The Cloud Has a Physical Address

At 4:30 in the morning Pacific time on March 1, something hit an Amazon Web Services data center in the United Arab Emirates. AWS's initial status page update described the cause as "objects." The objects created sparks, then fire. Local authorities cut power to the facility and its backup generators to contain the blaze.

Within six hours, a second availability zone in the same AWS region lost power. By morning in Bahrain, a third facility reported infrastructure damage from a drone strike nearby. Roughly 60 AWS services went offline across the Middle East. AWS told customers to enact their disaster recovery plans and migrate workloads to Europe.

Twenty-three hours after the first strike, Claude went down.

We noticed because our AI-powered newsroom stopped working. Our automated publishing pipeline runs on Anthropic's API and Google's Gemini API. Overnight on March 2, both services entered cooldown simultaneously. By morning, nothing had published. No summaries, no analysis, no automated workflows. The entire pipeline was dead.

The cause was 7,500 miles away.

The Objects

The "objects" were Iranian drones. On February 28, the United States and Israel had launched Operation Epic Fury, a massive strike campaign against Iran. Iranian state media reported that Ayatollah Ali Khamenei and senior military leadership were killed in the initial strikes, though independent confirmation remained difficult amid the fog of war. Iran retaliated the following day with strikes across the Gulf. The UAE defense ministry reported 2 cruise missiles, 165 ballistic missiles, and more than 540 drones launched at its territory alone. US military installations in Bahrain, Saudi Arabia, Kuwait, and Qatar were also targeted.

AWS data centers were almost certainly not the intended targets. They were caught in a broad-area barrage. But the result was the same. On March 2, AWS confirmed that "two of our facilities were directly struck" in the UAE, while in Bahrain, "a drone strike in close proximity to one of our facilities caused physical impact to our infrastructure." The company described structural damage, disrupted power delivery, and fire suppression activities that caused additional water damage from sprinkler systems soaking server racks.

Reuters reported it was the first time a major U.S. tech company's data center had been disrupted by military action.

Designed for One, Not Two

AWS builds its regions with redundancy. Each region contains at least three availability zones, which are physically separate data center clusters with independent power and networking. The design assumes that any single zone can fail without taking down the region. Customers spread workloads across zones so that if one catches fire or loses power, the others keep running.

ME-CENTRAL-1 in the UAE has three availability zones. Two of them went down.

S3, Amazon's storage service, is designed to survive the loss of a single zone. When two of three zones in a region are impaired, the math breaks. The Register reported that customers saw "high failure rates for data ingest and egress." Data going in failed. Data coming out failed. The redundancy model that underpins most cloud architecture was built for component failures and software bugs, not for 540 drones.

Mike Chapple, an IT professor at Notre Dame, told the AP: "Cloud computing isn't 'magical' and still requires physical facilities on the ground, which are vulnerable to all sorts of disaster scenarios." The AP noted that unlike previous AWS outages caused by software bugs that rippled globally, this physical damage produced "localized and limited disruption."

Localized to the Middle East, yes. Limited is debatable.

Claude Goes Down

On March 2 at 11:30 UTC, Anthropic's Claude began returning HTTP 500 and 529 errors worldwide. Claude.ai, Claude Code, the developer console, and the API were all affected. Bloomberg reported that Anthropic attributed the outage to "unprecedented demand."

By 3:25 PM UTC, Anthropic said a fix had been implemented. By 4:37 PM, the company told Mashable that "Claude is back up and running." By 4:50 PM, Claude went down again. Elevated errors on Claude Opus 4.6. Then, at 4:43 AM UTC on March 3, a third round of errors hit across the entire platform. Another fix, another monitoring period. At 10:27 AM UTC, Opus 4.6 errors spiked again.

Three outages in 24 hours. The official explanation each time: demand.

Anthropic runs its core inference infrastructure on AWS. Amazon has invested $4 billion in Anthropic and is building an $11 billion data center in Indiana. Anthropic's production inference runs mostly in US regions, not in the Middle East. So the question becomes: how does a drone strike in the UAE affect an AI service running out of Virginia or Oregon?

