compute is sold out through year-end, every additional token is real money.
The Thermal Lock-In
When a hyperscaler designs a cluster around Vera Rubin, the cooling is co-engineered from the earliest design stages. Nvidia’s documentation describes Vera Rubin as requiring “nearly double” the coolant flow of the prior generation. New cold plates, new manifolds, new quick-disconnect fittings, new coolant distribution units, all engineered for this specific platform.
Once specified into a deployment, switching requires re-engineering the entire thermal path. When the cluster costs hundreds of millions and the customer needs it operational on schedule, the qualified partner is not replaceable.
What Happens After the Data Centres Are Built
A thoughtful reader should ask: once the cooling systems are installed, does the demand stop?
It does not.
GPU generations turn over every 18 to 24 months. A system designed for 1,400 watts per GPU cannot serve a chip drawing 2,300 watts. Every generation of silicon creates a new generation of cooling demand inside the same building. The installed base also generates recurring revenue through maintenance and replacement of components under continuous thermal stress.
And the buildout itself is not a single event. The world operates roughly 80 gigawatts of data centre capacity today. That doubles by 2028 and grows at 14 percent annually into the 2030s. The hyperscalers are not building one wave and stopping.
The demand for cooling is not a one-time purchase. It is a compounding obligation.
The Simple Version
Every chip needs to be cooled. The hotter the chip, the harder the cooling. The harder the cooling, the fewer companies can do it. The fewer companies that can do it, the more valuable those companies become.
Vera Rubin is the hottest chip ever built. The one after it will be hotter. And the one after that, hotter still.
That is the entire thesis.
This is the first of two letters on the physical layer of AI infrastructure. The next will examine the companies that build the ground beneath the data centres, before the first server ever arrives. Subscribe so you do not miss it.
Neel Khokhani
Founder and CEO, Epochal Corporation
@neel_epochal
Great piece .. thank you🙌🏻