
What this covers
This is a High Yield walkthrough, built on SemiAnalysis's datacenter research, of what separates an AI datacenter from the traditional kind. It organizes the material around a four-part C3P Framework (Compute, Connectivity, Cooling, Power)). It begins with the origins of datacenters and why Ashburn/Sterling, near Dulles, became a hub: it sits close to population centers and its networking grew iteratively with each new building. It then argues that the useful measure of size is critical IT power rather than square footage. Retail sites provide under 10 MW, wholesale sites 10 to 30 MW, hyperscale sites 40 to 100 MW, and AI sites more than 200 MW.
The middle section covers the engineering. Location matters less because training) clusters are closed systems and inference) time dwarfs network latency. Rack-level density (compute)) takes over as the goal. Nvidia GPU power climbs from 250 W (Volta) to 1,000 W (Blackwell) and 2,700 W for a GB200 board. The GB200 NVL72 rack uses roughly two miles of copper because optical links would add 20,000 W. Cooling gets its own stretch: an air-cooled AMD MI300X is about 90% heatsink by volume, while liquid absorbs roughly 4,000 times more energy per unit volume and cannot be retrofitted into air-cooled buildings.
The later chapters turn to the industry and are the most contentious. Meta abandoned a half-built conventional datacenter in Temple, Texas in 2023, then moved to tents for speed and to the Prometheus and Hyperion sites planned at gigawatt scale. xAI runs 150,000 GB200 units partly on mobile generators. CoreWeave converts a failed Bitcoin site in Denton, Texas with a gas plant. Microsoft and Constellation Energy plan to restart Three Mile Island's TMI-1 in 2027. The speaker's claim is that a belief in a winner-take-all race to AGI pushes hyperscaler companies toward becoming energy producers, and the video closes on satellite imagery and SemiAnalysis's industry model.
AI datacenters represent a fundamental shift from traditional datacenters driven by compute density requirements, creating an insatiable demand for power that will make hyperscalers major energy producers and potentially consume global energy capacity at unprecedented scales.
- AI datacenters prioritize compute density over geographic distribution, fundamentally inverting the location-sensitivity that defined traditional datacenters
- Power requirements have escalated from 10-30 MW for traditional facilities to 200+ MW for AI facilities, with gigawatt-scale campuses already under construction
- The race to AGI creates unsustainable energy demand growth (40-50 GW estimated 2022-2026) that forces hyperscalers to acquire and operate nuclear power plants
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
Location does not matter for AI datacenters in the way it does for traditional datacenters because training clusters are closed systems with no customer access, and inference workloads are not latency-sensitive since compute time (multiple seconds) vastly outweighs network latency.
“A training cluster is a more or less a closed system. It literally doesn't matter where you place it, at least not in the sense of being close to customers. Because there are no customers accessing data.”
The relevant measure of datacenter size is 'critical IT power' (total power capacity) rather than physical building size, with traditional retail datacenters providing <10 megawatts, larger wholesale datacenters in the 10-30 megawatt range, modern hyperscale datacenters at 40-100 megawatts, and AI datacenters exceeding 200+ megawatts.
“It's not the actual size of the building we are talking about, it's the total power capacity of the datacenter, also called 'critical IT power'.”
Microsoft announced a deal to restart the decommissioned Three Mile Island nuclear reactor (TMI-1) in Pennsylvania with all energy going to Microsoft for the next 20 years starting in 2027, exclusively to power the next generation of AI datacenters.
“Last year, in 2024, Microsoft announced a deal with Constellation Energy, the owner of the site, to restart the still working TMI-1 reactor. The nuclear power plant is expected to resume operation in 2027, with all energy going to Microsoft for the next 20 years.”
Compute density is described as 'the holy grail' in AI datacenters because the optimization goal is maximizing GPUs per board, boards per rack, and racks per facility, with the constraint that each optimization enables marginal gains in power efficiency.
“Compute density is the holy grail when it comes to AI datacenters. You want your GPUs to use as much silicon as possible, have as many GPUs on a single board as possible and as many of these in a single rack as possible.”
AI datacenters are transitioning from air cooling to liquid cooling because liquid cooling is ~4000 times more effective at absorbing energy per unit volume than air, allows for compact GPU designs by eliminating large heatsinks, and enables running silicon at lower temperatures which increases both lifespan and energy efficiency.
“Liquid cooling can absorb about 4000 times more energy per unit of volume than air. If you have to remove a lot of heat, because you have to cool lots of GPUs in a very dense setup, it's the only option.”
