Today on Concrete Compute: what happens when missile defense starts dictating where you put a five-gigawatt AI campus? Reuters is reporting the UAE is quietly redesigning its flagship AI data center project after this year's war with Iran exposed just how vulnerable one giant campus really is. Before that, in the headlines: Microsoft is reportedly eyeing a tripling of its global data center capacity by 2032, Colorado's Weld County just approved a permit for what could become one of the state's largest AI data centers, Australia's NextDC went back to the capital markets for the third time in four months, and DISA opened bidding on a cloud contract with a ceiling of $21.6 billion. Welcome back to Concrete Compute, your daily brief on the AI infrastructure buildout. It's Friday, September 11, 2026. Let's get into it.

Picture Microsoft's entire current data center footprint — every leased and owned site combined — more than tripled. That's the plan Bloomberg is reporting, according to Yahoo Finance's write-up of that reporting, citing people familiar with the matter: Microsoft aims to grow its total compute capacity from roughly twelve gigawatts today to more than thirty-eight gigawatts by 2032. And the AI-specific slice grows even faster in that plan — from about two gigawatts now to roughly a third of that whole thirty-eight-gigawatt footprint, Yahoo Finance (Bloomberg reporting) says. Now, why the sudden hunger? The same reporting says hardware shortages have gotten bad enough that Microsoft has had to turn away cloud and AI business and deal with real disruptions — GitHub reportedly had an eight-hour outage tied to server constraints, and Xbox capped cloud game streaming time. Here's what should make you pause: this is an internal plan surfaced through anonymous sources, not a signed commitment — sources told Bloomberg the projections could shift given how long these builds take. Tripling a company's entire global footprint over the next six years is an enormous logistical bet, even before you get to permits, power contracts, and chip supply. File it as a target, not a blueprint.

In Windsor, Colorado, Weld County commissioners voted five to zero this week to approve a special-review permit for a data center on the old Kodak film plant site — a project Global AI says could eventually grow into a full gigawatt of capacity by 2029, though that figure is the company's own website talking, not anything on paper. What actually got approved is far smaller: the 41-acre parcel itself, not the full campus Global AI has floated. This one got contentious. Scott James, the commissioners' chair, called AI 'absolutely essential' for aerospace and military uses and said the county wants to stay friendly to those uses. Resident Patricia Nelson-Garcia pushed back, pointing out the company is 'only two years old' with 'no track record in Colorado.' And Hugh Gottschalk, an attorney representing a group called Windsor Renewal, zeroed in on power — he said the local utility's supply letter 'doesn't estimate or allocate the power that is needed' and only guarantees three years of it. So what's actually approved is a fraction of what's been floated — the gigawatt dream remains just that, with no construction timeline behind it.

Three capital raises in a little over four months — that's the pace Australia's NextDC is setting, and this week's installment is the biggest yet: DCD reports the company priced an AU$1.1 billion convertible notes offering, with the notes maturing in September 2031. Why does that matter beyond the balance sheet? NextDC counts OpenAI among the customers lined up for its upcoming S7 data center in Sydney. Now, three raises in four months tells you something about how capital-hungry this buildout has become — and my suspicion is that money isn't the binding constraint here anymore. Power is.

Twenty-one point six billion dollars — that's the ceiling value on a tender DISA just opened, but before you fixate on the number, know what it actually is: a bid invitation, not a signed check. DCD reports the Defense Information Systems Agency has begun accepting bids for the next round of Joint Warfighter Cloud Capability contracts, with proposals due October 6th. The last version of this contract, worth nine billion dollars, went to AWS, Microsoft, Google and Oracle back in December 2022 — before that, there was the JEDI contract, the single-vendor deal that was supposed to go to Microsoft alone and got tangled in an AWS lawsuit alleging presidential interference, no less. This new round covers edge compute at unclassified, secret and top-secret security levels. So here's the question worth asking: do the same four incumbents split it again, or does this open a door for someone else — Oracle's been climbing fast, and the neoclouds want in too. Hold that thought, because our main story is about infrastructure decisions shaped by forces well outside the data center fence.

