To state the obvious: a data center costs a lot of money.
When there’s a lot of money, there are contracts. Finish the foundation by this date, assemble the superstructure here. Get the zoning and land contracts ready. Acquire electricity, adequate data access, and other utilities needed to launch.
These all could be pretty straightforward. It might be difficult to negotiate and arrange the resources to accomplish them, but they’re one-off risks. Either the contract will be fulfilled or the penalties for non-performance will stop once the contract is completed or canceled.
Once the data center is launched, the contracts are ongoing. The sales teams guarantee a specific speed and reliability statistics for the managers to achieve. These might be evergreen contracts that are kept active while the owners offer those system advantages, If a data center consistently fails to meet those requirements, it can lead to business interruption and liability claims.
When I buy a new car, there might be a lemon law that applies. If it cannot be repaired adequately, the car is deemed a lemon and can be refunded. Would a data center that is consistently unable to meet contracted reliability levels be considered a lemon data center? Perhaps it might just be “lemon-flavored” since there’s no way to get a refund and return the massive air-conditioned box and its contents.
For most contractual obligations, a level of insurance can mitigate the risks of non-compliance. Since each stage of development has clear contractual obligations, there are insurers with an appetite for each stage of risk. Each policy, from start to finish, will have clear failure criteria and underwriters familiar with the issues can cover specific project phases. Well-placed insurers are a resource to guide the project to success and minimize their own exposure.
Many features of an active data center are also easy to assign risk for… Fire suppression systems, power failures, and cooling lapses that lead to inadequate temperature and humidity control. An insurance claim would be straightforward when something breaks like a cooling unit or power system.
Where the lemon-flavoring can come from is electronics and software configurations. Since a fault in a rack may not be identified precisely, insuring the capital behind such failures could be unattractive for underwriting. If there is a high risk of business interruption, the underwriters might change their expectations.
As a point of comparison, consider the backup company Backblaze. It collects detailed information about the reliability of their hard drives. They publish quarterly reliability data for around 350,000 drives. Their data allows outside researchers to evaluate drive model reliability. I noticed in one recent statistics chart that there can be a 10-fold range of reliability between different drive models.
One interesting message in their first quarter 2026 report was about some instances of a certain drive was having problems. Stephanie Doyle reported that the drives were having excessive failures. It was a complex situation that took an extended time to discover its solution. That reaction was not “we need to get rid of these drives”, but rather “we need to understand what was going wrong.” The problem ended up to be a complex interaction of several factors. Her conclusion was “our drive failure rates, and what we report on in general, is the function of people actively managing failure and risk.” The solution involved more than crunching numbers to make decisions. The problem was solved by careful analysis by the professionals maintaining the system.
This speaks to the problem of getting optimal results for a computing system. Building a data center is more than putting together a warehouse of GPUs as if they were a collection of high-end gaming machines. Different components in the center will have different failure profiles. Perhaps one model of memory card has more faults than demanded. A GPU might fail in a particular manner that is characteristic of a build of the computer board. Perhaps certain manufacturer facilities and date ranges were deficient, leading to problems. A 10-fold variation in reliability can create a few outliers on a critical path for the maintenance schedule. Insurance claims for normal events will be difficult to negotiate.
As Backblaze has found, a data center requires careful analysis of each feature of the hardware and the software installed on it. It’s more than just counting the frequency of unscheduled downtime. The hardware is enough of an investment that preserving the value is essential. The software configurations can lead to complicated failure patterns to decipher. Detailed diagnostic records should help identify when the problem is just a $2 connector vs. when it is a $10k GPU when both lead to the same failure patterns.
A lemon-flavored data center will have trouble meeting performance goals. It will take careful accumulation of data and human analysis to intervene. If the data is collected diligently (or compulsively?) an AI system might notice patterns in the failures. Skilled data center managers will need to develop tools to protect their investments.
With an advanced understanding of the varied configurations, some vexing failures could still have a solution. Having a robot able to quickly reboot a server will not be not enough. Diligent managers should be able to find a way to evolve their system into smooth iced tea rather than unpalatable warm lemonade that came with a pass code and a power bill.