Coreweave retrofitted an old Bitcoin mining datacenter in Denton, Texas with a massive gas-powered power plant that directly supplies power to the AI datacenters, demonstrating repurposing of failed crypto infrastructure for AI.
“Coreweave, a hyperscaler, acquired and retrofitted an old crypto mine datacenter in Denton, Texas, that was previously used to mine Bitcoin. If we take a look at satellite images, we can see a large cluster of buildings in this location. But one is not like the others. The center of the site is a massive gas powered power plant that directly supplies power to the AI datacenters.”
Meta is deploying AI datacenters in tents because it reduces construction time, and speed is critical in the AI race, with Meta's top AI locations being Prometheus in Ohio (planned to scale beyond 1 gigawatt) and the Hyperion supercluster in Louisiana (planned to reach 2 gigawatts by 2030 with room to grow to 5 gigawatts).
“Meta not only rebuilt the datacenter on our field in Temple, they are starting to place new AI datacenters in tents - because it reduces construction time. And aside from compute density, being fast is important in the AI race.”
The major AI players (Google, Microsoft, Meta, Amazon, Elon Musk's xAI) believe they are in a race to AGI where the winner takes the entire market, the prize is worth trillions of dollars, and this belief drives willingness to spend any amount of power and resources to get there first.
“The simple fact is that it doesn't matter if you believe in AGI or not. All that matters is that the major players very clearly believe it's a race to AGI and whoever gets there first takes the entire cake.”
AI datacenters run at close to full load constantly, unlike traditional datacenters which have fluctuating power demand based on usage patterns, resulting in higher effective power utilization and continuous strain on power infrastructure.
“While a traditional datacenter has fluctuating power demand based on usage patterns and rarely runs at full power, AI datacenters are more or less constantly running at close to full load. They not only have this massive critical IT power; they actually use it.”
The Ashburn/Sterling area near Dulles International Airport in Washington D.C. has become a major datacenter hub because it combines geographic proximity to major population centers with highly developed networking infrastructure that grew iteratively as new datacenters were built in the location.
“Datacenters, as we know them, are more location sensitive than you might think. A great example for this is communities like Ashburn or Sterling, which are located right next to the Dulles International Airport in Washington D.C. This small area contains a lot of datacenters because it combines a prime location, close to major population centers, with massive networking infrastructure that grew each time a new datacenter was built there. It's a perfect networking hub.”
Nvidia's GB200 NVL72 rack contains over 5,000 wires and two miles of copper, using copper-based NVLink interconnects throughout the rack because optical interconnects would consume 20,000 watts more power, illustrating the priority of power efficiency over distance in AI datacenter design.
“Nvidia's GB200 NVL72 compute rack contains over 5,000 wires and two miles of copper. If Nvidia would have used optics instead, it would have consumed 20,000 watts more than the current copper based NVLink solution.”
ChatGPT was released in November 2022, less than three years ago, and since that release everything in the AI and datacenter industry has accelerated, with the race for AGI creating an insatiable demand for compute that continues to accelerate.
“ChatGPT was released in November of 2022, less than three years ago. Ever since it feels like everything is speeding up. The race for AGI has created an insatiable demand for AI compute that only seems to accelerate.”
Elon Musk's xAI is the largest AI datacenter by GPU count, featuring 150,000 Nvidia GB200 units, with datacenters growing so rapidly they must be powered by mobile generators while main power sources take too long to connect.
“Elon Musk's xAI is the largest yet, with 150,000 Nvidia GB200. Datacenters are growing so quickly, they have to be fed with mobile generators because the main power sources take too long to get online. Everyone is scrambling as fast as they can.”
Nvidia GPU power consumption and performance has skyrocketed across generations: Volta 250W TDP, Ampere 400W, Hopper 700W, Blackwell 1000W, with GB200 superchip reaching 2,700W, and future GPUs approaching 2000W per chip, demonstrating exponential density increases.
“While Volta had an almost tame TDP at only 250 watts, Ampere, its successor, raised it to 400 watts. Next, Hopper increased the TDP to 700 watts and Nvidia's newest generation, Blackwell, is reaching 1000 watts.”
Microsoft announced a deal with Constellation Energy to restart the TMI-1 reactor at Three Mile Island nuclear power plant, with the facility expected to resume operation in 2027, providing all generated electricity to Microsoft for the next 20 years to power next-generation AI datacenters.
