enesptjazh-Hanszh-Hantfrdeitkoruarhiidtrnlplvithsvhemsukcselrohudafinobn

Musk says power, not chips, caps AI as Amazon parks $5 billion of Nvidia racks

SpaceX plans 13 gigawatts of AI compute by end of 2027. The same day, Amazon was reported holding $5 billion of Nvidia racks with no power.

By Published October 21, 2026 7 min read Grounded in the CEOInterviews.AI corpus

On SpaceX's first earnings call as a public company, on October 20, Elon Musk said the thing holding back AI is no longer the chip. It is the turbine, the transformer and the cooling loop. He then promised 13 gigawatts of SpaceX compute by the end of 2027. The same afternoon, a Business Insider report said Amazon had started taking ownership of about $5 billion of Nvidia GB300 racks that some AWS sites cannot power yet. Those two facts point the same way. Google's Sundar Pichai and OpenAI's Sarah Friar spent the same day saying they are short of compute. Ken Griffin used his own stage to explain what happens to GPU prices once the power finally arrives.

Stream of the SpaceX third quarter 2026 earnings call
The SpaceX Q3 2026 earnings call stream, October 20. YouTube thumbnail.

Musk moved the bottleneck

The numbers came first. SpaceX reported $14.6 billion of revenue, up 161% from a year earlier, and $4.4 billion of net income. The AI segment, which Musk calls SI, brought in about $7.5 billion of that, according to TechCrunch, and capex was $19.6 billion in the quarter. Then Musk changed the subject from GPUs to everything around them.

So as I think people are starting to realize, the constraint on deployment of super intelligence is more in power and cooling and infrastructure than it is in chips.

Elon Musk, October 20, 2026, SpaceX Q3 2026 earnings call

His supporting claim is a run rate, which is more useful than a target. He said SpaceX is turning on roughly 200,000 GB300s a month and expects to end 2026 with 2.5 to 3 gigawatts online.

We're tracking to be able to bring roughly 200,000 GB300s online per month for every month for the remainder of this year, and continuing that pace next year, with next year being primarily Vera Rubin compute.

Elon Musk, October 20, 2026, SpaceX Q3 2026 earnings call

He called 10 more gigawatts in 2027 a high-confidence number and 20 a stretch. He also floated $500 billion to $1 trillion of annual recurring revenue by the end of 2027. I would treat the gigawatts as a plan and the revenue range as Musk being Musk. He joked right after that he did not want to give his CFO a heart attack.

The turbine blade is the scarce part

The most concrete thing Musk said was about castings, not chips. SpaceX built a casting shop for rocket engines, and it is now pouring blades and vanes for gas turbines. Musk said turbine makers can ship everything else long before they can ship the blades.

And when we talk to the turbine suppliers, they tell us that the blades and vanes are 12 to 18 months behind the rest of the turbine.

Elon Musk, October 20, 2026, SpaceX Q3 2026 earnings call

If that is true, it explains the Amazon story better than any theory about demand. A GB300 rack is useless without a powered hall to sit in, and power plants wait on parts that take a year or more. Musk did not say how much of his 2027 power comes from his own turbines. He said the plan was "highly proprietary," which is also a convenient answer.

Amazon has the racks and lacks the power

The Business Insider report, based on internal August emails, describes an "Early Rack Receipt" program. Amazon takes title to finished Nvidia racks and stages them in warehouses until a site has utility power, and the emails projected about $3 billion more hardware a month. Amazon called the staging a routine "best business practice."

Ed Zitron read it differently on Prof G Markets on October 21. He has estimated that roughly half of America's AI chips are sitting in warehouses, and he argued Nvidia is pushing hardware onto customers early. That is his estimate, not a count, and the report does not prove it. What the report does show is the same constraint Musk described, from the buyer's side. The chips shipped and the power did not show up on time.

Larry Fink took that to the macro level in a Bloomberg interview the next day. He cited hyperscaler talk of $7 to $9 trillion of buildout and argued that slowing data centers would be the inflationary choice, because compute is expensive while it stays scarce.

Pichai says the shortage was set in 2024

Sundar Pichai was asked on the Big Technology Podcast, also published October 20, whether compute is sold out until 2028. He did not argue. He said a customer asking Google for compute in 2027 faces something "incredibly difficult to do," and that Google leased short-term capacity last year to help some customers. His explanation for the shortage is lead time.

So the compute we have today is a function of decisions I made in 2024.

Sundar Pichai, October 20, 2026, source
Sundar Pichai, chief executive of Alphabet
Sundar Pichai at the European Commission, 2023. Photo by Lukasz Kobus, European Commission, CC BY 4.0, via Wikimedia Commons.

On the bubble question he gave the bull answer and then a real caveat. Today's capacity is being used, he said, but an overbuild is possible because a data center needs thousands of parts to arrive together.

Yeah, like there are like 2,000 complements or more, right? And so you could get constrained at some point.

Sundar Pichai, October 20, 2026, source

That is the Amazon warehouse problem described by the person who signs Google's capex. One late part strands the rest.

Sarah Friar made the same point about OpenAI in a single line in Dublin. Asked whether its Dots product is coming to Europe soon, she said yes, and that it would have come "even sooner" if OpenAI had built more compute. That is a CFO admitting the company is supply bound.

Griffin's airline seat

Ken Griffin spoke at the Citadel Securities conference, in an interview Bloomberg posted on October 20. He was the only one of these speakers who described the downside in plain mechanics.

GPUs, like airlines, have a very low marginal cost to run.

Ken Griffin, October 20, 2026, source

An empty seat sold for twenty dollars is still profit, he said, which is why airline pricing collapses whenever there is spare capacity. He compared it to the fiber glut of 2000 and then said what that would mean for compute.

Well, that's the question mark, which is if we reach a point where the provisioning of GPUs exceeds demand for GPUs, marginal pricing will actually go into freefall. And that really upends all the investment assumptions that people are working with today.

Ken Griffin, October 20, 2026, source

He also questioned how long old chips keep earning. People are getting "extraordinary pricing on GPUs that are five or six years old," he said, and that probably does not last another five or six years.

SpaceX is betting on the spot price

This is where Musk and Griffin actually disagree. Musk said SpaceX will rent about half its compute to outsiders, and it will do so on short contracts at roughly the spot rate, not long fixed deals. He said renewals are coming in at higher dollars per watt.

Our belief is that the value of compute per watt will increase over time, not decrease.

Elon Musk, October 20, 2026, SpaceX Q3 2026 earnings call

His evidence is the same rental market we covered two days ago, when H100 rents rose 7.6% in a week. Musk even said old H100s now rent for more than they did new. Griffin looked at the same fact and called it temporary.

Short contracts are the right structure if the shortage lasts. If it ends, they are the worst structure to hold, because spot prices are the first to fall when capacity piles up. Musk said a six-month payback on compute next year is not impossible. That number only holds if the cheese shop stays empty, which is Musk's own Monty Python line about hyperscalers having no GPUs to rent.

Does the shortage outlast the power gap?

For 2027, I think it does. Pichai says customers cannot get 2027 capacity, Friar says OpenAI is supply bound, and Amazon is holding $5 billion of racks with no power to run them. None of that describes a glut. But the warehouse is also the risk Griffin is describing. Those racks are already bought. When the turbines, the transformers and the blades Musk says are 18 months late finally arrive, a lot of capacity switches on at once, and it lands on a market that SpaceX chose to price at spot. The figure to watch is not GPU shipments. It is how many dollars per watt SpaceX's renewals get in the quarter after Amazon's racks come out of storage.