Bùng nổ AI trị giá 3.000 tỷ USD đang kéo theo làn sóng xây dựng trung tâm dữ liệu toàn cầu chưa từng có

• Các ông lớn công nghệ như Meta, OpenAI và xAI đang triển khai các dự án siêu máy tính AI trị giá trên 100 tỷ USD mỗi dự án như “Stargate”, “Colossus”, và “Prometheus” – chỉ là phần nhỏ trong cuộc bùng nổ đầu tư trị giá 3.000 tỷ USD để xây dựng trung tâm dữ liệu toàn cầu đến năm 2029.

• Chỉ riêng năm 2026, Google, Amazon, Microsoft và Meta dự kiến sẽ chi hơn 400 tỷ USD cho trung tâm dữ liệu, vượt mức 350 tỷ USD của năm 2025.

• Nhưng sức mạnh tài chính của Big Tech cũng đang bị thử thách. Với doanh thu AI tạo sinh năm 2024 chỉ đạt 45 tỷ USD, nhiều nhà đầu tư bắt đầu đặt câu hỏi về lợi nhuận thực sự từ cuộc đua hạ tầng này.

• Khoảng trống 1.500 tỷ USD vốn đầu tư (trên tổng 3.000 tỷ) sẽ phải đến từ bên ngoài Big Tech – gồm quỹ đầu tư tư nhân, nợ vay ngân hàng, trái phiếu, và vốn chủ sở hữu từ các tổ chức như Blackstone, KKR, Apollo...

• Meta đã huy động 29 tỷ USD (trong đó có 26 tỷ USD nợ) từ nhóm nhà đầu tư do Pimco dẫn đầu để xây dựng trung tâm dữ liệu tại Ohio và Louisiana.

• Oracle không tự xây dựng mà thuê 2GW trung tâm dữ liệu tại Texas từ startup Crusoe, được tài trợ bởi 5 tỷ USD vốn chủ sở hữu và 10 tỷ USD vay từ JPMorgan, phục vụ hợp đồng trị giá 30 tỷ USD/năm với OpenAI.

• Mô hình “build-to-suit” đang phổ biến: bên phát triển chịu rủi ro đầu tư, còn Big Tech chỉ cam kết thuê dài hạn – nhấn mạnh vai trò ngày càng lớn của nợ vay trong lĩnh vực này.

• CoreWeave, khởi nghiệp từ khai thác tiền số, hiện trị giá 65 tỷ USD, là ví dụ điển hình cho xu hướng mới: huy động 10 tỷ USD từ Blackstone, thế chấp bằng GPU và hợp đồng cho Microsoft thuê sức mạnh tính toán.

• Tuy nhiên, nhiều chuyên gia cảnh báo rủi ro:

  • Công nghệ chip AI lỗi thời (đặc biệt là khi Nvidia liên tục ra thế hệ mới).

  • Hệ thống làm mát trung tâm dữ liệu có thể nhanh chóng lạc hậu.

  • Một số trung tâm dữ liệu có thể trở thành “nhà kho không dùng được” trong 10 năm tới nếu nhu cầu AI suy giảm.

• Các khoản nợ đang đổ vào các startup hoặc nhà phát triển chưa có khách thuê chính thức, làm tăng rủi ro vỡ nợ. Một số nhà đầu tư đã bắt đầu tránh các thương vụ chứng khoán hóa tài sản trung tâm dữ liệu vì lo ngại giá trị tài sản sẽ biến mất trước khi trái phiếu đáo hạn.

• Câu hỏi lớn: nếu doanh nghiệp không chi trả đủ cho dịch vụ AI như kỳ vọng, toàn bộ cấu trúc tài chính này có thể sụp đổ theo mô hình "dot-com" hoặc “bong bóng viễn thông” cuối thập niên 1990.

• Các nhà phát triển trung tâm dữ liệu đang gánh nợ cao và phụ thuộc vào mức thuê hiện tại, nhưng nếu chu kỳ sụt giảm xảy ra, họ sẽ là bên tổn thất đầu tiên – chứ không phải các tập đoàn lớn.