“Last year, in 2024, Microsoft announced a deal with Constellation Energy, the owner of the site, to restart the still working TMI-1 reactor. The nuclear power plant is expected to resume operation in 2027, with all energy going to Microsoft for the next 20 years.”
The Meta H-type datacenter was designed for energy efficiency and was already capable of running GPUs (Nvidia Hopper H100s), with Meta operating a 100,000 H100 GPU cluster across multiple H-type datacenters, but even this GPU-capable design was insufficient density to stay competitive.
“While it was designed for maximum energy efficiency, Meta's H-type datacenters are already capable of running GPUs. Meta has a massive Nvidia Hopper based AI cluster that combines 100,000 H100 GPUs across multiple of the same H-type datacenters.”
Liquid cooling requires complete redesign of datacenter infrastructure from scratch including waterpipes at rack and building level plus massive cooling towers, making it impossible to retrofit into existing air-cooled datacenters and requiring planning from the very beginning.
“An air-cooled datacenter is designed very differently to a liquid-cooled one. It's not your average desktop PC where you can just upgrade an air-cooler to a water-cooler, it completely changes the layout of the datacenter. You need to include waterpipes, from the rack- to the building-level and install massive cooling towers.”
Liquid cooling can absorb approximately 4,000 times more energy per unit of volume than air cooling, making it the only viable option for super high-density AI accelerator designs that must dissipate hundreds of kilowatts per rack.
“Liquid cooling can absorb about 4000 times more energy per unit of volume than air. If you have to remove a lot of heat, because you have to cool lots of GPUs in a very dense setup, it's the only option. Super high-density designs are only feasible with liquid cooling.”
Increasing compute and power density at rack scale directly drives increasing power consumption per rack, creating cascading requirements for supporting infrastructure (cooling, power delivery, generators).
“That's why the power requirements for a single rack are continuing to grow.”
GPU thermal design power (TDP) has increased dramatically across recent Nvidia generations: Volta (250W), Ampere (400W), Hopper (700W), and Blackwell (1000W for single GPU), with Nvidia's GB200 superchip combining two Blackwell GPUs with a Grace CPU reaching 2,700W TDP for a single board, and announced next-generation GPUs using four reticle-sized chips will approach 2000W per GPU.
“While Volta had an almost tame TDP at only 250 watts, Ampere, its successor, raised it to 400 watts. Next, Hopper increased the TDP to 700 watts and Nvidia's newest generation, Blackwell, is reaching 1000 watts. For a single GPU. A GB200 superchip, which combines two Blackwell GPUs with a Nvidia Grace CPU, has a whopping 2,700-watt TDP, for a single board.”
Meta abandoned construction of a traditional H-type datacenter in Temple, Texas halfway through in 2023, having invested an estimated 70 million dollars, because the design could not deliver the energy density required to remain competitive in the AI race even though it was already AI-capable.
“an estimated 70 million dollars just gone. Wasted it seems.”
Energy generation and power infrastructure are becoming central strategic focuses for hyperscalers and major tech companies, transitioning from purely software/compute companies to energy producers.
“That not only means more transformers, more generators and liquid cooling, but energy generation is more and more a focus point of hyperscalers and big tech companies.”
AMD's MI300X AI-accelerator GPU dedicates approximately 90% of its physical volume to heatsinks, with the actual GPU PCB occupying only ~10%, illustrating the space inefficiency of air cooling.
“about 90% of its volume is taken up by a massive heatsink. The small PCB below all that metal is the actual GPU.”
AMD's MI300X GPU accelerator requires a massive heatsink that occupies approximately 90% of the GPU's physical volume, with the actual silicon PCB representing only a small fraction of the total component size when air-cooled.
“This is AMD's MI300X AI-accelerator. One GPU, ready to be installed on a server blade. But about 90% of its volume is taken up by a massive heatsink. The small PCB below all that metal is the actual GPU.”
AI datacenters will eventually 'eat the world' in the sense that AI infrastructure will become the dominant consumer of capital, energy, and resources, and the first evidence of this dominance is already visible in current infrastructure transformation patterns if you know where to look.
“It truly starts to look like AI datacenters are going to eat the world. The first bites are already visible, if you know where to look.”
Germany's average power consumption is approximately 60 gigawatts, providing a reference scale for comparing individual AI datacenter power demands to entire national grids.
“For comparison, the country of Germany has an average power usage of about 60 gigawatts.”
