Those who advocate hitting the brakes on AI have driven up electricity costs across the United States.

Anyone who follows AI can clearly feel that the release of GPT6 has been an absolute smash hit. Whether it is official statements or real tests by various content creators, people feel that the AGI moment is truly just around the corner. Full-process automatic editing, pure voice interaction without the need for a keyboard and mouse — it seems that GPT6 can learn almost any software that humans can use, and the only problem left seems to be the high cost. As a Pro user, I am both happy and worried: the available AI capabilities have been upgraded again, but my monthly quota gap has grown larger.

Just a few days later, a group of top players in the global AI industry began to collectively call for a slowdown in AI development. First, Dario Amodei, CEO of Anthropic, publicly proposed to slow down the development of cutting-edge AI. Then Sam Altman of OpenAI, Elon Musk who led the development of Grok, and other figures successively expressed their support for this initiative.

Are these industry leaders deliberately going against profit? Do their investors agree? Do they still want to go public via IPO? Gao Zhikai, an international relations expert who is also a popular Bilibili content creator, said: In about ten months, the US AI bubble is very likely to burst.

Is this for real?

As a naturally skeptical person, I have long known that these AI industry leaders can make up any story for financing, and I highly doubt how much credibility their words have this time. I simply followed the sources of the news to dig around online to see what is really going on. Is the "voluntary slowdown in AI development" their usual deceptive trick, or have some people really started to hit the brakes?

The brakes have been pressed, but not fully engaged

OpenAI once suspended the reinforcement learning training of its latest model intended for deployment for about two weeks. Back in August, it "temporarily slowed down the pace of expansion" to meet new safety requirements.

On August 28, after the new safety requirements were implemented, the previously suspended large-scale cutting-edge reinforcement learning training resumed. However, as of September 1, some smaller experimental training tasks remained suspended. OpenAI also confirmed that in the past few weeks, part of the development and release of its then cutting-edge model Astra was postponed to strengthen and test protection mechanisms.

During the restricted period, the GPU allocation for Astra-class models dropped by 59.2%, while the GPU allocation for other model categories increased by 17.2%. Among the analyzed reinforcement learning tasks, other models absorbed about 85% of this reduced allocation, and the overall total GPU allocation remained roughly unchanged.

AI-related investment is still charging ahead at full speed

Anthropic is still preparing for its 2026 IPO, with the maximum fundraising amount discussed in the market reaching about 1000 billion US dollars, and the potential maximum valuation reaching about 2 trillion US dollars. Relevant computing power contracts have been signed for as far as ten years later. In the AWS cooperation announced in April, Anthropic promised to invest more than 1000 billion US dollars in the next ten years, to obtain up to about 5GW of computing power for training and running Claude.

OpenAI announced another listing arrangement at the same time. Sam Altman said that OpenAI will not launch an IPO in 2026, and the AI safety work that needs to be completed at present is one of the reasons.

The investment situations of several other large companies are also different. Alphabet has spent about 80.6 billion US dollars in capital expenditure in the first half of this year, compared with about 39.6 billion US dollars in the same period last year. When announcing its second-quarter results in July, the company raised its full-year plan to 195-205 billion US dollars to continue building AI infrastructure and data centers.

Amazon's full-year cash capital expenditure plan is about 220 billion US dollars, mainly invested in AI, as well as businesses such as robotics and satellites. Andy Jassy said at the time that even with this amount of spending, this year's capacity is still not enough to meet all demand.

As of the end of June, Meta's unleased data center, computer room hosting and some network infrastructure leases totaled about 279 billion US dollars, with some leases lasting up to 30 years.

Most of these capital plans and long-term contracts were formed before the round of slowdown discussions in September. Whether the next round of investment plans will be adjusted awaits new official disclosures.

Nvidia, the industry's workaholic, plays multiple roles

The large companies mentioned above can invest hundreds of billions of dollars, but some AI cloud companies and model companies do not have the capital, credit and infrastructure access capabilities of Microsoft and Google. After purchasing GPUs, they also need to secure long-term land, power and computer room contracts to raise funds for construction. Nvidia disclosed in its regulatory filings that it has helped some customers secure these resources and provided guarantees. By observing how these customers build their computing power, you can see what other transactions Nvidia is involved in besides selling chips.

CoreWeave is a company that builds AI clouds and provides computing power services to customers. CoreWeave is a major customer of Nvidia, which buys Nvidia's GPUs to build AI clouds, while Nvidia is also a shareholder and computing power customer of CoreWeave. On September 9, 2025, the two parties signed an order with an initial amount of about 6.3 billion US dollars: after conditions such as delivery and service availability are met, if the computing power covered by the contract is not sold to other customers, Nvidia is obliged to purchase it. This arrangement will last until April 13, 2032.

