Are AI Data Centre Builds Really That Impressive?

TL;DR

Two countries have repeatedly proven they can industrially scale a new technology at a world-beating pace: the United States since the 1940s, aircraft production, Apollo, now Starlink, and China since the early 2000s, EV manufacturing, high-speed rail. That’s not a one-off for either. It’s a demonstrated, repeatable capability, and both are now industrial peers in this specific sense, with a lot worth studying in how each did it. Everyone else currently attempting the same thing, Amazon’s satellite constellation, Anduril’s defence manufacturing, even China’s own newer satellite programs, is still building toward a standard those two have each hit more than once. Robotics is the live open case: forecasts claim a pace that would beat even that standard, but actual humanoid shipments today are still only in the tens of thousands. Ukraine’s volunteer IT Army shows a different kind of speed entirely, zero to 300,000 people in a month, once you remove the physical constraint altogether.

History already showed what the outer limit looks like. AI isn’t there yet.

AI is diffusing like every digital technology before it, concentrating faster than any of them. People just don’t realise how fast “fast” has already been.

eleven series indexed to each one’s own start

A quick definition, because it does real work here: the S-curve

Almost every ramp on this chart moves through the same three phases.

Early adoption is slow and easy to dismiss: a small base, unproven capability or high cost, output that barely registers next to anything established. Humanoid robots sit here right now, forecast shipments still in the tens of thousands. China’s EV industry sat here in 2015, 331,000 units a year that nobody outside the industry noticed.

Scale is the middle phase: the real constraint breaks, capital or capability or urgency align, and output compounds fast enough to look exponential. This is where Ireland’s and Australia’s data centre curves sit today, where China’s EV ramp spent 2017 to 2023, and where US aircraft production and Apollo spending both spent their fastest years. On an ordinary linear chart this is the steep part of the S. On this chart’s log scale, it’s the stretch that runs as a straight diagonal, which is the whole reason for using a log axis here: it’s the only way to fit a 1x series and a 60x series on one honest scale without squashing the small ones flat, and it turns the exponential middle phase into a line you can actually compare across series.

Maturation is the bend: growth decelerates as the technology saturates whatever it was feeding, a market, an industrial capacity, a war. Apollo flattens near 11x once its own funding ceiling arrived. Industrial robots, 60 years past their own disruption, barely climb at all, 8% a year, pure maturation. US aircraft production did something starker: it hit its ceiling and then the underlying demand disappeared entirely, so the curve doesn’t just flatten, it turns over and falls.

This is why Amara’s law holds: forecasters anchor on the flat early-adoption years and get run over by the steep scale phase, because they’re extrapolating the wrong part of the curve. And it’s why most series on this chart haven’t reached maturation yet. Scale is not the same as maturity. A series still climbing steeply has told you nothing yet about where its ceiling actually is, only that it hasn’t found it.

Index every series in the chart above to its own starting year and the shape becomes comparable across decades. From 1939, the United States multiplied aircraft production 45-fold in five years, 114% a year, by converting car plants that already existed. That is the ceiling: what a society can do when capital, labour and urgency align completely.

Apollo proved it doesn’t take a war

Indexed to 1961, Kennedy’s “before this decade is out” speech, NASA’s Apollo spending grew elevenfold by 1966, 61% a year, on nothing but a deadline and a budget, in peacetime, by democratic choice. A free society hit a meaningful fraction of the wartime ceiling voluntarily.

Starlink shows the commercial version of the same ceiling, and it’s why Amazon Leo isn’t on the chart

Operational satellites went from 120 in 2019 to 10,063 by 2026, 88% a year, close to the wartime pace, done by a private company with no war or moonshot speech forcing it. Amazon Leo, formerly Project Kuiper, is chasing the same market with comparable capital and the same physics. Indexed to its own first production launch, 27 satellites in April 2025, its 396 satellites by September 2026 would compute to above 500% a year, faster than every series on this chart, including the wartime ceiling. That’s not Amazon succeeding. It’s the tell that indexing lies once the starting base is tiny enough: almost anything looks like a steep S-curve in its first few months. The honest comparison is absolute, not relative. Starlink has 10,063 operational satellites. Amazon has 396 in orbit, but has already manufactured its 1,000th, building them at up to 30 a week, faster than it can launch them. Its constraint isn’t capability or capital, Amazon has reserved over 100 future launches, the largest commercial launch procurement in history. It’s cadence: roughly 600 finished satellites are sitting in storage waiting for a rocket, and Amazon still had to ask the FCC for more time because it couldn’t get half its first-generation constellation, 1,616 satellites, into orbit by its own July 2026 deadline. Amazon’s manufacturing line isn’t the problem, it’s actually outpacing Starlink’s own early rate, producing 30 satellites a week against Starlink’s roughly 17 a week average in its first full year. Same technology, same market, comparable capital, comparable production, and the gap still turned out to be the one thing money doesn’t buy outright: a reliable launch cadence.