AI datacenters often have very limited uninterruptible power supply (UPS) systems because the cost and permitting time for large emergency generators is prohibitive, and they simply stop working if main power fails, which is acceptable for training runs since GPUs already introduce frequent failures allowing continuation when power returns.
“AI datacenters have such a massive power demand, they need a large amount of generators, which not only have to be bought, but also require the proper permits. This not only adds a lot of additional cost, but takes time to set up. And since time to market is crucial in the AI race, AI datacenters often have a very limited UPS system.”
Even if ChatGPT was not the direct cause of Meta's Temple datacenter abandonment, the timeline suggests it served as a catalyst that made clear to hyperscalers the competitive urgency of deploying higher-density AI designs.
“So, is that our answer? Did Meta start building a traditional datacenter and realized half way through the construction that it was outdated? Well, kind of. But it's even more radical than that.”
The AI race is deliberately opaque, with hyperscalers keeping their actual compute capacity, future scaling plans, and datacenter efficiency metrics private to maintain competitive advantage and prevent competition from understanding their relative positions.
“Unlike supercomputers that are eager to get listed in the Top500, to show off what they achieved, AI datacenters are much more private. The AI race isn't happening out in the open, at least outside of PR announcements, if you don't know where to look. Hyperscalers don't want you or the competition to know how much actual compute they have, how much they will add over the next month and years and how competitive - or rather how dense - their AI datacenters are. And how much power all their datacenters consume.”
AI datacenters differ fundamentally from traditional datacenters in four critical components (connectivity, compute, cooling, power), with connectivity being location-insensitive for AI because training clusters are closed systems with no customer access and inference does not require low-latency or high-bandwidth connections.
“Calling it 'AI datacenter' might even be a bit misleading, but it has become the established name. In my opinion a more fitting name would be 'AI supercomputer', because that's what it actually is...From a high-level overview a datacenter has four main components: compute, connectivity, cooling and power (C3P).”
The major technology companies (Google, Microsoft, Meta, Amazon, and others) explicitly view the AI race as a race toward AGI, and the belief that whoever achieves AGI first captures the entire competitive value ('takes the entire cake') justifies any expenditure on compute infrastructure, including the massive capital and power commitments currently being made.
“The simple fact is that it doesn't matter if you believe in AGI or not. All that matters is that the major players very clearly believe it's a race to AGI and whoever gets there first takes the entire cake. And because that cake is worth trillions of dollars, they are willing to do everything in their power to get there first. And that quite literally requires power. A lot of power.”
The race for AGI requires both compute and power literally and figuratively, and power (energy) has become the constraining factor, not silicon or architecture.
“And that quite literally requires power. A lot of power. The race for AGI is not only about compute, it's about power. Both literally and figuratively.”
The C3P framework (Compute, Connectivity, Cooling, Power) is used to compare AI and traditional datacenters, with fundamental differences in all four areas.
“From a high-level overview a datacenter has four main components: compute, connectivity, cooling and power (C3P).”
AI datacenters require transformers to step down voltage from high voltage (>100 kilovolts) to medium voltage and then to low voltage (usually 415 volts for datacenters), and the demand for transformers has become so high that order books already have backlogs, with transformers previously bought by governments for cities now being purchased by AI companies.
“server racks don't run on high voltage, which is over 100 kilovolts, you need transformers to step down the voltage. First to medium voltage and then to low voltage, which for datacenters is usually 415 volts.”
Elon Musk's xAI is deploying the largest AI cluster yet with 150,000 Nvidia GB200 GPUs, demonstrating continued escalation in individual company compute deployments.
“Elon Musk's xAI is the largest yet, with 150,000 Nvidia GB200.”
Transformer equipment, previously purchased primarily by governments for large cities and industrial centers, is now in severe shortage due to AI datacenter demand, with order books already backlogged as hyperscalers require massive numbers of transformers to step down voltage from high-voltage power lines to datacenter-usable voltages.
“With the power consumption that rivals large cities, these massive AI datacenters require a lot of transformers. So many in fact, that the order books already have a back-log. Transformers, which previously were mostly bought by governments to serve large cities and industrial centers, are suddenly in high demand because of AI.”
Meta is deploying AI datacenters in temporary tent structures to reduce construction time, recognizing that speed-to-market is critical in the competitive AI race alongside compute density.
“Meta not only rebuilt the datacenter on our field in Temple, they are starting to place new AI datacenters in tents - because it reduces construction time. And aside from compute density, being fast is important in the AI race.”
