The race to bring AI data centres online is fundamentally changing who holds the advantage in contract negotiations. Projects like the Stargate AI data centre in Abilene, Texas, illustrate a landscape where deals are rarely identical. The entire process is a complex, multi-party negotiation involving land owners, power suppliers, data centre builders and operators, lenders, and major clients such as Microsoft and Google.
Historically, hyperscale cloud giants held the upper hand when leasing computing power from emerging providers like CoreWeave, Nebius, and Nscale, leveraging their vast capital and superior credit ratings. However, that dynamic is shifting. According to multiple data centre and credit executives, operators are now gaining leverage because cloud giants are so desperate to get their Nvidia server racks operational. This has allowed operators to push back on the risks they are expected to shoulder.
Previously, large cloud providers used their strong position to impose stringent service level agreements on smaller rivals. These core contract clauses dictate the performance standards AI servers must meet, with breaches resulting in financial penalties. Initial demands often required nearly 100% uptime for every rack. One data centre executive noted that developers were also forced to meet exceptionally tight temperature and humidity specifications. Some draft contracts even allowed the cloud giant to withhold six months of rent for a single rack experiencing downtime, regardless of the cause. Repeated SLA breaches could lead to lease termination. While insiders say SLA breaches are common, clients typically do not enforce the penalties.
Data centre leaders say SLA negotiations require a delicate balance: accepting tough terms for lower prices while also seeking "room for error." Even with lower rental quotes, the goal is to secure clauses with less severe penalties. The rise in operator leverage is also being fueled by chipmakers offering more innovative financing and deal structures. Executives report that Nvidia and AMD sometimes compete to provide credit guarantees for clients of data centre projects to ensure their capital-intensive hardware gets deployed. Nvidia, with its strong balance sheet, is reportedly the more aggressive of the two. This benefits developers, with Nvidia and AMD offering credit support on leases for up to 15 years, compared to the six-year terms Nvidia has previously disclosed for some emerging cloud providers.
Data centre operators are also feeling more confident in demanding stricter payment terms. In one case, a credit executive observed a situation where a client leasing a small portion of a large facility would be obligated to pay the rent for the entire building for a period if they were late on their payments. The data centre owner acknowledged this was a tough stance, saying, "We know it's an outrageous request, but we can ask for it."
This shifting power dynamic is also impacting utility companies. Power costs can represent over a fifth of a data centre's operating expenses, and sometimes much more. Surprisingly, larger and wealthier clients are often less sensitive to electricity prices. One credit executive noted that electricity price fluctuations of up to 400% could occur for different clients within the same month. Despite each party believing they are getting a good deal, it is often the largest cloud providers who end up paying the highest rates simply because they can afford to. This urgency is also leading some developers to buy gas turbines from what the industry calls "fly-by-night startups," which can drive up insurance and debt financing costs due to the difficulty in insuring equipment from unknown sources.
The true cost of a data centre is becoming incredibly difficult to predict. The price tag for building a one-gigawatt facility has more than doubled in the past few years. If costs continue to escalate, it could dampen market sentiment for major listed companies globally. I will be closely following how these contracts, terms, and negotiation strategies evolve. This dynamic was a key topic at Equinix's recent customer conference. An Nvidia executive highlighted that power access is now arguably the most challenging part of the supply chain.
A partial solution being explored is the construction of smaller, distributed data centres for inference workloads. These smaller facilities could alleviate the power gap while chipmakers and AI labs wait for larger computing hubs to come online. Nvidia, Equinix, and Together AI have announced a partnership where Together AI would deploy Nvidia hardware in Equinix's existing facilities to optimise and serve open-source AI models to clients other than the hyperscalers. While the executives did not disclose the investment or chip numbers, Nvidia CEO Jensen Huang views such collaborations as a way to balance the influence of major labs and cloud providers while also diversifying his own customer base. "Computing is going to be fragmented. The AI world is fundamentally going to be distributed, not hyperscale," he said, explaining that latency-sensitive tasks could run in the nearest Equinix facility while storage and some inference happen remotely. Executives from Google, Cisco, and emerging cloud provider Lambda Labs echoed this sentiment, suggesting AI's future is decentralised. While giant gigawatt facilities will remain necessary for training large models, a distributed architecture of thousands of smaller sites could become the standard for AI inference.
Comments