📌 Cơn sốt AI tạo ra một cuộc đua cơ sở hạ tầng trị giá 3.000 tỷ USD, với Big Tech và giới tài chính tư nhân ồ ạt rót vốn vào trung tâm dữ liệu và siêu máy tính. Nhưng rủi ro cũng tăng theo: công nghệ lỗi thời, chi phí tăng vọt, dư thừa công suất và mô hình tài chính dựa vào giả định rằng “AI sẽ dùng cho mọi thứ”. Sự bùng nổ này có thể trở thành bong bóng nếu nhu cầu không theo kịp kỳ vọng.

https://www.ft.com/content/efe1e350-62c6-4aa0-a833-f6da01265473

‘Absolutely immense’: the companies on the hook for the $3tn AI building boom
Private capital joins Big Tech in seeking to capture rewards from historic expansion of data centres

© FT montage/The Washington Post/Getty Images

Tabby Kinder in San Francisco
Published17 hours ago

Meta is building “Prometheus” and “Hyperion”, Elon Musk’s xAI has “Colossus”, and OpenAI is developing “Stargate” — each a more than $100bn project to build the world’s most powerful supercomputer and usher in a new generation of artificial intelligence.
But each of those gargantuan ventures is just a fraction of the spending required to build the data centres needed to power the AI era: one of the biggest movements of capital in modern history.
“The amount of capital required is absolutely immense,” said Rob Horn, global head of infrastructure and asset-based credit at private equity group Blackstone, which manages an $85bn data centre platform.
“The scale of the opportunity is exhausting the capital of [any one financial] market, and is requiring an all-of-the-above approach, with private capital playing a very large role.”
Google, Amazon, Microsoft and Meta will spend more than $400bn on data centres in 2026 — on top of more than $350bn this year.
For years, Big Tech’s capital spending grew steadily, focused on cloud, logistics and underlying infrastructure
Then came the launch of ChatGPT in late 2022 and with it a realisation that AI could upend their businesses
Now begins an arms race, where billions are poured into servers, chips and data centres to power generative AI
But as the money floods in, concerns are being raised about overcapacity, long-term profitability and energy demands.
“Lots of people who are trying to build data centres will fail,” said one banker who helps arrange financing for AI infrastructure projects.
“We are in that period where the capital markets are crazy enough to throw money at almost anything. I am curious to see the next phase and whether rationality prevails.”
Once seen as a niche part of the real estate market, the frenzied pace of construction has turned data centres into a sought-after asset class.
This year is forecast to break records for development. The US has about 20 gigawatts of operational data centre capacity. Before the end of the year, another 10GW of data centres are projected to break ground globally, and 7GW will reach completion, according to real estate group JLL.
Historically, most of the spending by the “hyperscalers” — Amazon Web Services, Microsoft Azure and Google Cloud — building data centres for their cloud services businesses was self-funded.
But the scale of computing power needed for generative AI is changing that.
The AI race

