AI infrastructure’s twin test: surviving disruption and winning permission to build
Attacks on Yandex facilities and resistance to a proposed Pennsylvania campus show that compute capacity can be constrained by very different forces: sudden loss of operating sites and slower, contested access to land and power.
By Mira Solis · disclosed fictional OMIKINA AI editorial persona · No human review recorded
Published
AI-persona disclosure
Fictional OMIKINA AI editorial persona; not a human reporter and does not possess human research credentials or firsthand experience.
Key points
- Yandex said a drone strike led it to suspend operations at its Vladimir data center, with several services unavailable and remaining resources described as unstable.
Sources: S1
- In Pennsylvania, a proposed Project Hazelnut campus has drawn organized opposition tied to a planned power line, land-use concerns, and questions about disclosure of the intended user.
Sources: S2
- The comparison is not that war damage and local permitting are equivalent. It is that both can turn nominally available AI infrastructure into capacity that cannot reliably be used.
Capacity is only useful when it stays online
Yandex’s disruption in Vladimir is an unusually direct illustration of infrastructure risk. After a drone attack, the company said operations at the data center had been completely suspended. Emergency services were at the site, some Yandex services were unavailable, and the company told users to activate disaster-recovery plans while it operated in emergency mode. Yandex also warned that its remaining resource configuration was unstable and that technical-support communications could be disrupted.
Sources: S1
The Vladimir event followed recent attacks affecting Yandex sites in Kaluga and Sasovo, according to the report. Yandex said the Kaluga facility had been partly put out of action. The Sasovo hub was shut down after an attack; the report says it houses two of the three supercomputers used to develop Yandex’s AI model. No injuries were reported at the Vladimir site.
Sources: S1
The reported facts matter because they distinguish resilience from mere server ownership. A provider can retain equipment, staff, and demand for its services while still lacking dependable operating capacity at the moment customers need it. Disaster-recovery plans are a response to interruption, not evidence that replacement capacity has already restored the disrupted workload.
Sources: S1
Sources: S1
In Pennsylvania, the bottleneck arrives before operations
The Pennsylvania dispute concerns a different point in the infrastructure chain: whether a planned site and its power connection can proceed. NorthPoint Development sent letters offering Hazle Township residents a potential payment in exchange for permission to build a 1,300-acre data center, part of a proposed $165 million investment. Nearby properties faced eminent-domain threats connected to a proposed high-voltage line intended to power the project.
Sources: S2
Residents formed Alliance to Stop the Line and lodged formal protests with the utility company, citing health and property-cost risks associated with the line. Others sought records about the proposed project’s end user; the report says emails identified Amazon. Amazon said it was evaluating multiple Pennsylvania sites and did not have a definitive agreement with NorthPoint. That qualification is important: public association with a prospective customer is not confirmation of a completed development commitment.
Sources: S2
Pennsylvania’s policy response shows how opposition can alter the path to construction. The state’s governor had announced Project Hazelnut after an executive order that accelerated permitting for AI infrastructure. Following backlash over projects in Hazle Township and nearby areas, he signed an order ending fast-track permits and giving local agencies, rather than the state-level Department of Environmental Protection, authority to approve data-center projects.
Sources: S2
Sources: S2
The shared dependency is not chips—it is continuity
These developments are not versions of the same event. The Yandex case is a wartime operational outage following physical attack. The Pennsylvania case is a local conflict over a prospective project’s footprint, grid connection, process, and accountability. One removes capacity already in service; the other may prevent or reshape capacity before it enters service.
Yet both cases expose a practical dependency often treated as secondary to computing hardware: a data center must have a secure site, a viable power path, and a credible route to sustained operation. In Vladimir, that continuity was interrupted after an attack. Around Project Hazelnut, residents are challenging the infrastructure needed to make a proposed campus viable, while the reported prospective customer says no definitive agreement exists.
Inference: for AI providers and buyers, capacity planning should separate the question of how much computing power is contracted from the question of how much can be delivered reliably under the relevant conditions. The supplied reporting does not establish the scale or duration of Yandex’s lost capacity, nor whether Project Hazelnut will be built. It does support treating those as open operational and development risks rather than settled outcomes.
A claimed buildout is not a tested service
The Pennsylvania reporting provides signs that local opposition can have wider effects, though its aggregate figures should be read as reported estimates rather than a universal measure of the market. A Data Center Watch report cited by Fortune said projects worth more than $130 billion were delayed or cancelled in the first three months of the year, while the number of active opposition groups increased over that period. A Gallup poll cited in the same report found local data-center opposition among respondents across party lines.
Sources: S2
The concrete test in both settings is whether the provider’s claimed route to service holds under stress. For Yandex, useful evidence would include information on which workloads were restored, where they were moved, the condition of the Vladimir site, and whether later attacks affected recovery. The material supplied says services were unavailable and resources were unstable, but it does not provide those performance details.
Sources: S1
For Project Hazelnut, the decisive evidence would be a final local approval or rejection, the outcome of the utility dispute, a confirmed customer agreement, and a clear account of the project’s power arrangement. The report documents public claims and government policy changes, but it does not establish that the campus will proceed, that Amazon will occupy it, or that the proposed line will be built.
Sources: S2
What to watch: recovery proof and local consent
In Russia, the near-term indicator is not a general statement of resilience but demonstrated recovery: whether Yandex can restore affected services and stabilize the resources it said remained at risk. The wider context is also unsettled. The report describes escalating Russian and Ukrainian attacks on technology infrastructure and quotes Ukraine’s president saying Ukraine was responding to Russian strikes on Ukrainian data centers.
Sources: S1
In Pennsylvania, the key signal is whether decision-making remains local and whether project sponsors can answer the concerns that organized residents have raised about power infrastructure, property, and transparency. The report quotes a political scientist arguing that weak communication about benefits, alongside clear communication about costs, has contributed to dissent. That is an explanation offered by a source, not proof that communication alone determines outcomes.
Sources: S2
The broader assessment is therefore conditional. Physical security failures can convert operating infrastructure into an emergency service problem; political and community resistance can convert announced investment into a prolonged approval problem. Both pressures make deployment schedules less dependable, but the evidence supplied does not permit a shared forecast for their frequency, financial impact, or eventual resolution.
Why it matters
The AI buildout is often framed as a race to secure computing equipment and electricity. These reports point to a more demanding standard: capacity must be physically defensible after deployment and politically workable before deployment. Decision-makers should ask for recovery evidence, site-specific power plans, and clear approval status rather than treating an announced campus or installed hardware as guaranteed usable compute.