The bargaining power dynamic in the data center sector is shifting: previously, cloud providers called all the shots, and now players like CoreWeave are starting to be far more assertive.

The dominant position of cloud computing giants in AI data center lease negotiations is loosening. Cloud vendors like Microsoft, in their rush to bring NVIDIA servers online, have instead given operators such as CoreWeave more bargaining power — terms of the Service Level Agreement (SLA) are becoming looser, payment requirements are getting more stringent, and NVIDIA and AMD have even joined the bidding process to provide credit guarantees for their customers.

Cloud computing giants used to be the undisputed dominant party in AI data center lease negotiations. In the past, large cloud vendors including Microsoft and Google, backed by abundant capital and top-tier credit ratings, took the leading position 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 vendors like Microsoft are eager to put NVIDIA server racks into operation as soon as possible, data center operators such as CoreWeave are gaining more bargaining chips, and the balance of contract terms is starting to tilt to the supply side.

This shift is reshaping the interest pattern of the entire AI infrastructure industry chain — from the strictness of SLA terms to the pricing of power procurement, and the involvement of chip vendors, the rules are being rewritten.

01

Data center operators gain higher bargaining power, payment terms are also tightening

The Service Level Agreement (SLA) is the core battlefield of disputes.

Previously, the starting point for large cloud vendors in negotiations was to require the uptime of each server rack to be close to 100%, while setting extremely strict standards for the temperature and humidity of the computer room.

A senior executive of a data center revealed that he had seen such contract terms: As long as one rack goes down due to power outage, overheating or switch failure, the cloud vendor can waive six months of rent. If SLA breaches accumulate to a certain level, the cloud vendor can even terminate the lease directly.

This executive pointed out that negotiating SLAs is essentially a trade-off between "optimal price" and "contract durability" — the stricter the terms, the higher the price, but the greater the risk. "If you can secure an SLA with lighter penalties, it is worth it even if the price is a little lower."

At present, as operators' bargaining power rises, these extreme terms are being gradually softened.

It is not just the SLA, payment terms are also tilting in favor of operators.

According to the report citing a credit executive, he has seen such a case: a customer only rents a small part of a large data center, but the contract stipulates that once the customer fails to pay on time, it must bear the full rent of the entire facility for a period of time.

Regarding this requirement, the data center owner himself admitted that it is very tough: "He said, 'Listen, we know this is outrageous... but we can make it happen.'"

02

NVIDIA and AMD step in: chip vendors turn into credit guarantors

There is another driving force behind operators gaining more bargaining chips: the active involvement of chip vendors.

According to the report citing another data center executive, to ensure that data centers hosting their own chips can be built smoothly, NVIDIA and AMD sometimes bid for the same data center project, competing to provide credit guarantees for customers.

"NVIDIA is more aggressive because its balance sheet is stronger," the insider said.

Such arrangements are particularly beneficial for data center developers: NVIDIA and AMD are willing to provide up to 15 years of credit support for leases, while the contract term previously disclosed by NVIDIA with some emerging cloud service providers is only 6 years.

03

Electricity costs: the price difference in the same month is as high as 400%

Electricity is the single largest cost component in data center operations, sometimes accounting for more than one fifth of the total cost, and sometimes far more than that.

The aforementioned credit executive said he observed that within the same month, electricity prices for different customers fluctuated by as much as 400%.

The paradox is: every customer thinks they have got a good price, but in fact, the largest cloud vendors often pay the highest electricity prices — because they can afford it.

The pressure of power shortage is still spreading. According to previous reports from *The Information*, Elon Musk expects a severe power gap in 2027 and is planning to build his own turbine blade factory. At the same time, some data center developers are purchasing natural gas turbines from unknown "unregulated companies" at a high premium, which further pushes up insurance and financing costs.

The combination of all these factors makes cost forecasting for data centers extremely difficult.

According to reports, the cost of building a 1 GW data center has more than doubled in the past few years. If costs continue to rise, it will put pressure on investor sentiment for many large listed companies around the world.

04

Computing power fragmentation: NVIDIA's other strategic layout

Against the backdrop of tight supply of large data centers, Jensen Huang, CEO of NVIDIA, proposed another path at the Equinix customer conference.

"Computing power is fragmenting," he said at the conference. "The world of AI will be fundamentally decentralized and highly distributed."

NVIDIA, Equinix and Together AI announced a joint launch of a program at the conference: Together AI will purchase NVIDIA hardware and deploy it in Equinix's existing data centers to provide open-source AI model inference services for small and medium-sized enterprises.

Huang described the logic of this architecture: the latency-sensitive parts of AI tasks run in the Equinix data center closest to the user; while memory and inference tasks can be completed further away.

Executives from Google, Cisco and emerging cloud service provider Lambda Labs also expressed similar judgments at the conference: AI inference will increasingly move towards decentralization, and a distributed network composed of thousands of small facilities may become the mainstream form in the future.

This article is from the WeChat official account "Hard AI", author: Long Yue, editor: Hard AI, published with authorization from 36Kr.