This is the third part in a series exploring the race for AI capacity and the data centres at the heart of billions of dollars in capital investment.
Part 1: Inside the relentless race for AI capacity
Part 2: Can data centres ever truly be green?
Part 3: Financing the data centre boom
While internal cash flows largely covered costs of up to $200bn last year, costs are projected to double this year and increase further next.
Some economists have started to question how much further hyperscalers’ cash reserves can be stretched, and investors want to know when their spending will translate to real revenues from AI services. Hyperscalers’ generative AI revenues were just $45bn last year, according to Morgan Stanley analysts — although they predicted revenues would exceed $1tn by 2028.
This has left a funding chasm that financiers are rushing to fill.
JLL estimates $170bn of assets will require construction lending or permanent financing this year. Between now and 2029, however, global spending on data centres will hit almost $3tn, according to Morgan Stanley analysts. Of that, just $1.4tn is forecast to come from capital expenditure by Big Tech groups, leaving a mammoth $1.5tn of financing required from investors and developers.
The gap will be filled by everything from private equity, venture capital and sovereign wealth to bank loans, publicly listed debt and private credit. But increasingly, the answer is debt.
About $60bn of loans are going into roughly $440bn of data centre development projects this year, twice as much debt as in 2024, according to a recent presentation by law firm Norton Rose Fulbright. More than $25bn of loans were underwritten in the first quarter of this year alone, according to a report by Newmark.
Funding data centres comes not just with the risk that costs overrun, but also that the technology becomes obsolete far quicker than anticipated, requiring new investment that decreases returns for its owner — or forces them to sell at a discount. That means even the deepest-pocketed tech groups may want to share the risk, particularly when debt is cheap and readily available.
Deals are being structured in myriad different ways, from structured debt solutions and project finance vehicles to construction loans, asset-backed securitisations and even green bonds to raise money and start building.
Meta raised $29bn — including $26bn of debt — from private capital investors led by Pimco this month to help fund data centres in Ohio and Louisiana, enabling it to offset high upfront costs and spend its cash on other initiatives with faster returns.
Investors including Apollo, Carlyle, Brookfield and KKR competed in a months-long bidding war to lend to Meta.
Oracle takes a different approach with the 2GW data centre it has signed up to lease in Abilene, Texas. The project is being built by start-up Crusoe and investment group Blue Owl Capital, which have raised about $5bn of equity from investors and borrowed almost $10bn from JPMorgan to fund the construction, backed by Oracle’s 15-year lease.
In turn, Oracle has agreed to provide OpenAI with 4.5GW of computing power — including from Abilene — in a deal worth about $30bn a year, which forms the first part of OpenAI’s Stargate data centre project in the US. Neither Oracle nor OpenAI will carry the debt raised to build the Abilene site on their balance sheet.
This data centre development model, known as “build-to-suit”, is being replicated by tech companies across the US.
A rendering of Meta’s plans for the Ohio Bowling Green Data Center
Meta raised $29bn — including $26bn of debt — from investors led by Pimco to help fund data centres in Ohio, pictured, and Louisiana © Meta
“All of the major hyperscalers have self-build programmes. Where third-party developers can add value is when we have sites that are shovel-ready and can deliver on an accelerated timeframe,” said Tim McGuire, capital markets leader for hyperscale data centre developer Rowan Digital.
He said the process of identifying a site, securing the necessary power and building the infrastructure can be a “three-year plus lifecycle. We can cut that cycle in half.”
To get comfortable with the risk involved in a build-to-suit project, lenders, equity investors and developers require hyperscale tenants to sign long-term leases or capacity commitments before they part with their cash. This means they are in effect lending against the creditworthiness of an investment-grade counterparty such as Microsoft or Oracle, a bet that is prompting a race among private capital providers to offer more and larger loans.
In some cases, it also means acquiring data centre developers themselves. Last year, Blackstone bought Australian data centre platform AirTrunk for $14.9bn, the second-largest data centre deal after KKR and Global Infrastructure Partners’ 2021 purchase of US data centre owner CyrusOne for $15.5bn. Last week, Apollo struck its own deal, buying a majority stake in data centre builder Stream.
Apollo said data centres would require “several trillion dollars of global investment over the next decade”; it has already deployed $38bn into data centre-related infrastructure.
But the scale of capital deployment has turned companies with access to land, power or the specialised computer chips used to power AI data centres into potentially very valuable players — if they can prove themselves capable of delivering results for hyperscale tenants.
The building site for a data centre in Abilene, Texas 
Oracle has signed a 15-year lease for a 2GW data centre in Abilene, Texas, from which it will partly provide OpenAI with 4.5GW of computing power © OpenAI
“We see multiple developers every week who have undeveloped land but think that they will be signing leases with hyperscale customers tomorrow,” said Sam Southall of Macquarie Capital.
“Essentially everyone with some land and a tenuous path to power is trying to raise capital, but there is a long way to go in order to have credibility with, and be trusted by, these types of tenants.”
Chief among those to have made this play is CoreWeave, a small company founded to mine cryptocurrencies in 2017, a function that required the high-performance Nvidia graphics processing units (GPUs) that have become a key element in training AI models. CoreWeave made the pivot to leasing and operating AI data centres years later, then listed its shares on the Nasdaq exchange in March. It is now worth $65bn.
The New Jersey-based company funded its shift to AI with large loans, including about $10bn from Blackstone. Blackstone took security both over CoreWeave’s GPUs — an increasingly popular form of asset-backed financing for AI data centres — and its contracts to lease computing power to Microsoft. Crusoe, which is building Oracle’s Abilene data centre, also started as a crypto-mining company with access to power contracts.
“Data centres are just a fraction of the capital needed,” said Blackstone’s Horn. “If you have a 1GW data centre, it will cost over $10bn, but all of the equipment costs another $30bn plus. There is not just a data centre financing opportunity, but an opportunity around equipment, inventory and supply chain finance.”
Powering AI makes up less than half of data centre demand at the moment, but it is responsible for almost all of the growth.
The pace of development has drawn comparisons with the telecoms bubble in the late 1990s, when companies laid more than 80mn miles of fibre optic cables across the US in a drastic overestimate of the demand required. The glut meant costs plummeted and many companies failed.
“People are making forecasts on the assumption that all enterprises will start to use AI technology and pay for it, and pay enough for it to justify the return on investment for all these training facilities,” said a banker who works on data centre deals.
“The conclusion is that we’re all going to be using AI all the time for everything. That’s an incomprehensible world, but one you need to believe in order to not see how this all ends up losing money.”
Big Tech companies stand to lose the most if forecasts about the potential of AI — and the money to be made — are overcooked. By self-funding and owning a large proportion of their data centre capacity, they take on the capital expenditure, operational risks and regulatory burden.
The AI race

