The bargaining power in data centers is shifting: previously, cloud providers had the final say, but now companies like CoreWeave are starting to assert themselves.

Author | Long Yue

    Editor | Hard AI

Cloud computing giants were once the absolute dominant parties in AI data center leasing negotiations. In the past, large cloud providers like Microsoft and Google, with their strong capital and top-tier credit ratings, held a leading position in negotiations with emerging cloud service providers and data center operators such as CoreWeave, Nebius, and Nscale.

However, according to The Information on September 7, as cloud companies like Microsoft are urgently seeking to bring Nvidia server racks online, bargaining power has been shifting toward data center operators like CoreWeave, tilting contract terms to favor the supply side.

This shift is reshaping the entire AI infrastructure value chain—from the rigidity of service-level agreements, to the pricing of power procurement, to the involvement of chipmakers in the process: all the rules are being rewritten.

Bargaining power for data center operators is rising, and payment terms are tightening

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

Previously, large cloud vendors would start negotiations by demanding near 100% uptime for each server rack, while also setting extremely strict standards for temperature and humidity in server rooms.

A data center executive revealed that he had seen contract terms like this:as long as a rack went down due to a power outage, overheating, or switch malfunction, the cloud provider could waive six months' rent. If the number of SLA violations accumulated, the cloud services provider could even directly terminate the lease.

This executive noted, negotiating SLAs is essentially a trade-off between "best price" and "contract durability"—the stricter the terms, the higher the price, but also the greater the risk. "Getting an SLA with lighter penalties is worth it even if the price is a bit lower."

Currently, as operators’ bargaining power rises, these extreme terms are being gradually softened.

Not only SLAs, but payment terms are also shifting in favor of operators.

According to a credit executive, he saw a case where a client only leased a small part of a large data center, but the contract stipulated that if the client failed to pay on time, they would be liable for the entire facility’s rent for a certain period.

Regarding this requirement, the data center owner admitted its strictness: "He said, 'Listen, we know this is outrageous... but we can do it.'"

Nvidia and AMD join: Chipmakers turn into credit guarantors

Another factor increasing operators’ leverage: the proactive entry of chipmakers.

According to another data center executive, in order to ensure data centers carrying their chips are built smoothly, Nvidia and AMD sometimes engage in bidding wars on the same data center projects, competing to provide credit guarantees to customers.

"Nvidia is more aggressive, as it has a stronger balance sheet," this insider said.

These arrangements are especially favorable for data center developers: Nvidia and AMD are willing to provide credit support for up to 15-year leases, while Nvidia’s previous contracts with emerging cloud providers disclosed only a 6-year term.

Electricity costs: Price gaps of up to 400% within a single month

Electricity is the single largest operational expense for data centers, sometimes accounting for more than a fifth of total costs, and sometimes much more.

The aforementioned credit executive noted that he observed price fluctuations as high as400% among different customers in the same month.

The paradox is: every customer thinks they got the best price, but in reality, the largest cloud providers often pay the highest rates—for the simple reason that they can afford it.

The pressure of electricity shortages is still spreading. According to a previous report from The Information, Elon Musk expects there will be a severe power gap in 2027 and is planning to build his own turbine blade factory. Meanwhile, some data center developers are buying natural gas turbines from unknown “fly-by-night companies” at a huge premium, further raising insurance and financing costs.

All these factors combined make cost forecasting for data centers extremely difficult.

According to reports, in the past few years, the cost of building a 1GW data center hasmore than doubled. If costs keep rising, it will put pressure on investor sentiment at many major global publicly traded companies.

Fragmented computing power: Nvidia’s other strategic move

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

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

Nvidia, Equinix, and Together AI announced a joint program at the conference: Together AI will purchase Nvidia hardware and deploy it in existing Equinix data centers, providing open-source AI model inference services to small and medium-sized enterprises.

Huang described the logic of this architecture: latency-sensitive AI tasks are run in Equinix data centers closest to users, while memory and inference tasks can be handled further away.

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