SpaceX Wants 10 Gigawatts of AI Data Centers by 2027. Watch the Suppliers Instead.

SpaceX plans to build 10 gigawatts of AI data center capacity by 2027, running on Nvidia's Vera Rubin platform. It's a striking figure for any single company's infrastructure plans, attached to a company still better known for rockets and satellites than server farms. The announcement landed in the same short news window as two quieter stories: Onsemi forecast upbeat revenue on surging AI data-center chip demand, and Siemens Energy posted forecast-beating results it attributed to AI data-center demand.

Line all three up and the obvious reading is "AI infrastructure buildout is accelerating, and here are three data points confirming it." We don't think that reading is wrong, exactly. We do think it flattens a distinction that matters more than the shared subject matter suggests — the SpaceX figure and the two earnings reports are not the same kind of claim, and they don't deserve the same amount of trust.

Part of why stories like these get bundled together is just how AI coverage tends to work right now: anything that fits the broad "AI infrastructure demand is real" template gets grouped with other stories that fit the same template, regardless of what kind of claim each one is actually making. A capacity target, an earnings beat, a permitting filing, an analyst note — different species of information, all filed under the same recurring headline logic. That's not a knock on how any single outlet covered these three stories. It's just worth naming as a pattern, because it's exactly the kind of pattern that makes stress-testing your own first read of a story worthwhile.

Two Different Kinds of AI Infrastructure News

Start with what each one actually is. SpaceX's 10-gigawatt figure is a target: a stated intention about what the company plans to have built by a date more than three years out from the announcement. Nobody has energized 10 gigawatts of anything yet, and nobody is being asked to verify that capacity exists today. It's a plan communicated with the confidence of an announcement — which is a different thing entirely from a plan that has already happened.

Onsemi's and Siemens Energy's numbers are the reverse kind of claim: backward-looking, describing a quarter that has already closed rather than one still to come. When a chipmaker says its revenue outlook is upbeat because of AI data-center chip demand, or an energy equipment maker says its results beat forecasts because of AI data-center demand, that's a company reporting on orders that were placed — not orders it hopes will materialize by some future date. A target and a result tend to get delivered in the same declarative sentence structure, with the same tone of certainty, but they are not equivalent evidence. Treating them as interchangeable proof that "AI infrastructure is happening" undersells how much more work the earnings numbers are doing.

Why Supplier Earnings Are the More Reliable Signal

This is where we'd go further than just noting the distinction, and make a real claim about which one deserves more weight if you're trying to gauge whether AI infrastructure spending is durable rather than a narrative running ahead of itself. Our answer: the supplier earnings, and it isn't close.

The reason is close to mechanical. Onsemi and Siemens Energy don't control the AI industry's narrative and gain nothing from talking up a future that hasn't arrived — their numbers are a function of orders customers have actually placed and, typically, already paid for. A company can publish an ambitious multi-year target more or less whenever it wants to. A company cannot simply decide its quarterly revenue will be larger. Suppliers get paid when customers show up with real money for real components, so an earnings beat attributed to AI data-center demand is downstream confirmation that somebody, somewhere, is buying right now — not announcing an intent to buy eventually.

Put differently: SpaceX's announcement tells us what SpaceX wants us to believe about 2027. Onsemi's and Siemens Energy's results tell us what their customers — plural, diversified, under no obligation to make any single company's roadmap look good — did in the most recently closed quarter. If we could only trust one category of story about whether AI infrastructure spending is real, the one where money already changed hands beats the one where a number got printed in a press release.

The Base Rate for Multi-Year Infrastructure Targets

None of this requires believing SpaceX is being misleading, and we want to be careful not to imply that. Multi-year capacity targets are a normal, legitimate thing for infrastructure-heavy companies to publish. Investors and partners want some sense of scale and timeline, and "we won't share our plans because plans sometimes change" isn't a realistic communications posture for a company operating at this size.

