I Source Electronics for a Living. Here's What 10 Gigawatts Actually Requires.

Elon Musk told SpaceX employees the company will go from roughly 1.4 gigawatts of AI compute to 10 gigawatts by the end of 2027. I don't read that as a vision statement. I read it as a bill of materials with no lead times on it.

---

On August 11, SpaceX released a thirty-minute all-hands video. In it, Elon Musk said AI revenue would exceed all other SpaceX revenue as soon as September, that AI would represent 99% of the company's value within five years, and that compute capacity would go from about 1.4 GW today, past 2 GW by the end of this year, to roughly 10 GW by the end of 2027. At $30 to $50 per watt per year, he put that at $300 to $500 billion in annual revenue.

Most of the commentary that followed argued about the revenue number. Almost none of it argued about the watts.

That's the wrong way round, and I say that as someone whose entire job is the boring half of this equation. I run a contract manufacturer. People send me designs and I turn them into shipped units. The gap between "we're going to build this" and "we built this" is not a motivational gap. It is a procurement gap, and it has a specific shape that shows up the same way on a 500-unit pilot run as it does on a multi-gigawatt data center program.

So let's do what I'd do if a client walked in with this. Let's treat 10 GW as an RFQ and go down the BOM.

Line 1: Power equipment

Before a single GPU draws current, you need to get medium- and high-voltage power to the hall and step it down. That means large power transformers, medium-voltage switchgear, protective relays, breakers, busway, and distribution.

None of this is a catalog purchase. Large power transformers in particular are built to order by a short list of qualified manufacturers, and the global order book has been congested for years — driven by grid replacement cycles, electrification, renewables interconnection, and, now, everyone else's AI buildout. Lead times that were once measured in months have been measured in years across much of this equipment class.

Here's the part that matters and that capital cannot fix: you are not buying a product, you are buying a slot in someone else's factory. When demand exceeds a supplier's annual output, money stops being the differentiator. Order position does. Relationship does. Willingness to take a non-preferred spec does. I have watched clients with more cash than their competitor lose six months anyway, because the competitor placed the PO in Q1 and they placed it in Q3.

Every serious multi-gigawatt schedule I'd believe starts with the transformer PO date. Not the chip PO date. The transformer PO date.

Line 2: Interconnection and generation

Getting power to the site is a separate problem from distributing it inside. Utility interconnection is a queue you join, not a service you buy. Studies, upgrades, and energization run on the utility's calendar and the regulator's, and in most major markets that calendar is measured in years.

SpaceX's answer so far has been to generate behind the meter. Gas turbines have powered the Colossus sites in Memphis, and that decision has drawn sustained local opposition on air-quality grounds. This is a rational engineering move — it converts a waiting problem into a building problem, which is the trade a fast-moving company should want.

But note what it converts into. Air permits, emissions litigation, and community consent are not schedule risks you can engineer around, and they don't respond to execution speed. They can also get worse with scale, because the political tolerance for one site is not the political tolerance for eight. If the 10 GW plan leans on self-generation, permitting becomes a critical path item, and permitting is the one line on this BOM where being SpaceX helps least.

Line 3: Cooling and mechanical

Rack densities for current-generation accelerators have moved past what air cooling handles economically. That means liquid: coolant distribution units, manifolds, quick-disconnects, piping, dry coolers or chillers, and in many designs a water strategy that has to survive local scrutiny.

Mechanical scope is chronically underestimated in these announcements because it isn't glamorous. But it is real fabricated equipment, sourced from the same congested industrial base as the electrical gear, installed by the same constrained trades, and — critically — it has to be commissioned. A hall with racks in it and an uncommissioned cooling loop is not compute. It is inventory.

Line 4: Silicon

This is the line everyone talks about, and it's not the one I'd worry about most — but the language around it deserves a closer read.

SpaceX has committed its AI stack to NVIDIA's Vera Rubin architecture, and Musk said on the August earnings call that the understanding with NVIDIA is that SpaceX will receive "a very significant percentage" of their GPUs next year. The filing describes a next phase at Colossus 2 bringing online at least 220,000 additional GB300 processors and over 400 additional megawatts.

Read the first of those as a sourcing person would. "Our understanding is that we will receive a very significant percentage" is not a delivery schedule with dates, quantities, and penalties. It is a relationship, characterized favorably, by the buyer. I have written that sentence in status updates myself. It means the supplier likes us and has indicated intent. It does not mean the units are allocated.