Anduril shows the same pattern, with sharper branding

Revenue went from $1 billion in 2024 to a projected $4.3 billion in 2026, about 107% a year, nearly matching the wartime ceiling, and its valuation quadrupled to $61 billion over the same window. That part is real. But its Ohio manufacturing campus is named Arsenal-1, a direct echo of the wartime “Arsenal of Democracy” this piece opens with, and the actual output doesn’t match the name yet: up to 150 Fury aircraft a year at full tilt, a Pulsar production line just doubled from 500 to 1,000 units a year, Barracuda cruise missiles not shipping until 2027. The company is projected to lose $1.2 billion in 2026 against that $4.3 billion in revenue, capital and contracts running well ahead of delivered volume. Judged against its own name, Anduril isn’t at scale yet. Judged against everyone else it’s competing with, that’s the wrong comparison. The traditional defence primes and most adjacent deep-tech sectors aren’t failing to hit Arsenal-of-Democracy numbers either, they’re mostly not attempting serial production at all, still selling low-rate, hand-finished systems on multi-year timelines. What Anduril has actually demonstrated is the harder, less visible achievement: standing up a real factory, a real production line and a real cadence problem to solve, which is a different kind of progress to a prototype on a stage. As Musk has put it often enough for it to be worth repeating, prototypes are easy, scale is hard. Most of the industry is still stuck on easy. Anduril is the one actually fighting the hard part, even if it hasn’t won it yet.

One distinction worth making explicit here: money raised is not the same signal as capability scaled, and defence is the sector most prone to confusing the two, because a funding round or a contract ceiling is a headline and a delivered unit count is a footnote. The deeper issue predates any one company. Western defence procurement has spent decades optimising for exquisite, low-volume, extremely capable platforms, carriers, fifth-generation fighters, precision-guided everything, over cheap, numerous, attritable ones. Exquisite systems are superb at winning engagements, and that’s a different thing from winning the actual objective, or from being able to keep fighting. The 2026 Iran war is the live, uncomfortable example, on both sides of that distinction. Seven weeks of war used up roughly 45% of the US Precision Strike Missile stockpile and close to half of both its THAAD and Patriot interceptor inventories, 3,700 to 4,500 advanced munitions in total. A Tomahawk takes around 47 months to build. Rebuilding what seven weeks used up could take up to four years, and the shortfall is already cutting into what’s available for Ukraine, the Gulf states and other allies. Iran, meanwhile, hit back at 20 to 28 US bases across the region, disabled the 5th Fleet’s own Bahrain headquarters for a time, and still disputes losing control of the Strait of Hormuz, with oil prices still up around 60% since the war began, something the IEA has called the largest supply disruption in the history of the global oil market. Overwhelming tactical dominance with exquisite, low-volume platforms didn’t translate into the strategic objective, and it left the stockpile too thin to comfortably fight the next one. Ukraine’s cardboard Corvo drones, part of this same story, sit at the opposite end of that spectrum: cheap enough to lose, numerous enough that losing some doesn’t matter, and fast enough to replace that a depleted stockpile isn’t a four-year problem. In peacetime especially, scaling the useful and attritable looks like it matters more than scaling the exquisite, and almost nothing in how Western defence budgets get allocated currently rewards building that way. This is the frame Anduril’s own numbers belong in. A $61 billion valuation and a 10-year, $20 billion Army contract are real money moving away from the traditional primes, and that reallocation is itself newsworthy. But redistributing defence spending from one incumbent to a newer entrant is not the same achievement as fixing the sector’s actual scale problem, and on the evidence so far, Anduril hasn’t fixed it either, it has absorbed a larger share of a budget that still isn’t buying the US the attritable, mass-producible capacity the stockpile numbers above show it actually needs.

Urgency alone isn’t sufficient, though

Germany had the same wartime pressure as the US and managed only 37% a year before its industrial base collapsed under bombing. China’s EV industry shows the other side: no war, no deadline speech, just a decade of patient industrial policy, and it still produced a 48-fold increase in NEV output since 2015, 47% a year, a case study for what structural depth achieves without existential urgency. China’s high-speed rail network shows the same patience taken further: 672km in 2008 to 48,000km by 2025, 71-fold over 17 years, only 28% a year, a lower rate than almost everything else on this chart but sustained for so long the total multiple nearly catches Starlink’s seven-year sprint. Rate and duration are different levers, and this is what pulling the slow one, for longer than anyone else here, actually buys.