Several mechanisms are plausible. When two AWS regions go offline, traffic from those regions reroutes to the nearest healthy infrastructure. If US regions were already running near capacity, an influx of rerouted Middle Eastern workloads could push them over. AWS doesn't publish capacity utilization numbers, but GPU inference capacity is known to be constrained globally. The CSIS flagged the Middle East as a key transit point for internet backbone traffic between Asia and Europe, with "fiber chokepoints such as the Bab al-Mandab Strait" carrying significant undersea cable traffic. Physical disruption to regional infrastructure could cause routing instability that touches continents.

There's also a simpler explanation. AWS itself told customers to "enact their disaster recovery plans and recover from remote backups into alternate AWS Regions, ideally in Europe." When the cloud provider tells everyone in the Middle East to move their workloads to other regions at the same time, that is a capacity event.

Gemini had problems too. DownDetector showed a spike in Gemini API reports starting at 4:12 AM Eastern on March 2. Third-party monitors flagged issues. Reddit users reported outages not reflected on Google's official status page. Google did not publicly acknowledge problems.

Whether the Claude outages were directly caused by the AWS Middle East damage, amplified by it, or purely coincidental is something only Anthropic and Amazon know. The timing is what it is.

What is not ambiguous is the dependency. Anthropic has one cloud provider. That cloud provider just lost physical infrastructure to military strikes for the first time in its history. The company's public explanation cited demand, not infrastructure. But the global cloud is not a collection of isolated rooms. It is a network. When part of the network burns, the rest of the network absorbs the load, and load has consequences.

Compute Is the New Oil

On February 27, one day before Operation Epic Fury launched, the Center for Strategic and International Studies published an analysis titled "If Compute Is the New Oil, War in the Gulf Significantly Raises the Stakes."

The piece warned that in previous Gulf conflicts, Iran and its proxies had targeted pipelines, refineries, and oil fields. The targets would now include "data centers, energy infrastructure supporting compute, and fiber chokepoints such as the Bab al-Mandab Strait off the coast of Yemen, which hosts critical undersea data cables."

The CSIS analysis referenced the State Department's Pax Silica Declaration: "If the 20th century ran on oil and steel, the 21st century runs on compute and the minerals that feed it." The Gulf states had been diversifying from petroleum wealth to compute wealth. AWS committed $5.3 billion to Saudi Arabia. Microsoft announced $15 billion for the UAE. The pitch was cheap energy, minimal regulation, and governments willing to co-invest at scale.

CSIS saw the flaw. If compute is the new oil, then compute infrastructure inherits the same targeting logic that has governed Gulf conflicts for 50 years.

CSIS published that warning on a Thursday. The strikes began Friday. The data centers burned Saturday.

The Pentagon Irony

Five days before Operation Epic Fury, Defense Secretary Pete Hegseth had given Anthropic a deadline. The Atlantic characterized the demand as: "Strip the ethical guardrails from your AI models by Friday or face the full weight of the state."

Anthropic rejected the ultimatum. The company dropped its Responsible Scaling Policy but maintained red lines against mass surveillance of Americans and autonomous weapons without human oversight. The New York Times called it "a decisive moment for how A.I. will be used in war." The Pentagon blacklisted Anthropic. Defense One reported it would take the military months to replace Anthropic's tools in its systems.

Then the war the Pentagon launched damaged the cloud infrastructure that Anthropic runs on, and the AI tools the Pentagon had been trying to commandeer went offline.

The Pentagon's two strategies during the same week were to coerce Anthropic into providing unrestricted military AI and to wage a war across the Gulf region where Anthropic's cloud provider operates critical infrastructure. Nobody in the room appeared to notice.

The cloud, in the defense establishment's framing, existed nowhere. It turned out to exist in the UAE.

The Category Change

Data centers have been damaged before. In March 2021, a fire at OVHcloud's facility in Strasbourg destroyed a five-story data center and took down 3.6 million websites, including government portals. Customers who lacked off-site backups lost data permanently. The cause was accidental: inadequate fire suppression and poor building materials.