This is the third part in a series exploring the race for AI capacity and the data centres at the heart of billions of dollars in capital investment.
Part 1: Inside the relentless race for AI capacity
Part 2: Can data centres ever truly be green?
Part 3: Financing the data centre boom
If demand for AI plateaus, or it emerges that models such as the Chinese start-up DeepSeek’s can be trained far more cheaply, they will be left with huge stranded assets.
Much of the current spending is on data centres capable of training powerful AI models, but as the technology shifts to inference — running those models — the demand for compute will probably drop and assets may become less valuable. Likewise, if power supplies dry up or AI chip supply chains are delayed, returns on investment may suffer.
“We view cloud services data centre build-outs as fairly robust. We are less confident long-term in the AI training-only locations,” said one executive at a large developer.
Another banker involved in the sector added: “In five years, you’re not getting a lease renewal at anywhere near the rates you have now. There is a massive overestimation of terminal values.”
The flood of capital into the sector also means increasingly speculative projects are securing funding, including sites that lack an anchor tenant. A larger portion of capital is already being allocated to building data centres for non-investment grade tenants, for example CoreWeave and OpenAI, as well as smaller AI start-ups.
Some landlords, such as Blackstone and infrastructure investor DigitalBridge, are increasingly seeking to extract cash from assets once they are fully built with contracts in place by turning to securitisation deals.
But some investors already view that debt as too risky. One large buyer of securitised debts told the Financial Times they were avoiding such deals because of concerns the properties would be obsolete by the time the debt matured, leaving lenders with assets of questionable value.
“People are lending at high loan-to-value ratios and there is obsolescence risk,” the person said.
Loans secured against GPUs are also at the riskier end of the data centre development financing market and are becoming more common after the success of CoreWeave, with more lenders financing former crypto companies that possess the chips. But Nvidia, the largest manufacturer and designer of AI chips, frequently produces faster and higher-performance chips that risk making its older generations obsolete and devaluing lenders’ security.
Obsolescence is also a risk for the data centres themselves. Many are being built to house Nvidia’s latest Blackwell chips, which require complex liquid cooling systems. The technology is changing so fast that it is feasible future chips will require newer cooling methods.
“There’s a risk in 10 years you just have a shed with obsolete GPUs and cooling infrastructure that is unfit for purpose and you may as well start again,” one person in the data centre sector said.
Still, the hyperscalers can afford the risk. They have the economies of scale to protect against at least some losses. The position of those who borrowed large sums to fund the build-out is more perilous.
“Microsoft and Amazon don’t know what they need. They’re just gobbling up everything because popular opinion is that it’s a winner-takes-all market,” said an executive at a data centre leasing group. “The people who are going to suffer from a pullback . . . are the data centre companies who are overleveraged.”