Our main story today: does a missile war change where you build a five-gigawatt AI campus? Reuters has an exclusive out this morning, citing six sources — a U.S. official, a Western diplomat, and industry executives — who say the United Arab Emirates is quietly rethinking its flagship AI data center project after this year's war with Iran. Here's the original plan: a single, ten-square-mile campus in Abu Dhabi, five gigawatts of capacity, built as essentially one enormous target. Now, according to those sources, the UAE is considering breaking that single campus into a dispersed network of facilities instead — spreading the risk out geographically. Think of it the way you'd think about not keeping all your savings in one bank: if a missile can only hit one site at a time, you don't want your entire AI capacity sitting behind one fence line. Now, it's worth being careful here, because Reuters describes a project still very much in flux — this isn't a groundbreaking or a finalized blueprint, it's a look inside a design conversation currently underway. G42, the Emirati AI company at the center of this, gave Reuters a statement worth reading closely for what it doesn't say: the project's specifics are under, quote, 'continuous review in line with the security, resilience and operational standards expected of critical infrastructure at this scale,' end quote. That's a company confirming it reviews things constantly — true of basically every major infrastructure project on Earth — without confirming or denying the specific redesign Reuters is describing. So why should you care about a redesign happening on the other side of the world? Because if hardening against military strikes becomes a real design input for gigawatt-scale AI campuses, that's a new variable in the siting calculus hyperscalers haven't had to weight this heavily before — alongside power, water, chip supply, and local permitting fights. And it raises the obvious follow-on question Reuters itself couldn't answer: how close is the UAE actually to finalizing any of this, and what does going from one campus to a dispersed network do to the cost and timeline of a project this size? Reuters says it couldn't determine either one. That's where we go next.

Let's dig into what a redesign like this actually means in practice, because 'dispersed network of facilities' sounds simple until you think about what makes a five-gigawatt AI training campus different from, say, five separate one-gigawatt facilities scattered across a country. The whole reason hyperscalers build enormous single campuses in the first place is interconnect — the ultra-high-bandwidth, low-latency links between racks of GPUs that let you treat thousands of chips as one giant computer for training the biggest AI models. Split that same GPU count across sites separated by real distance, and you introduce latency that can matter enormously for the largest training runs, even if it barely registers for smaller inference workloads. So a shift to a dispersed model isn't just moving walls around — it potentially changes what kind of AI work the facility is even good at. That's exactly the kind of detail Reuters says it couldn't pin down: the reporting doesn't tell us whether this dispersed network is still meant to function as one giant training cluster, or whether the UAE is accepting a different mission for the site altogether. Now, is this an overreaction to one war, or the start of a pattern? I think that's the real question underneath the reporting, and it's bigger than the UAE. The Gulf has become one of the most concentrated bets on AI infrastructure anywhere on the planet — enormous capital and chip allocations flowing into a small number of countries in a genuinely unstable neighborhood. If air-defense planning becomes a standard line item for gigawatt-scale AI sites — the way flood risk or seismic risk already is — that changes where the next generation of hyperscale campuses gets sited, not just in the UAE but anywhere near a live conflict zone. But here's where I want to plant my skepticism firmly, because this is a single Reuters exclusive built on six anonymous sources describing a plan that, by Reuters's own admission, is still in flux. Reuters says it couldn't determine how close the UAE actually is to finalizing any new masterplan, or how a shift like this would affect the project's cost or timeline. And G42's actual on-record response — that continuous-review language — didn't confirm the specific redesign details Reuters is describing. That's a company being careful, not a company confirming a pivot. So my read: treat this as a genuinely interesting signal about how the AI buildout's geopolitics are evolving, not as a finished decision you can bank on. Going from an announced ten-square-mile Abu Dhabi campus to a dispersed network of smaller sites is an enormous design and cost swing to execute on the say of anonymous sources alone. What would turn this from a credible story into a confirmed one is simple: a finalized masterplan, a permit filing that shows the new site plan, or G42 actually saying the words 'we are redesigning this project' on the record instead of talking about continuous review. Time for the Hype Check. I'm putting this one at a 5. The sourcing is serious — a U.S. official and a Western diplomat aren't nothing — and the strategic logic is sound: nobody wants five gigawatts of AI infrastructure sitting behind one fence line in an active-conflict neighborhood. But nothing here is finalized, G42 didn't confirm the specifics, and Reuters is upfront that it can't tell us the cost or the timeline. Real signal, no steel in the ground yet.

So the open question hanging over this one: if the UAE ends up rebuilding its flagship AI campus as a dispersed network, does that become the new normal for hyperscale sites anywhere near a conflict zone — or is this a one-country reaction to one war? If you're finding these episodes useful, go ahead and follow Concrete Compute wherever you listen, so it's just there for you tomorrow morning. 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 Space Stakes: the business of the new space race, every day. What actually flew, what the contract is really worth, and who has customers. Find it wherever you get your podcasts.