Signing a multi-year AI infrastructure contract usually draws public attention. At Compute USA, the more telling moment comes later, inside a data center, when a customer’s team puts a newly built GPU cluster through its paces and decides whether to accept it in writing. Until that sign-off happens, the customer isn’t billed, and the contract term doesn’t start.
That order of events says a lot about how the U.S.-based neocloud judges its own performance. Compute USA offers GPU cloud, hyperscale cloud, enterprise hardware sales and long-term capacity agreements, all built and operated in the United States, with clusters running NVIDIA platforms such as the GB300 NVL72. Its customers range from AI-native platforms and model developers to enterprises moving AI out of pilot programs and into production. Many of them arrive with multi-year roadmaps that only work if capacity shows up when it was promised.
Founder and CEO Mason Jappa said the company starts every agreement from a plain reading of what the customer is paying for.
“The customer is buying certainty,” Jappa explained. “When a team commits to a multi-year training or inference roadmap, the worst outcome is capacity that slips or disappears. So the core of every agreement is dedicated capacity. The cluster is reserved exclusively for that customer for the full term; it isn’t shared or reallocated, and it’s backed by firm service levels with real credits behind them.”
The company’s contract figures have climbed quickly. Inside Compute USA, however, the numbers that draw the closest scrutiny come after the deal closes, and they center on whether the company kept the commitments that won the business in the first place.
“Bookings tell us the market believes in us,” Jappa shared. “Whether we’re succeeding for customers shows up in different numbers, and those are the ones we watch most closely.”
Five Operating Metrics Tell Compute USA Whether a Customer Would Sign With the Company All Over Again
The first standard Compute USA tracks is on-time delivery of committed capacity. Jappa called it the first promise the company makes, and the one customers remember longest. Timing carries extra weight in this market because many customers come to Compute USA after being told to wait months for large GPU allocations elsewhere.
The second standard is clean first-pass acceptance testing, and the company’s bar for a passing grade is demanding.
“Every cluster goes through rigorous acceptance testing, including full GPU diagnostics, a sustained full-load burn-in and fabric performance tests,” Jappa said. “Clean first-pass acceptance means we built it right.”
Third on the list is uptime and network availability measured against what the contract guarantees. Customers get a full accounting every month.
“We report these to customers every month, along with every incident, its root cause and any credits due,” Jappa shared. “Credits are applied automatically, so our incentive is simply to avoid them.”
The fourth metric covers response and resolution times by severity level, along with how often an incident recurs. A repeat problem tells the company a fix didn’t hold. Customers have dedicated support channels around the clock, with direct escalation to the engineers who designed their cluster, which keeps accountability close to the people doing the work.
The fifth metric is the one leadership treats as the final word.
“Expansion is the clearest signal of all,” Jappa said. “When a customer adds capacity or brings us their next deployment, it means the first one worked.”
The company boils all five down to a single internal test: whether a given customer would choose Compute USA again tomorrow, knowing everything they know now.
“If the answer is yes, the bookings take care of themselves,” Jappa said.
Moving From One-Off Hardware Sales to Dedicated Multi-Year Clusters Rewrote How the Company Operates
Compute USA’s metrics grew out of a shift in what its customers wanted. Much of the company’s early work involved helping customers source and deploy GPU servers, backed by OEM and reseller relationships. As AI workloads grew larger and more permanent, the requests coming in started to look different.
“As demand grew, customers increasingly wanted something different: a single partner accountable for the whole thing, from power and facility to hardware, networking, operations and uptime, under a multi-year commitment,” Jappa said.
A server sale ends at delivery. A dedicated cluster under a three- or four-year agreement keeps the company accountable every month. Compute USA now structures those agreements with phased delivery, so customers don’t pay for capacity before they can use it, along with pre-agreed expansion rights and a formal change order process for upgrades and configuration changes.
Jappa described the move as the company’s biggest inflexion point and said it changed almost everything about how the business runs.
The shift also reframed the conversation about cost. Jappa believes buyers who fixate on the hourly GPU rate often miss where their money actually goes.
“What actually determines the cost of a workload is utilization, network performance, reliability and how much engineering time a team loses to infrastructure problems,” he said. “A cheaper cluster that underperforms or sits idle while engineers debug it is the most expensive thing you can buy.”
Compute USA Built Enterprise Service Levels, Acceptance Testing and Deal-Level Financing Before It Needed Them
Rather than wait until customers demanded formal guarantees, the company put the structure in place ahead of time.
“We put enterprise-grade service levels, acceptance testing and monthly reporting in place so customers could hold us to a standard, not a promise,” Jappa said.
That structure has proven especially relevant with enterprises in financial services, healthcare, industrials and the public sector. Many of those buyers carry strict requirements around data control, security and U.S.-based infrastructure before they move a production workload anywhere.
Financing follows the same logic. Each deployment is set up to stand independently.
“Dedicated project structures and financing matched to contracted demand let us scale without putting one customer’s deployment at risk because of another,” Jappa explained.
Pricing on committed capacity is locked for the full term. It covers power, cooling, connectivity and in-scope support, with no pass-through charges for utility rate changes or new tariffs. Compute USA absorbs that risk itself, a choice Jappa defends in blunt terms: “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,” he said.
The company also leans on data center operators, OEMs and network partners to meet its commitments, and Jappa was clear about where responsibility lands when something goes wrong.
“We select partners who can meet the commitments we make to customers, and we stay accountable for the outcome either way,” he said.
More Than $750 Million in Closed Contracts Now Puts That Operating Discipline to the Test at Real Scale
Compute USA’s operating model now supports more than $750 million in closed contracts, with roughly $1 billion in pipeline the company expects to close in the fourth quarter. The company has also signed more than 100 megawatts of data center capacity across multiple partners and regions, with more than 10 megawatts currently being deployed.
Growth at that pace makes the five metrics harder to hit, not easier. Power remains the binding constraint across the industry, so the company evaluates every site power-first and spreads its portfolio across locations to shield customers from any single site’s delays. It also means turning down work that doesn’t fit.
“In a fast-moving market, it’s tempting to say yes to everything,” Jappa said. “We’ve learned that the deals worth winning are the ones we can deliver exactly as committed.”
Jappa expects the industry to sort itself out quickly as AI compute becomes core national infrastructure, as essential to the economy as power grids and telecom networks.
“Over the next few years, the gap will widen between companies with energized power, real deployment experience and the balance sheet to build, and companies with announcements,” he said. “Execution becomes the moat.”
Compute USA intends to be one of the independent, U.S.-based operators at the center of that market, trusted for dedicated capacity because it has consistently delivered it. The company plans to keep expanding its power portfolio and deployments while holding to the same rule that shaped its first contracts: promise only what it can build, then build it.
“The lesson for us has been that scale isn’t just more racks,” Jappa said. “It’s building a company that customers can trust with infrastructure their business depends on.”
