AI最烧钱的战场：数据中心的真实账单

NASA's estimated budget of $500 billion for a human landing on Mars could buy 1.36 Alis, 3.5 NBA leagues, or build 100 Apple Towers. Park, buying 140 billion cups of coffee is only enough for OpenAI to build one Stargate data center.

But this may just be the beginning ; its ambition may be ten times or even more.

In addition to OpenAI , tech giants such as xAI and Meta have also started to spend money like crazy on AI data centers, starting a global infrastructure frenzy and betting on a new trillion-dollar market.

But behind the frenzy, we can't help but ask where all this money is going.

Hello everyone, welcome to "Silicon Valley 101".

I am Chen Qian.

In this video, we will take a look at the capital expenditure behind AI data centers.

What are the components of a data center?

Who are the main upstream and downstream companies and players?

How exactly is the money spent ? Interestingly, after reviewing various reports, we

? Interestingly, after reviewing various reports, we found that everyone's budget is different.

So who is right?

Some data centers have even been "forced" to go to space.

Why?

Why is capital still pouring in when AI is being questioned as a bubble ? Let's first look at the cost analysis of next-generation AI data centers by Bank of America on October 15 this year.

We divide the expenditure of data centers into four main categories : IT equipment, power supply equipment, cooling equipment, and engineering construction . For easier comparison,

. For easier comparison, we'll standardize the unit of measurement to expenditure per GW.

First, let's look at IT equipment directly related to computing , specifically servers, networks, and storage.

Servers account for the largest share , approximately $37.5 billion per GW.

Servers include crucial components like CPUs, GPUs, memory, and motherboards, typically supplied directly by ODMs (Original Design Manufacturers) , such as Foxconn.

These companies obtain server design standards from chip design companies like Nvidia and AMD and manufacture complete systems , supplying directly to hyperscale clients like Oracle, Meta, and Amazon.

ODMs hold a 46% share of the server market.

Other small and medium-sized enterprises (SMEs) need to purchase servers from OEMs (Original Equipment Manufacturers) like Dell, Supermicro, and HP.

In the networking sector, $3.75 billion per GW is needed for network equipment.

Major players include Arista, Cisco, Huawei, and Nvidia.

It's worth noting that although Nvidia only holds a 5% market share , some industry experts argue that while Nvidia's InfiniBand network communication standard is more expensive , its low latency... The advantages of delayed response and no packet loss risk are more suitable for AI data centers.

Finally, there's storage, specifically hard drives.

Each GW requires $1.9 billion in storage equipment.

Major players include Samsung, SK, Micron, and Seagate.

Adding these three items together , we arrive at $43.15 billion per GW for IT equipment —the bulk of data center spending.

While IT equipment is core , the entire data center's support system is equally crucial.

Next, let's look at the cooling system.

First, let me tell you a story: In 2018, an Atlanta data center suffered a cyberattack, forcing the closure of multiple city services, including courts, police stations, and airports.

Besides locking data with ransomware, the attackers also compromised the cooling system.

After the system was compromised, the ambient temperature soared to over 100 degrees Fahrenheit, damaging many chips.

The hackers even held the servers and cooling system hostage, demanding $51,000 in Bitcoin.

Later, attacks on cooling systems became increasingly common and sophisticated.

This story illustrates the critical importance of the cooling system for a data center.

Although the construction budget only accounts for 3% of the total cost, with the exponential upgrade of global AI computing power demand, traditional air cooling technology is struggling to meet the heat dissipation needs of high-density computing equipment. Furthermore

, for NVIDIA GPUs, heat dissipation capacity has become a core bottleneck restricting computing power.

Therefore, for data centers, liquid cooling has gone from an alternative to a necessity . For data centers equipped with liquid cooling systems,

. For data centers equipped with liquid cooling systems, the cooling equipment mainly includes cooling towers, chillers, CDUs (cooling distribution units), and CRAHs (data center air handling units).

To handle 1GW of cooling, these would require expenditures of $0.9 billion, $3.6 billion, $4.5 billion, and $5.75 billion respectively , totaling $1.475 billion.

Major suppliers are numerous and scattered across various stages, so we won't list them all but Vertiv, Johnson Controls, Stuttgart, and Schneider Electric are all major players in this field.

Having discussed cooling and power as core infrastructure, let's look at the power supply equipment.

Power supply equipment mainly includes emergency power supply, backup diesel generators, power distribution control switching equipment, and UPS (Uninterruptible Power Supply) to ensure uninterrupted power supply. The busbars and other power distribution equipment that supply power to each generator unit (Power Supply)

supply. The busbars and other power distribution equipment that supply power to each generator unit (Power Supply) are estimated by Bank of America to cost $400,000 to $550,000 per MW for a typical diesel generator.

Fuel tanks, fuel pumps, and installation costs combined are approximately $350,000 to $500,000 , making the cost of a generator per MW approximately $800,000.

Providing 1GW of power would require $800 million in emergency generators.

