The bargaining power in data centers is shifting: cloud vendors used to have the final say, but now companies like CoreWeave are standing up for themselves.

Cloud computing giants were once the undisputed dominant force in AI data center leasing negotiations. In the past, large cloud vendors such as Microsoft and Google, with their strong capital and top credit ratings, held the upper hand in negotiations with emerging cloud service providers and data center operators such as CoreWeave, Nebius, and Nscale.

However, according to a report by The Information on September 7, as cloud providers such as Microsoft are eager for Nvidia server racks to be operational as soon as possible, data center operators such as CoreWeave are gaining more bargaining power, and the balance of contract terms is beginning to tilt towards the supply side.

This shift is reshaping the entire AI infrastructure industry chain – from the tightness of service level agreements to the pricing of electricity procurement, and the role of chip manufacturers, the rules are being rewritten.

Data center operators are gaining more bargaining power, and payment terms are tightening.

Service Level Agreements (SLAs) are the core battleground of the dispute.

Previously, major cloud vendors started negotiations by requiring each server rack to be up to 100% capacity, while also setting extremely strict standards for data center temperature and humidity.

A data center executive revealed that he had seen contract terms such as: if a rack goes down due to a power outage, overheating, or switch failure, the cloud provider can cancel six months' rent. If SLA defaults accumulate to a certain level, the cloud provider can even terminate the lease outright.

The executive pointed out that negotiating an SLA is essentially a trade-off between "optimal price" and "contract durability"—the more stringent the terms, the higher the price, but the greater the risk. "If we can secure an SLA with lighter penalties, it's worth it even if the price is lower."

Currently, as operators gain more bargaining power, these extreme terms are being gradually softened.

Not only SLAs, but payment terms are also tilting towards operators.

According to a report citing a credit executive, he has seen cases where a client rents only a small portion of a large data center, but the contract stipulates that if the client fails to make timely payments, they must bear the full rent for the entire facility for a period of time.

The data center owner himself acknowledged the firmness of this demand: "He said, 'Look, we know this is outrageous... but we can do it.'"

Nvidia and AMD enter the market: Chip manufacturers transform into credit guarantors

Another driving force behind operators gaining more leverage is the proactive involvement of chip manufacturers.

According to another data center executive cited in the report, Nvidia and AMD sometimes compete for the same data center project, vying to provide credit guarantees to customers, in order to ensure the smooth construction of data centers that host their own chips.

"Nvidia is more aggressive because its balance sheet is stronger," the person familiar with the matter said.

Such arrangements are particularly advantageous to data center developers: Nvidia and AMD are willing to provide credit support for leases for up to 15 years, while Nvidia previously disclosed contracts with some emerging cloud service providers for only 6 years.

Electricity costs: Price differences as high as 400% within the same month

Electricity is the largest single expense in data center operating costs, sometimes exceeding one-fifth of the total cost, and sometimes much more.

The aforementioned credit executive stated that he observed electricity price fluctuations of up to 400% for different customers within the same month.

The paradox is that every customer thinks they've gotten a good price, but in reality, the largest cloud providers often pay the highest electricity prices—because they can afford it.

The pressure of power shortages continues to spread. As previously reported by The Information, Elon Musk anticipates a severe power shortfall by 2027 and is planning to build his own turbine blade factory. Meanwhile, some data center developers are purchasing natural gas turbines from dubious "fly-by-night" companies at exorbitant premiums, further driving up insurance and financing costs.

The combination of these factors makes data center cost prediction extremely difficult.

Reports indicate that the cost of building a 1-gigawatt data center has more than doubled in the past few years. If costs continue to rise, it will put pressure on investor sentiment at many of the world's largest publicly traded companies.

Fragmentation of computing power: Another move by Nvidia

Amidst a tight supply of large data centers, NVIDIA CEO Jensen Huang proposed an alternative path at the Equinix customer conference.

"Computing power is becoming fragmented," he stated at the conference. "The world of AI will fundamentally decentralize and become highly distributed."

At the conference, NVIDIA, Equinix, and Together AI announced a joint initiative: Together AI will purchase NVIDIA hardware and deploy it within Equinix's existing data centers to provide inference services for open-source AI models to small and medium-sized enterprises.

Huang described the logic behind this architecture: latency-sensitive parts of the AI task run in the Equinix data center closest to the user; while memory and inference tasks can be completed in a more distant location.

Executives from Google, Cisco, and emerging cloud service provider Lambda Labs also expressed similar views at the conference: AI inference will become increasingly decentralized, and distributed networks consisting of thousands of small facilities may become the mainstream form in the future.

Risk Warning and DisclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.