AI Data Center Bargaining Power Shifts Toward Supply Side as Firmus Signs Deal with OpenAI
nashnova research
Surging AI compute demand is shifting data-center lease negotiations in favor of operators. Nvidia-backed Australian firm Firmus has signed a multi-year deal with OpenAI, pushing its total contracted capacity past 900 MW — the supply side's pricing reset is underway.
Why can operators suddenly dictate terms?
The core driver: Microsoft and other hyperscalers need Nvidia server racks online as fast as possible. Demand far outstrips available facility supply.
This means → whoever holds ready-built floor space and power connections sets the terms. Operators like CoreWeave are in that seat now.
In plain terms = landlords used to chase tenants; now tenants chase buildings. The negotiating leverage has flipped.
How exactly have contract terms changed?
Previously, hyperscalers demanded near-100% rack uptime. If a single rack went down — power cut, overheating, switch failure — the cloud tenant could cancel six months of rent. Accumulated SLA breaches could trigger outright lease termination.
Those extreme clauses are now being softened. Payment terms are tilting too — one tenant leased only a small portion of a large facility, yet the contract stated: miss a payment, and you owe the full rent for the entire building for a period.
One data-center owner put it bluntly: "Look, we know it's outrageous… but we can do it." This reflects a supply side confident enough to write lopsided terms into contracts.
Why are Nvidia and AMD stepping in directly?
To ensure data centers carrying their chips get built, Nvidia and AMD sometimes bid against each other on the same project, competing to offer clients credit guarantees of up to 15 years.
This means → chip makers are no longer just selling silicon — they are using their own balance sheets to backstop operators' financing risk. Nvidia is more aggressive because its balance sheet is stronger.
Firmus's OpenAI deal was completed with Nvidia's backing. Firmus will also deploy Nvidia's next-generation Vera Rubin processors across the Asia-Pacific region.
How large is the hidden power-cost variable?
Electricity is the single largest line item in data-center operating costs. Power prices for different tenants in the same month varied by as much as 400%, and the biggest hyperscalers typically pay the highest rates — because they can afford to.
Building a single 1 GW data center now costs more than double what it did a few years ago.
In plain terms = power is both operators' biggest expense and their lever for screening tenants — the large clients who can pay top rates end up with the least room to negotiate on price.
What does "decentralized compute" actually mean?
Nvidia CEO Jensen Huang told an Equinix customer conference: "Compute is fragmenting. The AI world will be fundamentally decentralized, highly distributed."
Nvidia, Equinix, and Together AI launched a joint initiative: Together AI buys Nvidia hardware, deploys it inside Equinix's existing facilities, and offers open-source AI inference services to small and mid-sized businesses. Executives from Google, Cisco, and Lambda Labs voiced similar views.
This means → not all compute needs to crowd into hyperscale campuses. Spreading workloads across smaller, existing facilities is another path to relieving the supply bottleneck.
What does the Firmus deal tell us?
Firmus signed a multi-year agreement with OpenAI. Two data centers in Malaysia will supply the compute; OpenAI becomes Firmus's anchor tenant, lifting total contracted capacity past 900 MW.
Malaysia is now Southeast Asia's fastest-growing data-center market, but rapid expansion has raised concerns about power and water consumption.
Firmus is reportedly preparing an IPO; its latest funding round valued the company above $10.5 billion. This reflects the market pricing in a supply-side bargaining-power reset — whether that valuation holds depends on operators sustaining this negotiating advantage.
市场有风险，内容仅供研究参考，不构成投资建议。