We Share, Inspire, and Celebrate Outrageously Successful Ethical Businesses and their Leaders

Inside Compute USA’s Approach to Long-Term Contracts: Power, Pricing and Hardware Refresh

Inside Compute USA’s Approach to Long-Term Contracts: Power, Pricing and Hardware Refresh

Companies signing long-term agreements for AI computing power have several factors to consider before committing to three or four years of service. Their equipment needs could change, newer GPUs will become available and electricity prices may rise before their contracts expire.

Getting the equipment up and running can be just as complicated. Even when a company has the money to buy the hardware, securing enough electricity to operate AI servers can take months.

Compute USA addresses these concerns by offering multi-year agreements that reserve computing capacity for individual customers at fixed rates, with options to expand or upgrade their equipment as their needs change.

Since launching in May 2026, the U.S.-based AI infrastructure company has closed more than $750 million in contracts and signed agreements for more than 100 megawatts of data center capacity across multiple partners and regions. More than 10 megawatts are currently being deployed.

With contracts lasting several years, Compute USA has to think beyond the initial equipment installation. It also needs to determine when customers can access that equipment, how much they’ll pay over the life of their agreements and how to handle upgrades as newer technology becomes available.

Guaranteed GPU Access With Room for Changing Needs

Compute USA sets aside groups of GPU servers, known as clusters, for individual customers under multi-year agreements. This dedicated capacity remains available exclusively to each customer throughout the contract.

“We start from a simple premise: the customer is buying certainty,” the company said.

The agreements include guarantees for equipment availability and performance, along with financial credits if Compute USA fails to meet certain service requirements.

Customers also have some flexibility in how much computing power they can reserve. For example, an AI company that uses roughly the same amount every day might sign a multi-year agreement for a dedicated cluster, then purchase additional capacity for shorter periods when demand increases.

By offering both options, Compute USA allows customers to cover their regular computing needs without making the occasional extra capacity part of their long-term commitment.

Customers who expect their needs to change can also negotiate the right to add more equipment at the same data center or within a connected network. Those expansion rights come with terms agreed upon in advance, allowing customers to increase capacity without going through an entirely new purchasing process.

Other changes, such as switching equipment configurations, upgrading to newer hardware or adding services, are handled through a formal change-order process that documents what both parties have agreed to.

For larger projects, the company can deliver equipment in stages, allowing customers to begin using completed clusters while the remaining equipment is being installed and tested.

“Billing and the term start when the customer accepts the cluster in writing, not when the contract is signed,” the company said.

Before asking customers to accept their equipment, Compute USA tests the GPUs, runs the equipment at full capacity for an extended period and checks the connections between servers to make sure everything performs as promised.

Compute USA tracks whether clusters pass their initial tests and whether equipment is delivered by the agreed dates. After the equipment is running, customers receive monthly reports covering system and network availability, technical problems and any financial credits they’re owed.

Those credits are applied automatically. The company also tracks how quickly problems are addressed, how long repairs take and whether the same issues keep occurring.

Customers have dedicated support channels available around the clock, with direct escalation to the engineers who designed their clusters. Compute USA also prioritizes data center partners with SOC 2 Type II or comparable certifications, which many enterprise customers require.

Why Electricity Comes Before Equipment at Compute USA

Having enough GPU servers is only part of meeting a customer’s needs. Compute USA also has to find a data center with enough electricity to run the servers and enough cooling capacity to prevent them from overheating.

A single NVIDIA GB300 NVL72 rack uses well over 100 kilowatts of power. A facility housing several of these racks needs considerably more electricity, along with cooling systems capable of handling the heat they generate.

“Power is the constraint,” the company noted.

Compute USA considers electricity availability the biggest obstacle to getting AI infrastructure up and running in the United States.

When considering a data center, Compute USA looks at how much electricity the facility can actually provide, rather than relying on promises of additional capacity in the future. It also checks how reliable the power supply is, how long it would take to get new electrical connections and what redundancy is in place.

A facility might have enough room for the equipment but still be 18 months away from receiving the electricity needed to operate it. Even with servers ready for installation, that location wouldn’t work for a customer expecting its computing capacity to become available within the year.

Working with multiple data centers gives Compute USA more options when deciding where to install a customer’s equipment. It also reduces the risk of a delay at one facility holding up other projects.

The company considers each customer’s delivery timeline, the amount of equipment the facility can accommodate and the cooling requirements of the selected GPUs when choosing a location.

Several members of Compute USA’s team previously worked in Bitcoin mining and other large infrastructure projects, where finding reliable electricity and operating power-intensive equipment were regular parts of the job. They’ve brought that experience to the company’s AI data center operations, which face many of the same challenges.

How Compute USA Manages the GPU Refresh Cycle

Even after a customer’s equipment is installed and running, Compute USA has to consider what happens when newer, more powerful GPUs become available before the contract ends.

NVIDIA now releases a new GPU architecture roughly every year, while many computing agreements last three or four years. That means customers could see several generations of hardware released before their original agreement ends.

“The newest systems tend to go to frontier training, where every gain in performance matters,” the company explained.

Not every customer needs the latest equipment, though. Compute USA considers what customers plan to do with their servers when helping them choose hardware. Newer GPUs are particularly useful for training advanced AI models, while older generations can still handle tasks such as running existing models, fine-tuning them and conducting research.

Compute USA’s GPU Cloud offers NVIDIA B300 and GB300 systems to accommodate different computing needs.

Customers who need newer equipment can arrange to add capacity or move to another GPU generation as their requirements change. Compute USA also considers future upgrades when choosing data centers, checking whether facilities can accommodate higher rack densities and the electricity and liquid cooling requirements of future GPU systems.

The company keeps spare components available, replaces faulty parts and performs maintenance to help existing clusters continue meeting their agreed performance requirements.

Keeping Prices Fixed as Costs Change

Compute USA locks in pricing for the full term of its long-term agreements, including electricity, cooling, connectivity and support covered by the contract. Customers aren’t charged extra when utility rates change or new tariffs take effect.

A customer signing a four-year agreement knows the contracted rate they’ll pay in both month one and month 48. Compute USA takes responsibility for changes in its own costs by securing power and hardware on terms that match its customer commitments.

The company also finances deployments against long-term contracted demand and manages projects separately to prevent one customer’s deployment from putting another at risk.

Shorter-term and on-demand computing capacity is priced more closely to market conditions, giving customers the option to purchase additional resources without another multi-year commitment.

Compute USA also considers more than the hourly GPU rate when evaluating costs. Equipment use, network performance, reliability and engineering time spent resolving problems all affect what customers ultimately spend to complete their work.

“Pricing should be honest,” the company explained. “We’d rather quote a number we can stand behind for years than win a deal with a price we’d need to claw back later.”

Looking Ahead: Demand and Repeat Business

Compute USA expects inference to become the largest source of computing demand over the next five years as more businesses begin using AI in their daily operations.

The company is already seeing demand from AI platforms and businesses in financial services, healthcare, industrials and the public sector, some of which require greater control over their data and U.S.-based infrastructure. AI labs and model developers are also seeking newer, high-performance equipment for training and fine-tuning increasingly demanding models.

As demand grows, Compute USA plans to expand its data center capacity while remaining careful about the commitments it makes to customers.

Securing additional contracts is important, but getting customers to return for another deployment is the stronger sign that the business is growing the right way.