AI Servers Are Getting Too Hot

本站添加: 2026-05-13

[00:00:00]Nvidia's newest AI servers now consume up to 140 kilowatts of power. That's almost 10 times more than a normal enterprise server rack. [music] And now, traditional cooling systems are starting to fail. Not in the future. Right now.

[00:00:17]Companies like Microsoft and Meta are redesigning entire AI facilities around one problem, heat.

[00:00:26]Because AI scaling is no longer just creating smarter models. It's creating machines that are becoming physically difficult to cool.

[00:00:35]And this changes the economics of the entire industry. Because once cooling becomes infrastructure, only a few companies can afford to scale.

[00:00:44]And this is the part most people still don't understand. The AI industry solved compute and accidentally created a thermodynamics problem. Which raises a much bigger question.

[00:00:57]What happens when AI scaling starts depending on industrial infrastructure instead of software?

[00:01:04]For years, cooling was secondary. You built data centers around compute, networking, electricity.

[00:01:14]Cooling came afterward.

[00:01:16]That worked when racks consumed [music] 10 kilowatts, maybe 15.

[00:01:22]But AI scaling changed the math completely.

[00:01:25]Because modern AI systems concentrate enormous amounts of power into tiny physical spaces. More GPUs gives more electricity, [music] gives more heat, gives more thermal pressure, and eventually, normal air cooling starts reaching physical limits.

[00:01:43]That's why the industry is suddenly shifting toward liquid [music] cooling infrastructure.

[00:01:48]Not because it looks futuristic, because the old [music] systems are starting to break. And this creates a second problem. The companies that can redesign infrastructure around heat gain a massive advantage.

[00:02:02]Everyone else falls behind.

[00:02:04]This is no longer just a software race.

[00:02:07]It's becoming a physics race. And once thermodynamics becomes the bottleneck, money starts deciding who can scale AI.

[00:02:17]For years, data centers were designed around a simple assumption. Servers would keep getting more powerful, but cooling systems would remain mostly the same.

[00:02:28]That assumption is now breaking.

[00:02:31]Traditional enterprise server racks usually consumed around 10 to 15 kilowatts of power.

[00:02:37]The newest Nvidia Blackwell AI racks, up to 140 kilowatts. In just a few years, AI infrastructure jumped from normal server density to something closer to industrial scale power concentration.

[00:02:54]And this is where the system starts changing completely because AI scaling is not just creating bigger models.

[00:03:01]It's creating extreme physical density.

[00:03:04]More GPUs gives more electricity, gives more heat, gives more thermal pressure inside the same physical space. And the newest AI systems are pushing this to absurd levels.

[00:03:17]Some future AI racks are already projected to move toward 250 kilowatts and beyond.

[00:03:25]That is not a normal infrastructure upgrade. That is a completely different category of facility.

[00:03:31]And this creates a second problem most people still don't see.

[00:03:36]Traditional air cooling was never designed for this level of heat.

[00:03:42]For decades, air cooling worked because server density increased gradually.

[00:03:48]AI changed that.

[00:03:50]Now compute density is exploding faster than infrastructure can adapt.

[00:03:56]And once racks move past certain thermal limits, air cooling starts failing. That's why the industry is suddenly shifting toward [music] liquid cooling systems, not because it looks futuristic, because the old systems are reaching physical limits.

[00:04:12]And this is where the AI industry crossed into something very different.

[00:04:17]Companies are no longer just buying GPUs.

[00:04:21]They are redesigning infrastructure around thermodynamics.

[00:04:24]Nvidia redesigned Blackwell systems around liquid cooled architectures.

[00:04:30]Microsoft and Meta are redesigning facilities around thermal management.

[00:04:35]Cooling is rapidly becoming mandatory for high-density AI infrastructure.

[00:04:40]And this is where the economics become brutal.

[00:04:43]Cooling systems for a single Blackwell rack can cost around $50,000.

[00:04:49]Just cooling. Not the GPUs, not the networking, not the power systems, just thermal management.

[00:04:59]And this changes who can actually compete.

[00:05:02]Hyperscalers can redesign buildings, cooling infrastructure, power delivery systems, facility layouts. Most companies cannot because retrofitting existing infrastructure for 120 kW AI racks is massively expensive.

[00:05:21]At that point, you are not upgrading servers anymore. You are rebuilding industrial infrastructure around heat.

[00:05:29]And this is where the market starts closing. Money gives cooling infrastructure, gives thermal capacity, gives AI scaling capability.

[00:05:39]Most companies are locked out, not because they cannot build AI models, because they cannot support the physical infrastructure required to cool them.

[00:05:50]The AI industry is now entering a completely different phase because the newest AI infrastructure no longer behaves like traditional computing. It behaves like industrial infrastructure.

[00:06:04]Cooling systems for a single Blackwell rack can cost tens of thousands of dollars. Just cooling.

[00:06:11]Not the GPUs, not the networking, not the power systems, just thermal management. And the deeper this goes, the harder it becomes for smaller companies to compete. Hyperscalers can redesign buildings, cooling systems, power delivery, facility architecture.

[00:06:33]Most companies cannot. And this is where the bottleneck becomes visible.

[00:06:38]The companies that can manage heat will control AI scaling.

[00:06:44]The AI race is no longer just about better models. It's becoming a race to control infrastructure because once scaling depends on power, cooling, physical facilities, industrial capital, the market starts closing. Most companies are locked out not because they cannot build AI, because they cannot afford the infrastructure required to scale it. Money gives infrastructure, gives access, gives control.

[00:07:15]And that may become the defining power structure of the AI industry.

[00:07:19]AI is no longer just a software problem.

[00:07:23]It's becoming a thermodynamics problem.

[00:07:27]And this is the industry code.

[00:07:55]>> Mhm.