China’s own satellite constellations are interesting for a different reason: the same state that built the high-speed rail network hasn’t yet brought that capability to this particular technology

Qianfan, the Thousand Sails constellation, has about 256 satellites in orbit as of October 2026, well short of its own published target of roughly 600 by the end of 2025. Guowang has launched 163 satellites since December 2024, a rate of about 140 a year, a sixth of Starlink’s own pace in its first full year. Index either one to its own tiny first batch and the multiple looks dramatic, over 200% a year, the same statistical tell Amazon Leo’s number gives off. The honest read, on the absolute numbers, is that this particular programme is still ticking along, years behind Starlink, which is itself revealing: even a country that has done this before, repeatedly, has to rebuild the capability for each new technology. It doesn’t transfer automatically, even domestically.

Which is really the finding underneath all three of these cases, Amazon with private capital, China’s own newer satellite programme, Anduril with defence contracts. None of them are short of money, ambition or manufacturing capability. All three are building toward the same thing: a reliable, repeatable production and deployment cadence, the unglamorous E in STEM that doesn’t show up in a funding round or a state plan. On this chart, only two players have actually hit that standard, and each has hit it more than once, in completely different eras, political systems and technologies. The United States: aircraft production in the 1940s, Apollo in the 1960s, Starlink now. China: EV manufacturing and high-speed rail, both since the early 2000s. That repeatability, not any single programme, is the real signal. It’s a capability a country or a company can build, demonstrably, because these two have each built it more than once.

There’s a reason it keeps being these two, and it’s not luck: scaling is recursive. You need to already be operating at industrial scale somewhere adjacent before you can pivot into scale somewhere new, fast. The US in 1939 wasn’t inventing mass production from nothing, it was redirecting a civilian automobile industry that already ran at that scale. SpaceX spent the better part of a decade scaling Falcon 9’s own launch cadence before Starlink had a rocket to ride on. China’s EV and rail buildouts drew on four decades of general manufacturing scale, built since the 1980s, before either one took off. Amazon, Anduril and China’s own satellite programmes are each trying to build a genuinely new scale capability closer to from scratch, without an adjacent industrial base already running at the pace they need underneath them. That’s not a knock on any of them, it’s the actual mechanism. Everyone chasing the US and China, including parts of their own governments and closest allies, isn’t just behind on this one technology. They’re missing the base the US and China were each already standing on before they started.

Today’s AI data centre buildout sits well under that ceiling

Ireland’s actual electricity-share curve and Australia’s own AEMO forecast both compound around 20% a year, roughly a fifth of the wartime pace. The US, the poster child for AI creation and adoption, is slower still, about 15% a year, not because the commitment is smaller but because it is scaling an already-enormous base. Scaling 183 TWh of data centre demand by 15% a year adds more new capacity than Ireland’s entire grid; it just looks modest as a multiple of where the US started. Flow and stock, again: the number that gets published is the multiple, and the multiple hides the base.

the same eleven ramps, as one annualised growth rate each

Robotics is where this is still an open question

Industrial robots, a 60-year-old technology now past its own disruption, grow at 8% a year, global installations roughly doubling in a decade: this is the “no renewal event happening” baseline none of the historical cases can show, because none of them are old enough yet to have settled into one. Humanoid robots are at year zero. Bank of America’s own shipment forecast has them compounding at 127% a year to 2030, past even the wartime ceiling, if it holds; Goldman’s competing forecast for the same window is 86%, and Goldman’s own number was revised from 256,000 to 890,000 units within 2026 alone. Nobody knows which curve humanoid robotics will follow, industrial robotics’ slow one or the wartime-fast one. That uncertainty is the honest finding, not a prediction either way.

One case doesn’t fit on either chart, and the reason is the finding

Ukraine’s volunteer IT Army went from nothing to 300,000 Telegram members within a month of its February 2022 launch, with 3,000 to 10,000 people actively running operations at any time. Every series above took years to multiply because each was bounded by physical capital that has to be built before output can grow, a factory, a launch pad, a data centre. Ukraine’s mobilisation had no such constraint. It moved attention and software skill, not steel, so its clock ran in days, not years. That is the other ceiling: remove the physical bottleneck entirely and renewal stops looking like an S-curve at all.

The lesson isn’t that AI is moving impossibly fast

It’s that disruptive renewal has a known, recurring shape, a ceiling history has already reached more than once, and a floor set by whatever constrains the thing being renewed. AI’s data centre buildout is comfortably inside that shape, not outside it.

One number worth watching while the robotics question resolves: actual humanoid shipments this year are still in the tens of thousands, nowhere near the hundreds of thousands the forecasts above call for within the next year or two. Every ramp on this chart looked exactly this unconvincing early on, including the ones that went on to beat the wartime ceiling. So which is it: is humanoid robotics still early on the fast curve, or quietly sliding toward industrial robotics’ much slower one? I don’t think the forecasters know either. Worth checking this chart again in a year.

Also published on LinkedIn.