Traditional datacenters require uninterruptible power supply (UPS) systems with batteries and emergency generators to prevent critical failure during power outages, but AI datacenters often have limited UPS systems and no emergency generator backup, relying instead on simply restarting training runs after power restoration since GPU failures are already frequent in training operations.
“For a traditional datacenter, loss of power is a major critical failure point...But AI datacenters have such a massive power demand, they need a large amount of generators, which not only have to be bought, but also require the proper permits. This not only adds a lot of additional cost, but takes time to set up. And since time to market is crucial in the AI race, AI datacenters often have a very limited UPS system. In this case, the datacenter just stops working, if the main power source fails. Which, funny enough, isn't that big of a deal for training runs, as the GPUs already introduce somewhat frequent failures. When the power goes out, you can simply continue the training run when it comes back.”
Traditional datacenters measure size by building square footage or physical dimensions, but modern datacenters measure size by total critical IT power capacity (measured in megawatts), which has become the primary industry denominator for comparing datacenter scale.
“I'm talking about power at the level of the entire facility, which has become the number one denomination when we talk about datacenter 'size'. It's not the actual size of the building we are talking about, it's the total power capacity of the datacenter, also called 'critical IT power'.”
AI datacenters called 'AI supercomputers' more accurately describes their function than 'AI datacenters,' since their only goal is to deliver as much computational performance as efficiently as possible.
“Calling it 'AI datacenter' might even be a bit misleading, but it has become the established name. In my opinion a more fitting name would be 'AI supercomputer', because that's what it actually is.”
Meta purchased land in Temple, Texas and began constructing a datacenter in mid-2022, but by April 2023 abandoned the partially-completed construction despite investing an estimated 70 million dollars, because the H-type datacenter design was already becoming outdated relative to emerging AI compute density standards.
“Fast forward one year to August of 2022 and the entire field is gone. Replaced by a giant construction site...Just four months later – now in December of 2022 – we can see that the construction of a datacenter is well underway. But then something strange happens. Another five months later, in April of 2023, our curious field in Temple looks like this: all the previous construction gone, razed to the ground. Meta just deleted their entire datacenter halfway through its construction. An estimated 70 million dollars just gone.”
From ChatGPT's launch in November 2022 to next year, AI compute will add an estimated 40-50 gigawatts of global power demand, comparable to the average usage of entire countries like France and Germany.
“From the launch of ChatGPT to next year, AI compute will add an estimated 40 to 50 gigawatts of global power demand. These are numbers comparable to the average usage of entire countries like France and Germany.”
Nvidia's GB200 NVL72 rack contains over 5,000 wires and two miles of copper for interconnect; if optical interconnects were used instead, the solution would consume 20,000 watts more power than the current copper-based NVLink solution, demonstrating the power efficiency advantage of copper for rack-scale density.
“Nvidia's GB200 NVL72 compute rack contains over 5,000 wires and two miles of copper. If Nvidia would have used optics instead, it would have consumed 20,000 watts more than the current copper based NVLink solution.”
Coreweave, a hyperscaler, acquired and retrofitted a former Bitcoin mining datacenter in Denton, Texas, which included a massive on-site gas-powered power plant that directly supplies electricity to the AI datacenters.
“Coreweave, a hyperscaler, acquired and retrofitted an old crypto mine datacenter in Denton, Texas, that was previously used to mine Bitcoin. If we take a look at satellite images, we can see a large cluster of buildings in this location. But one is not like the others. The center of the site is a massive gas powered power plant that directly supplies power to the AI datacenters.”
Meta's Prometheus AI cluster in Ohio is currently powered by gas turbines and is expected to scale to over 1 gigawatt of critical IT power within the next year, establishing it as one of Meta's top two AI datacenter locations.
“Meta's top two AI locations are Prometheus in Ohio, an already existing AI cluster powered by gas turbines, that's supposed to scale to over 1 gigawatt in the next year.”
SemiAnalysis operates a datacenter model tracking over 5,000 datacenters worldwide, analyzing high-resolution satellite imagery in detail to assess power generation, cooling infrastructure, and other visible elements, reconstructing detailed specifications and competitive insights for each datacenter facility.
“SemiAnalysis is tracking over 5,000 datacenters world-wide. And by tracking I don't mean a simple Excel sheet with a name and an address. I'm talking about the true Sherlock Holmes stuff. The datacenter model not only tracks new construction projects, but also existing datacenters. High-res satellite images are analyzed in detail. And because the power stages, generators and cooling infrastructure are visible, the datacenter team can actually create very detailed insights for each datacenter.”
