When the Grid and the Missile Both Miss

What happens when the machine you built to never go down actually goes down for good? Today on Concrete Compute: Amazon just admitted it cannot restore two cloud zones knocked out by Iranian strikes back in March — months later, permanently offline. Before that, in the headlines: Scotland's parliament takes up a moratorium on AI data centres today, and the EPA scraps Biden-era power plant emissions rules just as utilities race to build generation for data center load. This is a two-lead day — both stories test the same idea, that the buildout is always-on, from opposite directions, one from a missile and one from the grid itself. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Wednesday, September 16, 2026. Let's get into it.

Let's start in Edinburgh, where lawmakers are voting today on whether to hit pause on the country's AI data centre boom. Scottish Greens MSP Patrick Harvie forced the vote, arguing the government has, in his words, wavered and failed to take the action communities need. According to the Google News aggregation of this story, there are 24 proposed data centre sites in Scotland that could together draw six thousand megawatts if every single one gets built — but that's a big if, and the same reporting notes experts think it's unlikely all of them will actually happen. Right now only fifteen AI data centres are even operating in Scotland, and the largest one draws just twelve megawatts, so there's a real gap between the pipeline number and what's on the ground. On the other side, Sandy Begbie, who runs Scottish Financial Enterprise, told a Holyrood committee that the word moratorium is proving quite negative with investors, and pointed to a Morgan Stanley estimate — again per that same reporting — of up to three trillion dollars in global AI infrastructure investment by 2028. Outlander actor Sam Heughan has also weighed in publicly backing the pause, as The National reported on X. The vote's outcome wasn't confirmed as of this recording, so we'll see where Holyrood lands.

Now, sticking with policy — the EPA has scrapped the Biden administration's greenhouse gas rules for power plants and is moving to eliminate the rest of them too. Administrator Lee Zeldin says the agency doesn't have authority under the Clean Air Act to regulate plant emissions for climate purposes, leaning on the repeal of the 2009 endangerment finding and the Supreme Court's Loper Bright decision. The Edison Electric Institute, which represents investor-owned utilities, welcomed the move, saying regulatory certainty helps as they plan long-term investments for grid reliability. That's one side of it — Utility Dive's reporting didn't include environmental or public-health pushback in what we've got, though local officials and environmental groups have vowed legal action elsewhere. Here's why it matters for this show: utilities are racing to add generation specifically for data center load, and this repeal changes the compliance math on new gas plants right as that race accelerates. Worth watching how fast that translates into actual megawatts, because that's the number that counts.

Our main story today: what happens to the trillions being poured into AI infrastructure when the infrastructure gets bombed. Amazon Web Services this week posted an update to its Health Dashboard — its first public word since April — confirming it cannot restore access to resources in one of its United Arab Emirates availability zones, called mec1-az2, and to its entire Bahrain region. The damage goes back to March, when Iranian drones struck two AWS sites in the UAE and a nearby strike hit a facility in Bahrain, during the broader US-Iran conflict. AWS said the damage in Bahrain, quote, spanned multiple availability zones and exceeded what our regional and multi-AZ services are designed to withstand, end quote. That's a hyperscaler admitting, in writing, that its own disaster-resilience engineering — the redundancy across zones that's supposed to survive almost anything — got beaten by a physical strike. Iran's Revolutionary Guard Corps claimed a second, more destructive attack on the Bahrain region in July, and per the source material, says it's now hit AWS facilities five times since March. Two other UAE zones are still described as under active recovery, not permanently lost — so this isn't the whole region gone, but it's the first confirmed case of a hyperscaler's cloud infrastructure being permanently disabled by military action, and that's new territory for how this industry prices risk. Reuters reports the UAE is already reconsidering the design of a major AI data-center project, including dispersing facilities and building underground, blast-resistant infrastructure — which tells you the industry itself is treating this as a design problem now, not a one-off. Amazon opened the Bahrain region back in 2019 and the UAE region in 2022, and says it's working on replacing the affected infrastructure, with further updates on Bahrain promised in early 2027.

Here's the deeper read. This is a single-facility, war-zone-specific event — it is not evidence that a campus in Virginia or Ohio faces comparable risk, and I want to be careful not to overclaim that. But the editorial question worth sitting with is what this does to how insurers and lenders price data-center risk in politically unstable regions going forward, and whether hyperscalers now avoid concentrating capacity in a single geopolitically exposed market, even at the cost of efficiency. My read: this is exactly the kind of tail risk that gets modeled wrong until it happens once, and now it's happened once, on the record, from the company itself. Time for the Hype Check. I'd put this at an 8 for substance — it's not a press release, it's AWS's own dashboard admitting permanent data loss in one zone and an entire region, corroborated across Reuters, CNBC and multiple regional outlets. Where it stays short of a 10 is scope — two other UAE zones are still recoverable, and this is one operator in one conflict zone, not a verdict on the whole buildout.

Now, the thread connecting today's two lead stories: both are stress tests of the same assumption, that the AI buildout is basically indestructible once the concrete's poured. AWS shows what an external shock — a literal missile — does to that assumption. PJM shows what internal strain does to it. A study commissioned by the Pennsylvania Public Utility Commission, done by Synapse Energy Economics, found that PJM's grid is projected to miss its own reliability standard from 2027 through 2030, largely because of data center load. In the study's reference scenario, the 2030 loss-of-load expectation comes in at nearly six times worse than PJM's planning criterion; in a worst-case scenario of higher load growth and constrained supply, it's over a hundred times worse, implying more than thirteen days a year with actual loss-of-load events. Only a scenario with no new data centers meets PJM's target. PUC Chairman Steve DeFrank called Pennsylvania at an energy crossroads, and the commission has already put in place a large-load tariff requiring customers needing fifty megawatts or more to bear the costs they add to the system — that's the beneficiary-pays principle I keep coming back to on this show, and it's good to see it actually written into a tariff rather than just promised. PJM itself, through spokesperson Jeff Shields, acknowledged new data center loads are growing faster than supply, while pointing to its Ratepayer Protection Pledge from developers. So: one story is a hyperscaler's redundancy failing against an external attack, the other is a regional grid's redundancy potentially failing against its own customers' growth. Neither one breaks the buildout thesis — but both say the same thing in different languages: assume less resilience than the marketing promises, and build the accounting for the failure mode in from day one. Hype Check on the pairing: a 7. Real, independently modeled numbers on the PJM side, real admitted damage on the AWS side — what's missing from both is PJM's own rebuttal to the study's methodology, which nobody's put on the record yet.

Two follow-ups worth flagging: we'll be watching what today's Holyrood vote in Scotland actually decides, and whether Pennsylvania's large-load tariff survives contact with the utilities that have to implement it. If you're finding these two stories useful together, follow Concrete Compute wherever you listen so you don't miss where they go next. This has been Concrete Compute, an AI-voiced podcast, created and built by a real human using today's cutting-edge technology. Nothing you heard on this show is financial advice. I'm Brian Lampert, and I'll catch you all tomorrow — take care!

I also host Quickly Quantum: a daily quantum computing briefing you don't need a physics degree to follow. The breakthroughs, the funding rounds, and how much substance is really under each claim. Find it wherever you get your podcasts.