SpaceX is adding 200,000 GB300 chips a month: Q4 spending jumps to $35-40 billion

SpaceX is bringing "roughly 200,000 GB300s online per month," Elon Musk said in his prepared remarks on the company's third-quarter call on October 20. That is the wording in both the Motley Fool and Investing.com transcripts, and we treat the prepared remarks as the source of record. The question-and-answer session is where the transcripts disagree: Investing.com has Musk saying "220,000 GB300s a month," while the Fool transcript has "20,000 GB300s a month." The Fool page warns it was partly generated with software and may contain errors, so we have not relied on either Q&A figure. SpaceX's recording and filings are the check.

We covered the quarter itself in our earnings story. This piece is about the build.

The spending numbers

The guidance implies fourth-quarter spending of 1.8 to 2.0 times the third quarter, by our arithmetic ($35 billion and $40 billion divided by $19.6 billion). Going from 1.7 gigawatts to as much as 3 gigawatts in a quarter would add about 1.3 gigawatts of capacity. Musk also gave 2027 targets, but the two transcripts record them differently ("incremental 10 gigawatts" in one, "less than 13 gigawatts" by the end of next year in the other), so we do not repeat either.

How big is that against the US buildout?

BloombergNEF's July forecast, published July 21, puts installed US data center capacity at 118 gigawatts in 2030 and 194 gigawatts in 2035. Three gigawatts is about 2.5% of the 2030 figure. That sounds small until you remember it is one company, and that most of it is arriving in roughly a year. The comparison is rough: SpaceX quotes nameplate compute capacity and BloombergNEF counts installed data center capacity, which are not identical measures. We have not found a clean national count of what is under construction today; published estimates vary widely because sources define it differently.

Half of it is rented out

In the Q&A, answering an analyst, Musk said that "half of the compute is used internally and half of it is rented to others." Per the Fool transcript, he described the rental as "effectively the spot market rate" and said short-term contracts are the smart move. Investing.com's transcript has him saying SpaceX could rent out its compute "several times over" if it said yes to everyone. The CFO said the company will "continuously evaluate the highest margin usage" for its compute and expects more new customers.

What it means for neoclouds and GPU renters

CoreWeave, Nebius and similar companies make money by renting GPU capacity, mostly on long contracts, largely financed with debt. A competitor adding hundreds of thousands of the same chips a month and offering them on spot terms is a pricing risk to that model. We cannot say it will lower rental prices, and Musk also said demand exceeds what SpaceX can supply. Both can be true.

The pressure shows up in three places we would watch:

• Spot rates. If SpaceX capacity sets the market price for short-term GPU time, contracts signed at higher prices look more expensive at renewal.
• Financing. Neoclouds borrow against contracted revenue. Our recent look at CoreWeave's bond yields and its credit default swaps shows lenders already charge a heavy premium.
• Chip supply. Spending at this pace competes for the same Nvidia output. SpaceX's own GPU debt plan is the other side of that trade.

The counterpoint is that a hungrier buyer can also tighten supply for everyone, and a landlord who says it turns customers away is describing a market with more demand than capacity. Which effect wins is not settled by one call.

Sources: Motley Fool and Investing.com call transcripts, Oct 20 2026 (transcripts differ in the Q&A and may contain errors); BloombergNEF; SEC EDGAR filings. This is market information, not investment advice.