Germany's average power usage is approximately 60 gigawatts, which provides a comparative anchor: Meta's Hyperion supercluster, designed to reach 5 gigawatts, would represent a single facility using 8% of an entire nation's power consumption.
“For comparison, the country of Germany has an average power usage of about 60 gigawatts.”
Early datacenters were often located inside business properties in basements as afterthought server rooms, but grew into massive construction projects that fueled Web 2.0 and became the 'factories' of internet companies like Google, Microsoft, Meta, and Amazon.
“Early datacenters were often located inside business properties, where a few rooms, for example in the basement, were converted to server rooms, housing the IT equipment.”
Traditional datacenters are highly location-sensitive because they must be close to major population centers and must have access to massive networking infrastructure to provide fast, low-latency access to content like YouTube, Netflix, and cloud services.
“when it comes to content like YouTube, Netflix or basically anything else in the cloud that you want to access, a datacenter not only has to provide a massive amount of data storage, but also a lot of network bandwidth and good latency. And that's where location comes into play.”
If AI becomes more responsive (lower latency inference), then latency and bandwidth requirements for AI datacenters may increase, changing the current location-agnostic deployment model.
“This might change in the future, once AI becomes more responsive. But for now, neither training nor inference have strong latency or bandwidth requirements.”
Ashburn and Sterling near Dulles International Airport in Washington D.C. became major datacenter hubs because they combine prime location close to major population centers with massive networking infrastructure that grew as each new datacenter was built there.
“communities like Ashburn or Sterling, which are located right next to the Dulles International Airport in Washington D.C. This small area contains a lot of datacenters because it combines a prime location, close to major population centers, with massive networking infrastructure that grew each time a new datacenter was built there.”
Server GPUs lack individual cooling fans and rely on massive heatsinks cooled by industrial high-performance fans that cool every component on a blade, unlike consumer GPUs.
“Unlike consumer GPUs, server GPUs don't have individual fans. They just come with massive heatsinks and are cooled by industrial high-performance fans that cool every component on a single blade.”
The vast majority of internet video content (YouTube, Netflix, etc.) is stored and served from large datacenters located relatively close to end users to minimize latency and ensure reliable delivery.
“Right now, you are watching a YouTube video, which means it's stored inside a Google datacenter somewhere. And that somewhere is hopefully not too far from you, because you want fast and uninterrupted access.”
Early internet datacenters were often located in building basements and converted server rooms, with only limited space and infrastructure, but grew into massive industrial facilities as the internet scaled and datacenters became the critical infrastructure for Web 2.0 companies.
“But for a long time, that's basically what a datacenter was. Early datacenters were often located inside business properties, where a few rooms, for example in the basement, were converted to server rooms, housing the IT equipment. With the rise of the internet, datacenters became more and more important. And they grew in size. Gone were the days of being an afterthought in a basement somewhere. Datacenters became massive construction projects and fueled the rise of the Web 2.0.”
Server GPUs do not have individual fans and instead rely on high-performance industrial fans cooling all components on a blade simultaneously, unlike consumer GPUs which have dedicated cooling solutions.
“Unlike consumer GPUs, server GPUs don't have individual fans. They just come with massive heatsinks and are cooled by industrial high-performance fans that cool every component on a single blade.”
Traditional datacenters outside hyperscalers rarely use more than 10 kW per rack, with 3-7 kW being typical and anything above 10 kW considered high-performance, whereas hyperscaler traditional datacenters reach 15-20 kW per rack.
“If you pick a random server in a traditional datacenter, outside of hyperscalers like Google, Meta and so on, you'd be hard pressed to find a rack that uses more than 10 kilowatts. The typical rack power consumption is in the range of 3 to maybe 7 kilowatts. Everything above 10 kilowatt per rack is already considered high-performance for a traditional datacenter. And while hyperscalers are building racks in the 15 to 20 kilowatt range”
Knowledge of AI datacenter projects comes from a combination of extensive knowledge bases, high-quality research, and professional high-resolution satellite imagery (not Google Earth updates), used to analyze power stages, generators, and cooling infrastructure visible in images.
“The answer is a combination of a large knowledge base, lots of high quality research and actually spending the money on high-resolution satellite images. And I don't mean the kind on Google Earth that maybe gets a low-res update every half year. I mean professional satellite images.”