And accelerator supply has its own upstream constraints — advanced packaging capacity and high-bandwidth memory chief among them — which sit at suppliers NVIDIA doesn't own either. Then there's the quieter line item: optical transceivers, cables, and the network fabric. Interconnect has been a real bottleneck for large clusters, and it scales with the square of ambition.

Line 5: The people

The one nobody puts on the BOM.

Data center buildouts at this scale need high-voltage electricians, controls technicians, pipefitters, and commissioning agents — and they need them concentrated at specific sites for specific windows. Skilled trades don't scale on a quarterly cadence, and when several hyperscale programs run concurrently in overlapping regions, they bid against each other for the same crews. Labor availability has quietly become a gating factor on more programs than component supply has.

The principle underneath all of it

Here is the thing I'd want any reader to take from this, whether or not you care about SpaceX:

A schedule is set by its longest-lead item, not by the average of its items, and not by the enthusiasm of the person presenting it.

You can parallelize the short poles all you like. If the transformer arrives in month 20, the site energizes in month 20, no matter how fast the concrete went in or how many GPUs are sitting in crates. This is why "we'll build it fast" is a weaker claim than it sounds — speed on the controllable 80% of a project buys you very little when the uncontrollable 20% sets the date.

And critically: scaling is not repetition. The constraints that bind at 300 MW are not the constraints that bind at 10 GW. At small scale you're competing for your own attention. At large scale you're competing for the world's transformer output. The failure mode of ambitious hardware programs is almost never that the team was slow. It's that they modeled the second build as a copy of the first, and the second build ran into a constraint the first was too small to hit.

The fair counterargument

I want to state the other side properly, because it's strong.

This company built Colossus 1 in roughly 122 days. It then built Colossus 2 in a repurposed Electrolux factory on a comparable timeline. Those are extraordinary numbers by any industrial standard, and dismissing SpaceX's execution has been a losing trade for a decade. Serial builders get faster: they standardize designs, pre-order long-lead items against a program rather than a project, and develop supplier relationships that jump the queue legitimately.

They also have capital to deploy against the problem — AI capital expenditure ran $15.83 billion in the second quarter alone — and they have a demonstrated willingness to solve power by building their own, which most operators won't do.

So the honest position is not "10 GW is impossible." It's that 10 GW by end-2027 requires doing the 122-day miracle roughly seven or eight more times, concurrently, against the most congested industrial supply base in a generation, while every hyperscaler on Earth bids for the same equipment and the same crews. Possible. Not the base case I'd put in a plan I had to defend.

What I'd actually watch

Revenue projections are downstream. If you want early signal on whether the buildout is tracking, watch the physical layer:

• Long-lead equipment orders. Transformer and switchgear commitments show up before anything else does. They are the earliest honest indicator of intent converted into schedule.
• Interconnection and permitting filings. Substation energization notices and air permit applications are public, boring, and predictive.
• Supplier commentary. What NVIDIA and the power-equipment makers say about allocation tells you more than what buyers say about their expectations.
• Hiring. A surge in commissioning-engineer and high-voltage electrician postings across specific metros is a build actually starting.
• Disclosed energized capacity, if it ever appears. Nameplate is installed rating. Energized and utilized capacity is the number that bills. If a company starts reporting both, it's confident in the gap between them.

Why this generalizes

I've spent this whole piece on someone else's gigawatts, but the lesson is the one I give clients on a 2,000-unit production run.

Every hardware roadmap dies at its longest-lead component, and the longest-lead component is almost never the one the founder is excited about. It's the connector with a 26-week lead. It's the custom magnetics. It's the one MCU that went on allocation in the same month you froze the design.

The discipline that saves a program is unglamorous: identify the long poles before you commit to a date, place those POs first, and treat every schedule that hasn't done this as a wish. That's true at 2,000 units and it's true at 10 gigawatts. The physics of procurement doesn't care about the scale of the ambition.

---

Figures cited are from SpaceX's Q2 2026 results and IPO filing, the August 11, 2026 all-hands video, and reporting by CNBC, Reuters, and Data Center Dynamics. Capacity and revenue targets attributed to Elon Musk are his stated projections, not filed results. I have no position in SPCX.