But the honesty of a target and its reliability as a predictor of 2027 reality are two separate questions, and the base rate on the second one isn't generous — not because of anything specific to SpaceX or to AI, but because that's how large infrastructure projects behave generally. Power plants, chip fabs, transmission buildouts, big data-center campuses: projects that require physical construction, permitting, equipment lead times, and years of capital markets cooperating routinely slip their timelines, shrink from the number in the original announcement, or get restructured into a different shape before the target date arrives. That's true across infrastructure megaprojects broadly, well before anything AI-specific enters the picture. A 10-gigawatt figure stated for 2027 is a plan sitting at the very start of the multi-year window where that kind of slippage typically shows up. That doesn't make it fake. It does mean a reader should hold it more loosely than a same-quarter earnings result.

This is familiar enough that it's close to conventional wisdom among people who cover infrastructure for a living, even outside of AI: energy projects, transportation megaprojects, large manufacturing buildouts — the number attached to the groundbreaking announcement and the number attached to the ribbon-cutting are, more often than announcements would suggest, two different numbers. None of that requires any special skepticism about SpaceX, Nvidia, or the AI industry specifically. It requires only the ordinary skepticism anyone should bring to a multi-year capital commitment stated in a press release, in any industry, at any point in the last several decades of infrastructure reporting.

Steelmanning the 10-Gigawatt Number

It's worth taking the other side seriously too, because we don't want this to read as "big target bad, boring earnings report good." A 10-gigawatt figure paired with Nvidia's Vera Rubin platform is genuinely useful information, even if it ends up realized on a different timeline or at a different scale than announced. Targets signal where a company is directing capital and attention. They give suppliers and partners something to plan against. And they're often directionally right even when the specific number moves — a company that eventually builds a smaller number, or hits 10 gigawatts later than 2027, hasn't necessarily been dishonest. It's run into the ordinary friction that multi-year builds run into.

So the fair framing isn't "distrust the target." It's "hold the target and the earnings beat to different evidentiary standards by default, because they're answering different questions." One answers "what does this company intend to do." The other answers "what already happened." Reading AI infrastructure claims well is mostly a matter of noticing which question a given headline is actually answering, instead of letting shared subject matter and a shared confident tone blur two different kinds of claims into one.

There's also a coordination function a big public target serves, separate from whether it's hit exactly on schedule. Suppliers, partners, and capital markets calibrate their own planning off of publicly stated demand signals — a number like 10 gigawatts by 2027 gives Nvidia, its component suppliers, and prospective energy partners something concrete to build capacity against, even before a single facility is finished. In that sense the target is doing real work in the present, just not the kind of work that functions as proof the number will be hit.

What Would Actually Confirm the Plan Is On Track

So what would move us from "SpaceX has stated an ambitious target" to "SpaceX's buildout is verifiably on track"? Not another restatement of the 10-gigawatt figure — a number repeated in a follow-up story isn't new evidence, it's the same claim again. What we'd want to see is more of what this week already handed us in miniature: suppliers to that specific buildout — chipmakers, energy equipment providers, the companies that get paid when construction actually happens — reporting increased orders they attribute to it, in their own results, independent of anything SpaceX itself says.

That's a meaningfully higher bar than it sounds, because it requires corroboration from a party with no incentive to inflate SpaceX's numbers. We don't have that yet, specifically tied to this buildout, and its absence isn't evidence against the plan — it's simply too early in the timeline for that kind of confirmation to exist. What we do have, this week, is a useful habit worth adopting: keep two separate columns for AI infrastructure news. One for stated targets, which tell us about intent. One for realized supplier results, which tell us about money that already moved. The SpaceX announcement belongs in the first column. Onsemi and Siemens Energy belong in the second. We'd weight that second column more heavily, right up until the two columns start describing the same thing.

We'd rather leave it there, honestly, than force a tidier ending. The 10-gigawatt target might turn out to be conservative, or it might turn out to be the kind of number a company quietly revises well before 2027 arrives. Either way, the more useful habit for readers isn't guessing which outcome is more likely. It's noticing, every time a plan and a result show up in the same news cycle, which one is actually which.