The Texas freeze of 2021 knocked power offline across the state, taking some data centers with it. The 2011 Japan earthquake forced Tokyo-area facilities onto backup generators for weeks. These were infrastructure failures caused by natural disasters or accidents. They shared a common feature: nobody intended to hit the data centers.

What happened in the Gulf was different in kind. The Iranian drones were not aimed at AWS. But they were aimed at the UAE, and AWS chose to put buildings there. The Indian Express described it as "one of the first instances of a major global cloud provider's data centre being directly hit during military action": "Just as oil fields and power plants once defined economic warfare, data centres may increasingly sit in the crosshairs during geopolitical crises."

Google, Microsoft, and Oracle all operate data centers in the UAE. None reported damage from the March 1 strikes. All of them are within range.

What Breaks Next

AWS, Microsoft Azure, and Google Cloud together host roughly 65 percent of global cloud infrastructure, according to Synergy Research Group. Anthropic runs its inference on one of those three. A single provider, a single dependency chain, a single point of failure that extends from a server rack in Abu Dhabi to a chatbot answering questions in Tokyo.

The cascade traveled fast. Abu Dhabi Commercial Bank's platforms and mobile app went down. Snowflake attributed its own regional disruptions to the AWS outage. Physical infrastructure to cloud provider to SaaS platform to end user to our dead publishing pipeline, 7,500 miles from the nearest crater. Each hop involved a different company, a different contract, a different risk model. None of those risk models priced in Iranian drone swarms.

AWS warned customers that the "ongoing conflict means broader operating environment remains unpredictable" and recommended migrating workloads to alternate regions. The word "unpredictable" was doing a lot of work. AWS was telling customers, in corporate language, that their data might not be safe and the company could not say when it would be.

Direction, Not Speed

There is a policy conversation happening about cloud resilience, and it has been happening for years. Multi-cloud architectures. Geographic distribution of workloads. On-premise inference for critical systems. These are not new ideas. They are expensive ideas, which is why most organizations treat them as optional until the moment they are not.

The Gulf strikes did not create the vulnerability. They revealed it. You give up control of where your data lives in exchange for lower costs and convenience. That bargain works until geography reasserts itself.

For AI companies, the dependency chain is unusually tight. Most startups are locked to a single cloud provider through GPU allocation agreements and volume pricing. When that assumption of always-available compute broke for 48 hours in March, organizations that had no backup plan discovered they had no backup plan.

The alternatives exist — multi-cloud, on-premise inference, geographic diversity. Every resilience engineer knows this. The obstacle is cost. Redundancy is expensive during the 99.9 percent of the time when nothing is wrong. It becomes cheap in retrospect during the 0.1 percent when everything is.

The CSIS analysis, published the day before the bombs fell, asked the right question: what happens when the new oil fields are in the same place as the old oil fields? The Gulf states bet that compute would be different from petroleum. That data centers would bring stability in ways that oil derricks never did. That the 21st-century resource would not inherit the 20th-century geopolitics.

For one weekend in March, that bet lost.

The Cloud Has a Physical Address

Our publishing pipeline came back online on March 3 as Anthropic's services stabilized. We wrote this piece using the same tools that went down.

Three days ago, a think tank published a paper warning that Gulf data centers could become military targets. Two days ago, they did. The AI services that millions of people and thousands of businesses depend on went intermittent for 48 hours, and the company running them attributed it to demand while its cloud provider was pulling server racks out of sprinkler water in Abu Dhabi.

Every organization running on cloud infrastructure has a physical dependency it cannot see. The API endpoint is an abstraction. Behind it sit buildings, in specific cities, in specific countries, connected by specific undersea cables that run through specific chokepoints. Those buildings can lose power. They can catch fire. They can be hit by drones.

The cloud is not a metaphor. It is concrete and steel and fiber optic cable. It has a physical address. And last weekend, that address was in the blast radius.