A Day of Lateness in AI Data Centers Now Costs $5 Million, and Insurers Are Building Policies Around It

Tenants renting AI data center capacity are now demanding more than $5 million a day in penalties for late delivery of power and finished buildings, according to a report from The Information, roughly five times the $1 million a day being written into contracts as recently as this spring. The figures describe fresh negotiations, not every signed lease, but the direction is unambiguous: schedule risk is being repriced across the fastest-building infrastructure market in history.

The benchmark case came through the insurance desk. Mike Mathews, who runs the global digital infrastructure practice at broker Marsh, said his team assembled a custom delay-in-delivery policy this spring for a developer whose tenant wanted $1 million a day if the first phase of a gigawatt-scale campus slipped. Within months, tenants in new deals were asking for $5 million and up. Marsh's Nimbus facility, launched in June 2025 with up to $1 billion in construction all-risks limits and $365 million for delay in start-up, raised its ceiling to $2.7 billion in January.

The sharpest public example of tenant leverage sits in Ohio. SB Energy, the SoftBank-backed developer building a 10-gigawatt campus for OpenAI, disclosed in its IPO filing that OpenAI receives one day of free base rent for every day of delay, doubling to two days once the delay passes 90 days. A full year of slippage would return roughly 640 days of rent, nearly two years of income. The campus depends on a 9.2-gigawatt gas plant that SB Energy will neither own nor operate, and the company itself has no operating data centers, recorded $139 million of revenue in the first half, and postponed a planned $5 billion IPO in September after prospective buyers resisted a $50 billion valuation.

Tenants can write these clauses because they know the odds. A May analysis by Goldman Sachs researchers, built on facility-level tracking from Aterio, estimated that only about 60 percent of US data center capacity scheduled for the next year will arrive on time, and only about half of the two-year pipeline, against a historical rate of 72 percent. Generator and transformer lead times have stretched to 12 to 24 months, forcing developers to order equipment before projects are locked in.

The workarounds show how thin the margins have become. Oracle has started trucking compressed natural gas to sites in Utah and Texas at roughly four times the cost of piped supply to keep construction moving, and issued a force majeure notice in September on its Project Jupiter campus in New Mexico. Microsoft walked away from a deal to rent 23,000 Nvidia chips at Nscale's planned Loughton site in England after grid delays pushed power availability into the 2030s.

The financing layer is absorbing the same stress. CoreWeave, the GPU cloud carrying about $50 billion of debt, saw its bonds yield close to 15 percent earlier this month, and its interest expense hit $640 million in the June quarter, up from $267 million a year earlier. At $5 million a day, a seven-month slip of roughly 210 days generates more than $1 billion in penalties before the developer loses a single day of rent.

The deeper point is that penalty clauses do not make transformers arrive faster. They convert physical grid bottlenecks into balance-sheet risk for developers and a new product line for insurers. As gigawatt-scale financing stacks are assembled, a Marsh-style delay-in-delivery policy is becoming a standard line item, and the next place this squeeze will surface is developer credit spreads.