However, according to our guests, the actual cost is far higher due to redundancy.

If your IT (equipment) capacity (power consumption) reaches, for example, 1GW, you often need more than 1GW of diesel generators.

There are several reasons for this , but the most important is ensuring redundancy.

For some data centers with particularly high reliability requirements, the number of diesel generators may be twice the data center's computing power.

For example, a 1GW data center might require 2GW of diesel generators.

The biggest players in the diesel generator market are Caterpillar , Cummins, and Rolls-Royce, which are roughly equal . In addition, a 1GW data center also requires $615 million in switchgear, $ 985 million in UPS systems, and $300 million in power distribution equipment.

The three major players in these electrical equipment are Schneider Electric, Vertiv, and Eaton.

Therefore, the total cost of the power supply infrastructure is calculated per GW... $2.7 billion is only 1/13th of the cost of IT equipment , which seems quite cheap.

Although power supply is free , it has become a core bottleneck for many data centers in the United States.

We will discuss this in more detail later in the video.

Let's first draw the complete pie chart of this expenditure.

The last item is construction costs. Construction costs include

costs. Construction costs include building costs, installation costs, general contractor fees, etc. The estimated cost for each GW of construction is approximately $4.28 billion . We calculated

. We calculated that the total expenditure for building a 1GW data center is approximately $51.6 billion , with IT equipment accounting for the highest proportion of the cost, reaching 84%.

Therefore, OpenAI's 10GW Stargate project would cost $516 billion.

This is very close to the officially announced $500 billion investment.

However, while reviewing various research reports, we found something interesting: the data provided by different institutions varies greatly.

For example, taking Stargate as an example, the total budget estimated by different institutions differs by as much as $200 billion.

Why is this?

How should we interpret these calculation discrepancies ? Let's elaborate a bit more.

? Let's elaborate a bit more.

First, let's look at the predictions from several different institutions.

Bernstein's report released on November 1st stated that the cost of each GW of AI data center is approximately $35 billion , and the expenditure breakdown for each item also differs from Bank of America's prediction.

For instance, the total expenditure for IT equipment related to GPUs, networks, CPUs, and storage is 56%, far lower than the 84% calculated by Bank of America.

Barclays Bank 's report at the end of October indicated that the expenditure for each GW of AI data center is $50-60 billion , with 65-70% used for computing and networking.

And in August of this year, Morgan... Stanley's research model

Morgan... Stanley's research model estimates a cost of $33.5 billion per GW , with computing equipment accounting for 41% and the remaining 59% allocated to infrastructure such as power and cooling.

Why then are the predictions so disparate?

There are two main reasons.

First, the assumptions differ.

Bank of America's calculations focus on Nvidia's Rubin architecture , released in early September , which is expected to be available by the end of 2026.

Bernstein and Morgan Stanley's calculations focus on the Blackwell architecture, to be released in March 2024.

Bernstein's GPU costs $13.65 billion (13.65 billion USD), while BofA's (Bank of America) Future (future data center) costs $37.5 billion (37.5 billion USD).

This difference alone is over 20 billion USD , which I feel is the biggest difference . Therefore, the biggest difference

. Therefore, the biggest difference in the calculated amounts lies in the different chip prices, resulting in a $20 billion difference per GW.

The costs of other facilities like power supply and cooling don't differ significantly.

However, this also suggests that Jensen's next-generation chips will be more expensive, a new " Jensen's Mathematics." Math

Jensen's Mathematics." Math estimates that the total cost of a 1GW AI data center is between $600 and $800 billion , even higher than predictions from other institutions. The

"computing cost," which represents Nvidia's potential revenue, is approximately $ 400 to $500 billion . I think Nvidia's assessment is probably accurate

. I think Nvidia's assessment is probably accurate ; they know what price they want to set, right?

They've calculated the energy consumption, the price, and how much profit they expect to make.

The second reason is the difference in the scope of calculation.

BofA (Bank of America) calculates the cost of the data center building itself , while Bernstein calculates the cost of the entire data center campus, including the power distribution system and generators.

I feel that BofA's generators are primarily backup diesel generators, while Bernstein's generators are gas generators.

Turbines are essentially self-generating generators .

Therefore, considering all factors, our guests believe that the budget provided by Bank of America for the future data centers built by tech giants is closer to the reality.

So, this episode's estimate is based on the Bank of America report. As

we mentioned earlier, electricity will become a bottleneck for data centers.

This is why you see generators both inside and outside the data centers in our animation.

There will be a significant hidden expense in electricity investment . Let's continue discussing this.

. Let's continue discussing this.

We previously did a video and podcast discussing the power shortages caused by AI and why the US is so short of electricity . A year and a half later,

. A year and a half later, the situation hasn't improved.

Now, tech giants have to invest in building their own power plants to obtain electricity. OpenAI or its partners like Oracle need to find ways to create new capacity.

Many tech companies now ... They have to build their own generators, power plants,

... They have to build their own generators, power plants, substations, and distribution network facilities , and even some shorter power transmission lines, to meet their own needs.

