A single AI factory requires as many as 1.5 million GPUs: NVIDIA sees a $600 billion massive business opportunity.

How capital-intensive is AI exactly?

In the past, we measured computing power by "how many GPUs" were deployed, but now OpenAI and NVIDIA have started using another unit: Gigawatt.

In the PORTS-Pike project disclosed on August 17, OpenAI will build world-class AI infrastructure in Ohio, USA. OpenAI officially revealed that the company has locked in approximately 8GW of IT capacity for the park, where NVIDIA will provide credit support for the initial 4.25GW of related construction, and the entire park will be dedicated to deploying NVIDIA AI computing infrastructure.

According to the calculation in NVIDIA's press release, in the initial 4.25GW deployment, each generation of NVIDIA AI Factory system may correspond to a scale of about 1.5 million GPUs, which may represent a revenue opportunity of 150-200 billion US dollars for NVIDIA.

If future cooperation is further expanded, the related computing power scale of OpenAI and NVIDIA is expected to advance to about 16GW.

The figures given by NVIDIA are even more staggering:

By 2030, the corresponding NVIDIA computing opportunity may reach about 600 billion US dollars.

Note that this does not mean that OpenAI has signed a 600 billion US dollar GPU order today, but a long-term market opportunity calculated by NVIDIA based on the potential computing power expansion scale.

Even so, this figure still reveals a very important signal:

The truly capital-intensive era of AI may have only just begun.

What does 1.5 million GPUs mean?

Don't rush to look at the 600 billion US dollar figure first.

The 1.5 million GPUs themselves are already astonishing enough.

Many investors already consider AI clusters with 100,000 or 200,000 cards to be extremely large in scale.

But now the industry is starting to discuss million-level GPU infrastructure.

And special attention needs to be paid here:

It does not mean that 1.5 million GPUs are installed permanently at one time and the work is done.

The PORTS-Pike project has a planned cycle of up to 20 years, and multiple generations of GPU upgrades will be experienced in the future.

In other words, the current generation may be Blackwell or Rubin, and the next-generation architecture will emerge in a few years.

AI data centers are not built once and used for 20 years, but may go through an expensive "heart replacement" every few years.

This is also why NVIDIA estimates the long-term revenue opportunity corresponding to a 4.25GW project to be 150-200 billion US dollars.

What is really valuable is not just the first sale of GPUs.

Instead:

Continuous upgrades of the first generation, second generation, third generation, and so on.

This is somewhat similar to the previous smartphone replacement cycle, except that what is being replaced now is not a mobile phone worth thousands of yuan, but AI infrastructure worth tens of billions of US dollars.

Why do AI factories start to be measured in "gigawatts"?

Because there are so many GPUs that we can no longer only look at the number of chips.

What does 1GW mean?

Simply understood, it is a power load of 1 million kilowatts.

The approximately 8GW of IT capacity locked in by OpenAI in PORTS-Pike is already close to the power consumption scale of a super-large industrial base. The first batch of about 800MW is expected to be ready for use in 2028, after which new power generation facilities, transmission lines and supporting infrastructure will need to be constructed.

This shows that a very important change is taking place in the AI industry.

In the past, when an Internet company expanded, the core question was:

Are there enough servers?

Now when AI companies expand, the first question to ask is:

Where does the electricity come from?

Then it comes to:

Are there enough GPUs?

Are there enough transformers?

Are there enough optical modules?

Can the cooling system withstand the load?

Can the power grid deliver so much electricity?

Therefore, the AI industry is increasingly resembling heavy industry.

The model is written in code, but the means of production behind the model have become:

Land + power plant + power grid + GPU + HBM + network + refrigeration.

This is also why NVIDIA likes to use the term "AI Factory".

It really looks more and more like a factory.

The only difference is that traditional factories produce automobiles, steel and chemical products, while AI factories produce Tokens and intelligence.

Why does NVIDIA dare to target 600 billion US dollars?

Because OpenAI's appetite for computing power is still expanding.

As early as September 2025, OpenAI and NVIDIA announced a strategic cooperation with the goal of deploying at least 10GW of NVIDIA systems, corresponding to millions of GPUs; NVIDIA also stated that it will invest up to 100 billion US dollars in OpenAI as the computing power is gradually implemented.

Now PORTS-Pike has further expanded the physical infrastructure footprint.

OpenAI officially revealed that this Ohio project is expected to create about 35,000 construction jobs and 2,500 long-term operation jobs during the six-year construction period; the data center will be constructed, owned and operated by SB Energy, which will provide capacity to OpenAI through a 20-year lease.

NVIDIA also announced on August 17 that it will invest 1.5 billion US dollars in SB Energy.

You will find that NVIDIA's role has quietly changed.

In the past, it only sold GPUs.

Now it is deeply involved in more areas:

Chip supply, networking, software, data center design, capital investment, and even project financing support.

Why?

Because if AI computing power really becomes a market of hundreds of billions of US dollars in the future, for NVIDIA, the most important thing is no longer just "let customers buy my GPUs".

Instead:

Ensure that customers really have the money, electricity and data centers to deploy the GPUs they purchased.

This is the sign that competition in AI infrastructure has entered the next stage.

The biggest winners may not only be GPU vendors

When many investors see the 600 billion US dollar figure, their first reaction may be:

How much more money can NVIDIA make?

But if 10GW or even larger AI computing clusters are really built, what is truly driven will never be limited to just one company.

The first layer is of course GPUs

It is the most expensive core equipment of the entire AI factory.

Training models requires GPUs, inference requires GPUs, and scaling up new models still requires GPUs.

As the demand for computing power continues to rise, GPUs are naturally at the core of the industrial chain.

The second layer is HBM and advanced packaging

GPUs do not work in isolation.

