OPINION

In recent years, when announcing its annual resiliency research results, the Uptime Intelligence analyst team has highlighted an apparent contradiction. While the number of data center outages continues to rise, the outage rate — measured as outages per data center or per megawatt of data center capacity — has steadily declined.

The reason is simple enough: there are more data centers in operation every year. A simple analogy is aviation — there were fewer air accidents in the 1960s than in the 1980s, but many more miles were flown in the latter period.

Greater dependence on data centers can nevertheless create the perception that there are more outages. As more consumer-facing services depend on IT, the effects of downtime are felt more publicly — and sometimes more painfully. This has prompted some regulators to propose greater interventions to reduce outage risks, even though overall availability remains extremely high.

During a recent webinar on the topic, Uptime Intelligence's analysts raised an important question: will the steady improvement in resiliency continue in the years ahead (see Annual Data Center Outages Analysis 2026)? Figure 1 shows the overall outage trend reported in our global annual survey. The recorded improvements can largely be attributed to better technology, more resilient data center designs and, most importantly, effective operations, training and processes.

Figure 1 Outage frequency has shown a gradual decline

Has your organization had an impactful outage in the past three years ("Yes" respondents)?

However, there are some counter pressures building that might cause outage rates to rise from 2026 onwards — at least for a few years — as the current wave of expansion, innovation and pressure to deploy AI infrastructure poses increasing risks to availability.

These newer and growing risks — both internal and external to the data center — are discussed in the report Resiliency will be re-examined, but few will compromise published in February 2026. These include worsening climate conditions, increasing pressure on power grids, supply chain problems, staffing shortages and the introduction of complex, unproven technologies.

By maintaining and investing in high levels of resiliency and best practices, organizations can reduce the likelihood of downtime and mitigate the impact caused by these emerging factors. However, not every site will be able to do this all the time.

If outages become more frequent, their impact will register more strongly than in the past. Data centers — whether hosting cloud or AI — are much larger than they were a decade ago and host many more services, often public-facing ones. The blast zone of an outage is therefore larger, increasing the risks of consequential losses and systemic failures or "referred" outages. Larger facilities also have more stakeholders with an immediate financial interest. These include not only the operators themselves, but also traditional mission-critical tenants, organizations making substantial AI investments and newer AI-focused clients. Some of these stakeholders are highly leveraged, while others depend on business models that require very high levels of utilization.

Investors and insurers are also paying attention. At the Datacloud Global Congress in Cannes in June 2026, resiliency emerged as a central issue as participants discussed how to absorb, transfer and manage the risks associated with large data center and infrastructure investments. One advisor predicted that "service level agreements are going to harden as tenants demand more guarantees"; another suggested that insurance would increasingly be used to "manage rather than transfer the risk"; while a global law firm advised that a requirement for "five nines" (99.999% availability) is "just around the corner".

This last prediction is striking for two reasons.

First, over the past two years, many in the data center industry have been saying almost the opposite. The common refrain was that investing in full IT and facility infrastructure — including Tier III and Tier IV data centers, which would likely be required to achieve five nines availability — is simply too expensive for large AI data centers. As a result, AI training was expected, by economic necessity, to operate with lower levels of availability. It is now clear that sentiment has changed: availability has always carried a commercial premium, but that premium is scaling alongside facility size and cost.

Second, reaching these levels of availability will be challenging for many operators. AI service availability, for example, remains well below this standard. It is not only below traditional IT and cloud services, but also the availability levels achieved by the best systems and data centers of many mission-critical service operators.

As an example, most AI inference services (used by most enterprises) embed AI cloud services provided by one of the foundation model suppliers, such as OpenAI, Anthropic and Google. To date, these providers have offered levels of around 99.9% availability. Anthropic, despite being admired for its innovation, offered just 99.05% in the first half of 2026 (see As AI models improve, availability lags behind).

In 2015, Uptime introduced the term "creeping criticality" to describe how services and systems are often introduced without an initial requirement for high levels of robustness or resilient — echoing theories of disruptive product innovation from the 1990s. Over time, however, customers develop critical services on these platforms — the requirements change and the resiliency is improved.

Within the data center sector, most of those in the engineering and facility community have never relaxed their vigilance or commitment to resiliency. Now, a broader community of stakeholders — driven by hard economics — is falling in behind them. Despite that, it is probable that outages will tick up before they decline again, as the operators grapple with so many challenges at the same time.