If you want to equip a 10GW data center with a power plant, the cost could reach $120-200 billion.

So we also see that the large-scale infrastructure construction of AI data centers has boosted the power stocks , which were considered a "sunset industry."

Their gas turbine orders are even booked up for three years.

Google once spent $3 billion to upgrade two hydroelectric power plants in Pennsylvania in order to obtain 3,000MW of power.

This means that it would cost $1 billion to acquire 1GW.

And this is just the cost of upgrading . Musk

. Musk also acquired a power plant for the Colossus 2 project.

Data centers are fiercely competing for power.

Some analysts believe that companies like GE have the ability to pay a premium (to purchase) 1GW (generators), which may be equivalent to $2.5 billion (US dollars).

$2.5 billion is a potentially high figure.

I think if you use $2 billion, it'll be around the CapEx (capital expenditure).

Here's the problem: data centers already have emergency generators, so why not just use them for power ?

There's a fundamental difference between emergency diesel generators and large natural gas turbine generators.

Diesel generators are primarily used as backup power, so all their components are optimized for high power and short bursts.

Therefore, they can't handle 24/7 operation.

Natural gas turbine generators, on the other hand, are designed for scenarios like running almost every hour of the year.

There are other reasons as well.

Diesel is a relatively expensive fuel, while natural gas generators use natural gas , which can be piped . The cost of generating the same amount of electricity from a diesel generator is likely 3 to 8 times that of a natural gas generator.

So, the construction of data centers is currently stuck on obtaining power.

The US power grid can't provide enough power, and natural gas turbine generators are unavailable.

This has spurred the development of other methods , such as fuel cells, which are becoming increasingly popular.

Even tech giants are being "forced" into space.

Google's latest news is that they plan to send a data center into space in 2027.

The main reason is that using solar panels in space can generate electricity at eight times the efficiency on Earth , and it can also solve the problem of no solar energy at night, which is essentially a free and unlimited energy supply.

In addition, vacuum and radiation heat dissipation in space can reduce the need for cooling systems. In addition to Google, Microsoft, Amazon , and Musk's SpaceX , all have started exploring this area.

So how much does it cost to build a space data center?

On LinkedIn, we saw that some people predict that the cost of building a 1MW space data center , including launch costs, is about $35.5 million.

If it is 1GW, it will be $35.5 billion . However, this does not seem to be much more expensive than on Earth.

. However, this does not seem to be much more expensive than on Earth.

We plan to make a separate video to discuss the challenges and opportunities behind the feasibility of space data centers.

Don't forget to follow us.

Since building AI data centers is so costly , and the market is full of doubts and doubts about a bubble, why does this infrastructure boom only increase?

According to the guests we interviewed, there are two main reasons . First, underinvestment is riskier than overinvestment.

. First, underinvestment is riskier than overinvestment.

Most companies now realize that underinvestment is riskier than overinvestment.

The risks of insufficient investment far outweigh those of overinvestment. Why? Because whoever acquires the best AI model or AGI first will likely capture a large market share, rapidly shrinking the survival space of other companies. Now let's look at the risks of overinvestment. You

've essentially bought more land , more electricity , and more buildings to build data centers.

In the end, you might find that you've overspent , but you can use it for internal efficiency improvements , lease it out , or sell the land and electricity to other companies.

Overall, the risks of overinvestment have a ceiling.

The second reason is that as long as there's computing power , tech companies will always find a way to utilize it.

There's a saying in Silicon Valley : "Bill will always eat Andy.

" As long as you have infrastructure hardware and servers, there will always be a way to use it.

Earlier this week, OCP... Meta's representatives stated that their current GPUs already require significant computing power just for internal AI tasks like filtering inappropriate content for Instagram or Facebook . Even with surplus idle computing power

. Even with surplus idle computing power , it could be fully utilized for internal cost reduction.

Therefore, I believe mainstream companies aren't worried about overinvestment.

We see that even with market skepticism about AI overinvestment, giants are still investing heavily .

The final question is, where does the trillion-dollar demand come from?

It comes from the self-generated money of these hyperscale cloud service providers.

They reinvest their profits and debts.

Ultimately, it relies on the public market— the US investment-grade or high-yield bond market—and the emerging shadow banking system , also known as private credit.

These large financing channels support the entire AI build-up . The infrastructure boom isn't unprecedented in American history.

I think AI is more like a large-scale global infrastructure cycle . As long as you (AI) can make money,

. As long as you (AI) can make money, you're a global growth driver, so you don't really need to worry about money.

This seemingly crazy investment is essentially a game of "who gets to the future first."

This path may be risky , but for tech giants, the cost of "absence" is higher than the cost of "investing wrongly."

Okay, that's all for this video.

I'm Chen Qian, co-founder of "Silicon Valley 101."

Your likes, comments, and shares are the best motivation for us to create in-depth technology and business content.

See you in the next video! Bye!