High-performance AI chips require a large amount of HBM high-bandwidth memory, and advanced packaging to efficiently connect GPUs and storage.

When the number of GPUs reaches the million level, the consumption of HBM is also an astronomical figure.

That is why storage is receiving increasing attention from the market in this round of AI boom?

The answer lies right here.

AI not only consumes computing power, but also devours memory at a frantic pace.

The third layer is high-speed networking and optical communications

It is easy to connect a few GPUs.

How to connect 1.5 million GPUs?

This is no longer an ordinary network problem.

High-speed data exchange is required between GPUs, and interconnection is also required between data centers.

Therefore, switches, optical modules, optical fibers, and high-speed network chips will all grow as the scale of the cluster expands.

The fourth layer is becoming more and more critical:

Electricity.

Without electricity, all GPUs are expensive iron boxes.

The biggest bottleneck for AI in the future may not be chips, but electricity

The most noteworthy sentence in the PORTS-Pike project is actually not the 1.5 million GPUs.

Instead, OpenAI officially clearly stated:

The initial 800MW can mainly utilize existing power infrastructure, but subsequent expansion will require the construction of new power generation facilities and transmission lines, including natural gas power generation.

This sentence is extremely valuable.

In the past, when people discussed AI investment opportunities, the first thing that came to mind was semiconductors.

But when the project enters the multi-GW level, the constraints begin to change.

Chips can be purchased.

But power plants cannot be put into operation the day after you place an order.

High-voltage transformers are not supplied infinitely.

The construction of transmission lines takes time.

Data center site selection also needs to consider water, electricity, land and approval procedures.

Therefore, the truly scarce resource for AI computing power in the future is likely to spread further upstream from GPUs:

Whoever can provide stable, low-cost large-scale electricity will be qualified to build the next generation of AI factories.

This is also why US technology companies have been discussing natural gas, nuclear power, energy storage and power grids more and more frequently in recent years.

AI has finally circled back to one of the oldest businesses:

Energy.

What is truly worth guarding against in the 600 billion US dollar figure is "computing power depreciation"

Of course, this frenzy is not without risks.

The biggest risk comes precisely from the extremely fast iteration of AI chips.

A traditional power plant may operate for 30 years.

An office building can be used for decades.

But today's most advanced GPUs may be significantly surpassed by next-generation products in a few years.

This means that AI data centers face a very special problem:

The asset scale is huge, but the core equipment depreciates extremely fast.

If the energy efficiency and performance of next-generation chips are greatly improved three or four years later, the GPUs purchased for tens of billions of US dollars today must be considered for upgrade.

For NVIDIA, this is a good thing.

Continuous product replacement by customers means continuous revenue.

But for OpenAI and data center investors, this means a very realistic problem:

The revenue earned must be sufficient to cover the huge capital expenditure.

OpenAI previously estimated earlier this year that cumulative computing expenditure may reach about 600 billion US dollars by 2030, while the company needs to rely on revenue and financing to support continuous expansion.

Therefore, the AI industry must ultimately answer a question:

Can the Tokens created by these GPUs really generate enough revenue?

This is the core of whether the 600 billion US dollar story can really come true.

The AI investment logic has changed

In the past two years, the capital market's favorite question was:

Whose large model is the best?

In the future, investors will increasingly need to ask several other questions:

Who owns the most electricity?

Who can build the largest-scale data center?

Who controls GPUs, HBM and networks?

Who can truly operate an AI factory?

This shows that AI competition is rapidly escalating from a "software war" to a war of industrial infrastructure.

OpenAI locking in about 8GW of capacity in PORTS-Pike and NVIDIA entering the investment and financing system both point to the same trend:

Computing power is becoming the infrastructure of the new era.

Just like in the 20th century, the competition revolved around railways, power grids, highways and oil.

What the AI era in the 21st century is starting to compete for may be:

GPUs, electricity, data centers and high-speed networks.

The biggest enlightenment of the 600 billion US dollar figure: don't only focus on NVIDIA

So why does an AI factory have a number as high as 1.5 million GPUs?

Because AI has gone beyond the stage of "buying thousands of cards to train a model".

What leading companies are building now is global computing power infrastructure that can continuously train, continuously perform inference, and serve hundreds of millions or even more users.

The investment chain formed behind this is very long:

HBM comes after GPU;

Advanced packaging comes after HBM;

Optical modules come after servers;

Liquid cooling comes after data centers;

Power equipment comes after liquid cooling;

Even further back are natural gas, nuclear power, power grids and energy storage.

Therefore, what is really worthy of investors' attention is not only:

Can NVIDIA capture the 600 billion US dollar market.

Instead, if AI infrastructure really evolves from 10GW to a higher level, how much electricity, networks, servers and data centers will the world need to rebuild?

This may be the most astonishing part of this round of AI capital expenditure.

In the past, we thought AI was a software revolution.

Now it is increasingly obvious that:

AI may eventually be an infrastructure construction wave on a scale comparable to the energy and industrial revolutions.

The model determines what AI can achieve.

But what ultimately determines how far the model can run are those most basic things:

Chips, electricity, land, networks and capital.

And the million-GPU plan of PORTS-Pike has for the first time presented the scale of this computing power arms race more intuitively to everyone.

References:

1. Cls.cn: Report on the expansion of potential computing opportunities between OpenAI and NVIDIA

2. OpenAI: Official announcement of the PORTS-Pike project

3. NVIDIA: 10GW strategic cooperation announcement between OpenAI and NVIDIA

4. Reuters: Report on OpenAI's computing power expenditure and related developments of NVIDIA and SB Energy

This article is only an analysis of the technology industry and business logic, and does not constitute any investment advice.