Unprecedented trillion-scale investment—is NVIDIA the only winner?

[[**The global AIDC construction boom is driving up computing power investment; demand is real but profitability is not yet proven; Nvidia's DSX POD shortens deployment, the market worries about returns, and the only certain winner may be Nvidia, which sells the equipment.**

**Key points**

**1. Nvidia uses DSX POD to shorten the plant construction cycle, but the market still isn't buying it**

The solution pre-integrates the Vera Rubin system and its supporting components, with delivery planned for early 2027; deployment can be compressed to 14 days, but power, cooling, and networking must be ready in advance; after GTC, Nvidia gave back its gains over two days.

**2. AIDC investment surges, with much capacity not yet in use**

Total computing power investment in 2026 is expected to be $830 billion, up 79% year on year; as of July 2026, more than 95% of Nvidia's Grace Blackwell capacity had not been put into use, and most of the largest announced projects have not been completed.

**3. Demand is real, but about 30% to 40% is capacity pulled forward**

Industry consensus is that more than about 60% of the giants' capital expenditure corresponds to clear customers; Anthropic and OpenAI revenue is growing fast but they are still loss-making, and real demand does not equal having made money.

**4. The construction boom is hard to stop in the short term; profitability and financing conditions have not yet been verified**

CoreWeave had second-quarter revenue of $2.575 billion but a net loss of $626 million, and Oracle once issued a force majeure notice; whether single-park revenue covers costs, whether utilization and rental prices are stable, and whether financing can continue have not been broadly verified, and alternative nodes may still need three to five years. Original 2026-10-24 13:52 An unprecedented trillion-dollar investment, is Nvidia the only winner? Chen Yifan_YF Quick Overview

Nvidia has already begun thinking for customers about how to help them build AI factories faster. At the Berlin GTC in October, Jensen Huang no longer emphasized CUDA; Nvidia released DSX POD: partners pre-integrate the Vera Rubin system and its supporting facilities into modular containers, and after verification they are shipped to the site, compressing deployment time from six months to as fast as fourteen days. Of course, this 14 days has two conditions: the site's power, cooling, networking and other facilities must be ready in advance; and this solution is planned for delivery in early 2027, so it is still a delivery commitment. Either way, it partly solves the data center deployment problem.

Also selling GPUs, in the past CUDA helped Nvidia keep developers in its ecosystem; today, it also has to organize computing, networking, power supply and distribution, and liquid cooling into a factory that can be replicated. CUDA is still the foundation, but Nvidia wants to control more links.

This is obviously a natural and highly profitable business.

The market has a different view of this business. On the day of GTC, Nvidia closed up 2.5%; the next day the Nasdaq fell 2.6%, the biggest single-day drop since June, and Nvidia itself gave back 4.4%, returning all the gains from the launch event over two trading days. Another backdrop is the sharp fall in global tech stocks.

Selling factories can of course shorten the construction cycle, but it cannot shorten the time the capital market is willing to wait.

Jensen Huang found a manufacturing explanation for this business: input power and equipment, output tokens. This metaphor is easy to understand, and it brings a problem that manufacturing cannot avoid: after the factory is built, how much can the product sell for? Have these builders, who spend huge sums, even borrow, to buy cards and AI containers from Nvidia, actually made money from AIDC? Where will this construction boom stop?

Crazy investment

AIDC, intelligent computing centers, are harder to build than traditional data centers, and the construction logic is different too, from the original multi-site layout to raising the computing power density of a single site.

According to statistics from third-party data consulting firm Epoch.ai, total computing power investment in 2026 reached $830 billion, up 79% year on year, implying an investment scale of about more than $400 billion in 2025. In the single-site computing power ranking, xAI's computing center ranks first. With only two data centers, it entered the top 10 in the global AIDC total computing power ranking.

Global AIDC site computing power ranking

Data source: Epoch.ai

GPUs installed in racks still have a way to go before generating revenue. Storage must continuously supply data, networking must let chips work together, and power supply and cooling must withstand high-density operation; if any one link falls behind, expensive equipment cannot be fully utilized. Moreover, the requirements for operators are higher: with the same set of Huawei equipment, some commercial entities can achieve a power efficiency score of 100, while others only reach 70.

Intelligent computing centers have floor load-bearing requirements 50% to 100% higher than traditional general computing centers, and power demand for the same area exceeds the latter by more than three times. Therefore, traditional data centers are also hard to upgrade into AIDC through renovation.

One problem caused by complex infrastructure construction is higher deployment difficulty and cost. According to the latest "AI Report Computing Index" released by third-party institution Air Street Capital, which tracks major AI computing deployments worldwide, as of July 2026, more than 95% of Nvidia's Grace Blackwell GPU capacity had not yet been put into use. Most of the largest announced projects are still not built. Saudi Arabia's HUMAIN project accounts for about 600,000 of the announced Grace Blackwell GPUs, while projects supported by Stargate Abilene, OpenAI, Oracle and Crusoe account for about 450,000.

