Applied Digital’s $10M Renewal Gap

Applied Digital’s quarterly revenue rose 22% to $53 million—and still missed estimates by $10 million after customers delayed lease renewals. At the same time, Microsoft, Alphabet, Amazon, and Meta kept pouring capital into the same AI data-center boom. The numbers came from one buildout but two financing systems.

The buildout split before the headline did

At first, investors treated every new campus, power agreement, and GPU order as another expression of the same demand. Scarce compute made the shortcut useful: if the largest buyers wanted more capacity than suppliers could deliver, announced megawatts looked close enough to occupied megawatts, and occupied megawatts close enough to durable revenue.

The spending continued even as operator financing tightened. Microsoft, Alphabet, Amazon, and Meta reported combined 2024 capital expenditures of $246 billion, up 63% from $151 billion, and indicated that 2025 spending could exceed $320 billion. The four later forecast roughly $650 billion of combined 2026 capital expenditures, driven by data-center construction.

Applied Digital’s shares fell more than 13% after delayed renewals pushed quarterly revenue below expectations.

Together, the figures separate broad demand from operator financeability. Hundreds of billions of dollars in platform investment confirmed demand for infrastructure; Applied Digital’s miss exposed the risk of carrying capacity on uncertain commitments. Capacity demand remained strong, but balance-sheet strength and contract quality determined who could carry it.

Balance sheets buy time; leases must buy certainty

Hyperscalers can absorb mismatches between construction schedules, customer adoption, and utilization because diversified businesses and multiple capital sources fund their capacity. Microsoft added more than 500 megawatts of data-center capacity after July 2023 and surpassed five gigawatts of total capacity in the first half of its latest reported fiscal year. It did not need every incremental megawatt to arrive with an external tenant attached.

Hyperscalers also use leverage. Microsoft’s data-center finance leases rose by nearly $100 billion in two years to $108.4 billion, largely for leases beginning between fiscal 2025 and 2030. The company can combine diversified cash flow with leases and other capital; an operator dependent on customer contracts has fewer ways to fund the interval before a tenant pays.

A large platform can own a facility, sign a finance lease, rent third-party capacity, or shift workloads among regions while other cash flows carry the transition. A contract-dependent operator must persuade lenders and investors that a specific customer will pay for a specific block of capacity long enough to cover construction, equipment, and financing costs. The operator’s financing rests on a commitment stack: customer credit, guarantees, term, termination rights, delivery milestones, and the cost of funding the interval before revenue begins.

The hyperscaler buys time and decides later where to place the workload. The operator sells time in advance so the facility can be built. A slipped renewal can therefore become the operator’s central business risk.

Investor capital cannot renew a customer lease

Applied Digital made the reversal clear within seven months. In September 2024, the company raised $160 million from Nvidia and other investors to expand its data-center and AI cloud-computing business. In April 2025, delayed lease renewals helped produce the $53 million quarterly revenue result.

Investors initially treated the funding round and revenue growth as consecutive proof of demand. Capital raised appeared to validate demand because investors were willing to fund supply. Revenue growth appeared to validate the capital because customers were using that supply. Delayed renewals broke the chain where intent had to become another signed term.

The $160 million round could fund expansion, but the next customer signature remained outside Applied Digital’s control. The company still posted year-over-year growth while delayed renewals weakened its claim on future cash flow.

A bankable megawatt needs power and a customer

The gigawatt became the preferred unit of the buildout because it compressed an industrial system into one legible number. Each data center, however, must combine power delivery, grid access, cooling, network capacity, and utilization. Facility construction costs are substantially tied to the maximum power draw the site must support, even when customers are not yet paying to use that maximum.

By 2025, US AI data-center capacity classified as built, underway, planned, or stalled had topped 80 gigawatts. That total put a powered hall and a stalled project in the same pipeline even though one could host workloads and the other could not. It measured ambition across several stages rather than a uniform stock of revenue-producing infrastructure.

Overloaded grids pushed operators toward on-site power plants, where permitting and supply-chain limits introduced new delays. A site could have land without grid service, a power agreement without completed generation, or completed capacity without enough customer utilization. Capacity headlines conflated each condition because the unit had no field for execution risk.

Dispatched electricity sold under an offtake agreement services a power plant’s debt. Data centers have reached the same structural threshold. A powered megawatt is more real than a planned one, yet it remains an option on demand until a customer consumes it or signs an enforceable commitment to pay for it.

Lenders should divide contracted revenue, adjusted for renewal risk, by powered capacity, excluding secured land, campus master plans, and maximum eventual buildout. Contracted revenue per powered megawatt keeps execution risk visible and helps an underwriter price the gap between construction spending and lease payments.

Lease terms divide capacity into financial classes

Long duration can turn a specialized facility into something lenders can underwrite. Anthropic’s 20-year, roughly $19 billion lease for a TeraWulf facility covers a site expected to have about 400 megawatts of capacity. If the reported lease value is spread evenly over the full term and capacity, it implies roughly $2.4 million per megawatt-year before adjustments.

Underwriters can use the ratio only with its limitations in view. The facility is not expected to begin delivering power until the second half of 2027, so the operator still carries construction and delivery risk. The lease also concentrates substantial revenue in one counterparty. Underwriters must examine guarantees, milestones, termination rights, and the customer’s capacity to pay through several technology cycles.

Even the meaning of “term” can become contested. Reported language around another Anthropic arrangement ranged from a 180-day lease with 90 days’ notice to monthly fees extending through May 2029. That difference determines how much debt the revenue can support and how quickly the operator can be left carrying capacity alone.

The market has already distinguished leased infrastructure from proposed infrastructure. Digital Realty agreed to pay $7.8 billion for a majority stake in three fully leased Northern Virginia data centers. Their fully leased status put those assets in a different financial class from a planned campus still seeking power and tenants.

Gigawatts worked as a proxy while scarcity made every credible block of capacity look spoken for. Once the proxy became a target, operators announced and investors rewarded planned capacity, allowing gigawatt headlines to display demand while concealing the distance to cash.

Applied Digital’s $53 million quarter showed that revenue can rise even as the next lease becomes less certain. Even when a campus plan says 400 megawatts, the bankable asset begins where the power meter and the renewal clause share the same address.