Nvidia has similar arrangements with some of its AI cloud partners: cloud companies buy Nvidia's data center products, and Nvidia promises to purchase part of their cloud computing power; after third-party customers buy the relevant capacity, Nvidia's own procurement obligation will be reduced accordingly. As of July 26, 2026, the commitments corresponding to such agreements are about 36 billion US dollars, with a usual term of six years.

Nvidia made it very clear in its regulatory filings: Customers may postpone the procurement of new-generation hardware due to insufficient funds, and shortages of land, power and computer rooms will also delay deployment. According to Nvidia's own judgment, some AI cloud and model companies have training and inference needs, but it is difficult for them to obtain long-term infrastructure contracts, and lack the ability to finance under investment-grade credit conditions. Companies with weaker capital strength are particularly vulnerable to such restrictions.

Nvidia has helped some customers lock in land, power, computer rooms and data center capacity, and provided guarantees. For large cloud vendors and enterprises with investment-grade credit, it is expected that they will still secure these resources on their own.

The PORTS campus in Ohio is developed, owned and operated by SB Energy, with OpenAI as the long-term user of the campus, and the relevant lease term is about 20 years. Nvidia provides AI infrastructure, and also provides credit support under agreed conditions. The first batch of credit support from Nvidia corresponds to an IT load of about 4.25GW, with a cumulative upper limit of payment obligations reaching 1050 billion US dollars.

This liability covers the land, power and computer rooms within the agreed scope. The 1050 billion US dollars is a conditional cumulative upper limit of payment, not cash already paid, a loan to OpenAI, or the total cost of the entire project. As one of the relevant conditions, the campus will in principle only deploy Nvidia AI infrastructure, with limited exceptions reserved. The project is expected to be put into use in phases starting from 2028.

These GPU procurements, long-term leases and credit supports all correspond to data centers that are to be built and put into operation. Data centers need land, access to the power grid and guaranteed water supply. At the construction site, apart from capital and computing power, people also need to figure out: whether the land is sufficient, whether the power grid can be connected, where the water comes from, and who will build and pay for the new supporting facilities.

Texas (not Shandong Province in China) is very concerned about electricity and water bills

By August 2026, the total capacity of large new power access applications tracked by ERCOT, the grid operator of Texas, has exceeded 474GW, about 90% of which comes from data centers. For reference, the actual peak record announced by ERCOT at that time was about 91.1GW, and the applied capacity was more than five times that figure. However, 474GW is only the total of applications, which cannot be simply counted as all projects that will be completed and actually consume electricity.

The Abilene campus under construction is one of the first important projects of Stargate, with a planned power capacity of about 1.2GW, and 8 buildings adding up to about 4 million square feet, equivalent to about 370,000 square meters. Google also announced that it will add about 40 billion US dollars of investment in Texas by 2027, including building new data center campuses in Armstrong and Haskell counties.

It is still impossible to estimate how many projects in the 474GW applications will finally be implemented based on the application volume. But even if only part of them are completed, the power grid needs to evaluate the power generation, transmission and grid connection capabilities in advance, and figure out which supporting facilities need to be expanded. Who will build these facilities and who will bear the costs must also be arranged before the projects are connected to the grid. The state government's requirement for data centers to bear the cost of new power infrastructure brought by themselves is to deal with how this cost is shared.

In August, the state government required relevant data centers to undergo additional audits. Developers need to report the estimated power consumption, peak load, and how much power they plan to generate on their own; where the water comes from, how much water they will use, and what technology the cooling system adopts.

Putting all the above facts together, several things are happening at the same time. OpenAI did suspend specific training, and delayed development and release, and resumed these operations after meeting the new safety requirements, with part of the computing power allocated to other models. The slowdown actions disclosed so far have specific targets and resumption conditions.

Anthropic is still advancing its IPO and long-term computing power arrangements. Most of the huge investment plans of Alphabet, Amazon and Meta were made before this round of discussions, and there is no new official disclosure yet to confirm whether the next round of plans will be adjusted. Among GPU suppliers, AI cloud companies and model companies, the identities of shareholders, suppliers, computing power customers and credit support providers have begun to overlap. The implementation of data centers also requires securing land, power grids, water sources and the cost of new facilities, and Texas has started to ask these questions one by one.

We can continue to watch the actual actions in the next stage: whether the cutting-edge training plans will be adjusted, how Anthropic's IPO will be priced, whether the next round of capital expenditure of large tech companies will change, and whether the announced data center projects will be delayed, scaled down, or continue to expand.

This article is from "Feng Dabai", authorized for release by 36Kr.