This is also why Nvidia launched its DSX POD solution. However, building data centers involves not only deployment difficulty but also approvals for water, power, land and so on.

The most typical case is Oracle. Its Project Jupiter in New Mexico has a total investment of $165 billion, and its main tenant is itself. On September 24, it sent a force majeure notice to developers: if the park cannot come online by the 2028 target, it can delay rent payments.

On the day of the news, Oracle's stock fell 3%. Those who borrow money to build factories have first left themselves a way out. The state government has already rejected this natural gas pipeline twice, and in mid-October it was still asking federal regulators to speed up approval.

Similar cases are more than one. Amazon issued about $54 billion in investment-grade bonds in March this year, Alphabet's free cash flow has turned negative, and at the same time it raised its capital expenditure guidance. According to Forbes statistics, by the end of 2026, AI-related debt issuance has exceeded $500 billion.

How much is real demand?

How much of the crazy AIDC expansion is real demand?

In its October 19 report, SemiAnalysis estimated that global data center new capacity in 2027 will be about 50GW. Its basis includes construction progress tracked building by building, the number of cloud vendors' self-built projects started, and leases already signed.

Behind this round of construction is a financing mechanism. According to SemiAnalysis's observation, when developers have no customers, they usually only invest 5% to 10% of the project capital first; only after obtaining pre-lease agreements are they able to raise subsequent funds. The large leases listed in the report usually have terms of 15 to 20 years, with the total contract value for the full lease period per gigawatt at about $30 billion to $50 billion.

Huxiu has discussed this issue with multiple investors and data center builders: in this round of global AIDC expansion, how much comes from real customer demand? How much is preempting capacity in advance?

The basic consensus is that more than about 60% of giant companies' capital expenditure corresponds to clear customer demand, and the remaining 30% to 40% belongs to advance layout for grabbing output, power quotas, and land.

Computing power and capital expenditure changes

Data source: Epoch.ai

Evidence supporting this judgment is also emerging one after another. In the storage field, companies such as SK Hynix have already signed a large number of long-term agreement orders with downstream customers, locking in demand in advance; the backlog orders disclosed in the financial reports of leading cloud vendors are even scheduled two years out. At the industry level, the demand cycle that can be clearly seen is generally 18 months to 2 years. Whether all orders can ultimately be delivered still has variables, but the feedback from downstream is already clear enough.

Three years after ChatGPT came out, applications have extended from chat to code generation, and both training and inference are consuming computing power. Anthropic's third-quarter revenue was $16.8 billion, up about 45% from the previous quarter, with an annualized rate close to $67 billion; OpenAI's annualized revenue by the end of September was about $50 billion, with a year-end target of $70 billion. These two revenue curves are the most direct support for this round of computing power orders.

But both are still losing money. Anthropic's net loss in the first half of the year was $80.7 billion, of which about $72 billion came from the book valuation of convertible bonds, while computing power expenses in the same period were about $9 billion. Revenue is rising, and losses are rising too. The difference is that revenue is money paid in by customers, while a large part of the loss is only accounting treatment. Real demand does not mean this business is already making money.

Liang Wenfeng said at an investor exchange meeting that DeepSeek's API pricing is anchored at "buying a batch of equipment back from the market and recovering the cost in ten months," corresponding to about six times profit.

McKinsey predicted in early last year that global computing power demand would grow 350% from 2025 to 2030, with AI workloads accounting for more than half of total data center workloads. Judging from the actual situation this year, this forecast is even conservative.

This demand shortage has also spread to the entire supply chain. First high-end GPUs were in short supply, then gaps in storage and optical communications were exposed, and the doping materials thulium, ytterbium, and erbium for optical chips all became bottlenecks; the market size of optical communications was temporarily expanded about tenfold because of this round of demand. Transformers take four years to wait for, and grid connection queues are even longer. Cloud vendors therefore place orders in advance, GPU vendors lock HBM and advanced packaging in advance, and this passes all the way to upstream materials and multilayer PCBs, with every link stocking a little more.

Based on current demand estimates, plus orders visible within two years, this construction wave cannot stop in the short term.

Who has already made money?

Demand is real and growing, but who has actually made money?

There are mainly several business models for AIDC. In the United States, the construction entities are basically commercial companies: large cloud vendors, model vendors, and third-party infrastructure builders. China has a mixed model, with local governments, local state-owned capital platforms, the three major telecom operators, cloud vendors, and small third-party players all building.

The first is the colocation model, mainly local governments, state-owned capital platforms, telecom operators, and third-party infrastructure builders. They hold land, substations, diesel generators, and liquid cooling pipelines, and charge rent by rack power.

Another type of business model is GPU leasing. The ones putting up the money are cloud vendors and GPU clouds. They buy cards, buy servers, install machines into racks, and charge by card-hour or by month, earning the portion of utilization above the break-even point.

Further up is scheduling service. This layer hardly holds assets; it pools computing power scattered in several places with different specifications, splits it by task, and charges scheduling fees or earns the price difference. At the top is model invocation, where model companies charge users by token.

