Meta's 14GW ambition and Blackstone's abandoned 22 months: AI infrastructure shifts to competing on resources and speed

[[**Meta plans to invest $145 billion to deploy 14GW of computing capacity, but Blackstone abandoned the world's largest data center project due to timing and policy risks. AI infrastructure has shifted from "comparing capital" to "comparing resources and speed."**

**Key Points**

**1. The AI infrastructure bottleneck has shifted from chips to power and time**

57% of data center projects worldwide are delayed by more than 3 months; compressing the cycle can save up to 22 months. Chips can be used as soon as they are bought, but data center construction involves land approval, community litigation, and power infrastructure. Any delay in any link will turn computing assets into invalid cash flow.

**2. Blackstone abandons a costly flagship project**

Blackstone's Digital Gateway project, which persisted for 22 months, ultimately failed due to resident opposition and litigation. Even with a $165 billion data center platform, it could not withstand time costs and policy risks, indicating that Wall Street's requirements for infrastructure return efficiency have risen significantly.

**3. Data centers begin building their own power plants and supporting generation**

In the South Korea investment plan announced by Trump, a 6.47GW natural gas power generation project in Texas will supply power to the AI park and feed back into the grid. In the past, land was found first and then power; now developers choose to build their own power plants. AI infrastructure has changed from a pure power consumer to a power supplier.

**4. Capital favors only "hard projects" with mature conditions**

Oracle's Jupiter project serving OpenAI was also forced to delay due to power issues. Only projects with ready power, customer contracts, and construction conditions can continue to receive investment. Projects with only land and grand plans are rapidly losing capital patience. 2026-10-07 18:02 Meta's 14GW ambition and Blackstone's abandoned 22 months Cui Peng CP Quick Overview

This article is from the WeChat public account: Cui Peng CP, author: Clooud

Over the past two years, Nvidia's chips have been in short supply, Micron rode HBM to a rocket-like stock price, and capital is very concerned about what will happen next.

Now the computing power frenzy is surging, and tech companies represented by Meta are all scrambling for computing power at all costs. This giant is expected to invest $145 billion in AI infrastructure this year, with an installed scale of 14GW (gigawatts), equivalent to the power generation capacity of more than a dozen large nuclear power plants.

Not long ago, Blackstone Group, which had been investing madly in data centers, abandoned Digital Gateway, the world's largest data center project. Blackstone had persisted with this project for 22 months, but ultimately yielded to a massive wave of public opposition.

The era when you could just go big and fast as long as you had money is over. Welcome to a new era of competing for resources.

Chips can be used as soon as they are bought, but data centers cannot

Meta is also adjusting its chip mix and increasing CPU procurement. It believes that with the development of agents, pure GPU training can no longer meet demand, and more work requires CPUs to handle continuous tasks.

In March this year, Meta announced joint development of a new generation of data center CPUs with Arm. In April, it announced cooperation with AWS to incorporate tens of millions of Graviton CPU cores into Meta's computing resources.

In July, Reuters cited informed sources as reporting that Meta may invest $145 billion in AI infrastructure in 2026, deploying infrastructure with a power scale of about 14GW.

This scale is no longer the traditional sense of buying a few servers; behind it is a complex large-scale industrial infrastructure system. After all, chips and servers can be used immediately after purchase, but data centers cannot be delivered as soon as an order is placed.

The star project Digital Gateway, which Blackstone was forced to abandon, is a typical case.

The data center project was first approved by the Prince William County government in December 2023 and was planned for construction near Gainesville, Virginia. The final planned area of the project exceeds 22 million square feet and involves about 2,000 acres of land.

Local residents and historical preservation organizations successively filed lawsuits against it. This year, the Virginia Court of Appeals ruled that the local government did not meet statutory public notice requirements during the project approval process.

Subsequently, the two co-developers successively withdrew from the litigation, leaving only QTS still appealing in an attempt to save the project. Until early July, QTS decided to withdraw its appeal to the state supreme court and announced the termination of the project.

(Blackstone's flagship project ultimately failed. Image source: Prince William government official website)

It should be noted that Blackstone has always been a very aggressive investor in data centers. It disclosed in the middle of this year that the scale of its global data center platform had reached about $165 billion.

This star project, which was dragged back and forth for 22 months, was reluctantly abandoned by Blackstone, showing that Wall Street is increasingly concerned about the return efficiency of infrastructure investment. AI competition is no longer limited to chips and storage.

At the beginning, what everyone lacked was GPUs, and advanced packaging, HBM, and servers successively became market favorites. Now there are many projects on the market that lack neither customers, capital, nor chips, but what they lack most is time and power.

Because the construction cycle of data centers itself is a very big problem.

A report released by JLL, a leading global commercial real estate investment management company, shows that in 2025, 57% of data center projects worldwide will be delayed by more than 3 months.

In the construction plans given by JLL and InfraPartners this year, through prefabrication (80% of power supply, cooling, and IT modules) and parallel survey and construction, the construction cycle of traditional hyperscale data centers can be compressed by up to 22 months.

If a project is delayed for one or two years due to land approval or community litigation, like the Digital Gateway abandoned by Blackstone, then computing resources that could have quickly generated revenue can only become invalid assets that occupy cash flow for a long time.

Whoever can lock in land in advance, whoever can compress the construction cycle, whoever can obtain enough power, can truly turn chips into computing power.

(U.S. electricity demand is surging. Image source: Blackstone official website)

This change has begun to be reflected in the latest U.S. energy investment plan.

On September 30, Trump announced a group of large energy projects in South Korea's strategic investment plan for the United States, with a total investment of up to about $200 billion. About $120 billion of that will be used to build 8 large nuclear power plants, more than $50 billion will be invested in an Alaska LNG project, and a natural gas power generation project located in Texas.

Trump called this group of investments "one of the largest energy infrastructure investments in American history."

The highly watched Texas project will begin commercial operations in 2029. It has 6.47GW of natural gas power generation capacity and is responsible for supplying power to the surrounding infrastructure park, with surplus power transmitted back to the grid.

The most direct change brought to the entire industry is that in the past, building data centers required finding land first and then power; now developers choose to build their own power plants.

Behind this change are long-term energy investments such as power stations and transmission lines, causing some capital expenditures that originally belonged to tech companies to be transmitted to the entire energy infrastructure system.

Not every data center can survive

But on the supply side, capital investment in data centers has begun to be "picky." Even if a project has potential returns of tens of billions of dollars, it may lose economic viability due to time costs and policy risks.

In September this year, the Jupiter data center project cooperated by Oracle and Blue Owl was affected by power supply issues. Reuters reported that Oracle had issued a force majeure notice regarding the project delay.

This project originally served OpenAI and is also a star case in the current AI infrastructure financing market. Even it has encountered construction schedule problems, showing that the market's divergence trend is already very obvious.

Only projects with power, customers, long-term contracts, and mature construction conditions can continue to obtain capital investment. For projects that have only land and grand plans but cannot quickly resolve power approval and construction cycle issues, capital patience will only become less and less.

After all, AI data centers are standard heavy-asset infrastructure, often requiring tens or even billions of dollars in upfront investment, and also facing a series of issues such as chip iteration speed, customer contract terms, electricity prices, interest rates, and equipment depreciation. Problems in any link will affect the return efficiency of capital investment.

10 articles published on Huxiu. This content is authorized by the author for publication. The views represent only the author and do not represent Huxiu's position. If you have objections or complaints about this manuscript, please contact tougao@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]]]