Alphabet is studying the possibility of putting data centers in orbit. That sentence sounds like a pitch meeting for a science fiction film. It is not. It is a real estate problem wearing a spacesuit, and it is the clearest evidence yet of the argument worth making about this entire buildout: the constraint on artificial intelligence is not intelligence. It is land, power, and water, and the companies furthest ahead are the ones that started counting kilowatts before anyone forced them to.
Here is what is actually happening. The AI buildout has run into something no amount of software elegance dissolves, which is physics. Land costs money and takes time to acquire. Power grids sit at capacity in most of the markets that matter, and the queue to connect a new load is measured in years rather than quarters. Water for cooling is increasingly contested by people who live near the sites and vote in local meetings. The companies racing to build the next layer of AI infrastructure are not constrained by ideas, by models, or even by capital. They are constrained by kilowatts and square footage. When $GOOGL starts exploring orbital data centers to support long-term AI growth, the correct read is not that somebody is being visionary. The correct read is that every other option has already been examined and found short.
This is the part of the AI story that does not fit on a slide. The revolution is being paced by concrete, copper, and coolant, and none of those three respond to a faster release cycle.
What 25 Million Copilot Seats Actually Mean
Microsoft’s Copilot has reportedly crossed 25 million seats, and analysts see Azure growth above 40 percent alongside it. The seats number is the one collecting headlines, and it deserves them. Enterprise adoption at that scale means AI tools have cleared the pilot-project stage and entered the procurement budget, which is a far more durable signal than a usage spike. Pilots get cancelled quietly. Budget lines get defended.
Look at what the seats number implies on the other side of the ledger, though, because that is where it stops being a software story. Every one of those seats is a demand signal landing on the compute layer. Azure growing at 40 percent is not a separate fact from Copilot growing; it is Copilot growing, observed from the other end of the same pipe. The software story and the infrastructure story are one story, and the only reason they get reported separately is that different analysts cover them.
$MSFT is succeeding right now because it built demand and supply at the same time. Copilot created enterprise appetite, and Azure scaled to absorb it, and neither half would have been worth much without the other. The companies that will stumble are the ones that built the demand before the capacity, or the capacity before the demand. Timing those two against each other is harder than it looks from outside, and most companies miss in one direction or the other: either they sell something they cannot serve, or they pour concrete for a customer who never arrives.
Supermicro and the Unglamorous Middle
Supermicro’s pitch, validated building blocks for end-to-end AI infrastructure, is easy to overlook when the conversation is about language models and agents. It deserves attention precisely because it is dull. The infrastructure layer is where real money concentrates during a technology transition, not the application layer where everyone is visibly competing and margins get bid away by the fifth entrant.
Think about it in terms of which stories get told. The loudest stories in AI right now are about what the models can do. The quieter stories are about who manufactures the hardware those models run on, who cools it, who powers it, and who builds the buildings that hold all of it. The second set usually turns out to matter more, and for much longer, than the first. Modular components that can be assembled and validated as a unit are not interesting, but they compress the time between a signed power contract and a working rack, and in a market where the binding constraint is time-to-capacity, compressing that interval is worth more than a benchmark win.
What stays with me from reading through this period is a point about managing limits before demand forces your hand. The companies doing that right now, whether by exploring orbital compute, building modular infrastructure, or locking in power capacity before it becomes scarce, are not being dramatic. They are thinking one cycle ahead while everyone else is still thinking about this one, and the gap between those two postures is where advantage accumulates.
Apple’s Supply Chain Is the Constraint Nobody Names
$AAPL‘s AI pivot is being watched closely, and the valuation questions are legitimate. The company is trying to reframe itself as a platform for AI while managing a hardware supply chain that runs through territory it does not control. Analysts framing the story as Apple’s AI shift and Huawei chips testing valuation and growth expectations are pointing at the structural part rather than the narrative part, and the structural part is the one that binds. Apple’s ability to deliver on its AI promises depends on supply chains subject to pressures that have nothing to do with how good the product is or how well the software team executes.
That is not a crisis, and treating it as one would be an overreaction to a normal industrial fact. It is a constraint, and constraints compound quietly until the day they stop being quiet. The orbital data center study and the chip supply question are the same category of problem seen from two different balance sheets: a company discovering that the thing limiting its ambition is physical, located somewhere specific, and not for sale on the timeline it would prefer.
There is a shape to all of these stories that keeps recurring. Every large system is eventually limited by something physical and structural that cannot be abstracted away, no matter how much capital is aimed at it. Hitting the limit is not the impressive part, because every growing system hits one. The impressive part is whether you saw it coming before it arrived and bought yourself options while options were still cheap.
The difference between Alphabet exploring orbital data centers today and scrambling for power capacity in three years is only timing. One is a choice made from a position of slack; the other is a reaction made from a position of need, and they cost very different amounts. Systems that manage their constraints on their own terms tend to survive transitions intact. Systems that wait for the pressure to show up tend to get reshaped by it, on someone else’s schedule.
The AI story everyone is watching is about intelligence. The real story is about whoever owns the cooling system.
Leave a Reply