Among these, the integrated model of owning assets and owning workloads is the most ideal: model companies or cloud vendors build it themselves, use it themselves, and sell it themselves.

The speed of construction and delivery matters because the money for civil construction is paid out at the start of construction, while revenue only begins after racks are powered on, customers are installed, and leases start. Every extra month of delay means an extra month of interest and depreciation, with no corresponding revenue entry on the books. Equipment does not wait for the factory building either; servers and GPUs start accruing from the day they arrive, while prices decline by generation. If the data center is not built, the cards first grow old in the warehouse. The lease window is not wide either; the demand customers can clearly see only extends 18 months to 2 years. If construction is slow, by the time it is completed, this round of pricing may already be over.

Nvidia has compressed the on-site deployment of DSX POD to 14 days, managing another segment: racks, networking, power supply and distribution, and liquid cooling are pre-installed and pre-validated in the factory, then shipped to the site and connected to already-ready power, cooling, and networking. Only these two segments together make up the complete construction period; civil construction, grid connection, and approval are not included, and transformers still take four years to wait for.

The representative of the first model is the US data center operator Digital Realty, which in the second quarter of 2026 had revenue of about $1.9 billion and net profit of about $458 million. However, in the quarter there were special gains such as insurance payouts, and the group also operates a huge existing data center portfolio, so the net profit cannot all be attributed to newly built AI factories. More informative are the operating details: in the quarter, renewal rents rose 25.4% on a cash basis. The landlord did receive the money, and still has to wait for new projects to start collecting rent.

The second model, buying GPUs and then leasing them out: CoreWeave in the second quarter of 2026 had revenue of $2.575 billion and adjusted EBITDA of $1.510 billion, which looks quite substantial; but under US accounting standards it had an operating loss of $49 million, net interest expense of $640 million, and ultimately a net loss of $626 million. In between there are also depreciation, amortization, adjustments and taxes. CoreWeave has already sold a large amount of services, while still bearing the cost of expansion.

The situation is somewhat better for large integrated cloud vendors: the cloud businesses of AWS and Alibaba Cloud are themselves profitable. But these segment figures cannot prove that every newly built AI factory has recouped its investment, because they also include traditional cloud services.

In the second quarter, Google Cloud grew 82% year on year, Azure grew 43%, AWS grew 37%, Oracle OCI grew 93%, Alibaba AI cloud and computing power services grew 45%, and Tencent Cloud grew 24%.

But on the other side, more was spent. In the first half of the year, the capital expenditures of Amazon, Microsoft, Alphabet and Meta totaled $301.7 billion, about 2.1 trillion yuan; in the same period, cloud revenue was $79.8 billion for AWS, $74 billion for Microsoft Intelligent Cloud, and $44.8 billion for Google Cloud; Meta has no external cloud business. These columns are total cloud business revenue, and also include traditional cloud hosting, storage, databases and software subscriptions; AIDC revenue is not yet shown separately.

Prepayments are substantial: Microsoft $678 billion, Oracle $638 billion, Google $514 billion, CoreWeave $104.2 billion, with lease terms of fifteen to twenty years, while equipment is depreciated over three to five years. On average, Digital Realty's newly signed leases wait nine months from signing to rent commencement; Oracle's New Mexico campus is targeted to go online in 2028, and in September it also issued a force majeure notice to developers.

Cash flow has also run into problems. Meta's second-quarter free cash flow was less than $800 million, Alphabet's turned negative, and Amazon's fell 95% year on year in the first quarter. No matter how it ends, the money has really been paid out.

As for the signals of overinvestment in this round of AIDC construction, industry insiders generally believe that three indicators are more reliable than current stock prices. One is domestic chip sales; chip demand can be measured, and if the top vendors' cards cannot be sold, it means downstream can no longer absorb more computing power; obviously, supply still falls short of demand now. One is the rack-up rate, and the utilization rate calculated by racked power; a continuously declining rack-up rate means investment has run ahead of demand. And one is cloud vendors' capital expenditure, especially that of domestic cloud vendors; once mainstream vendors slow down, it means they themselves already feel the cooling; if non-market entities are still building at that point, excess supply will be very clear.

When exactly this frenzied AIDC construction wave will end depends on when large models' capabilities achieve true labor substitution in multiple industries. This node will appear first in the United States, and at least three to five more years are needed.

Jensen Huang is shortening factory delivery times. Whether this construction wave can continue depends on three conditions: whether the revenue of a single campus can, before the equipment is replaced, cover power, operations and maintenance, depreciation, interest and principal; whether utilization and rental prices can both remain stable; and whether subsequent financing can keep being renewed until the project generates cash on its own. None of these three has yet been widely verified today.

At present, AIDC is still in the stage of investing in the future, and it is not yet known how long the secondary market will keep paying for the future. In this round of construction, the certain winner may only be Nvidia, which sells equipment and factories.

Article title: An unprecedented trillion-scale investment, is Nvidia the only winner?

Has published 133 articles on Huxiu. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Huxiu APP[ ]