The Biggest IPO in History Is Also an AI-Compute Disclosure. SpaceX's S-1 Surfaced the Anthropic-Colossus Lease.

The SpaceX roadshow opens today. The deal prices June 11, debuts June 12, and trades on Nasdaq under SPCX. It raises up to $75 billion at a valuation of at least $1.8 trillion, more than twice the largest IPO ever recorded. That makes it the biggest public offering in history, full stop. And here is the part the space-finance coverage keeps burying: the single most consequential disclosure in this filing is not about rockets. It is an AI-compute lease that surfaced inside an S-1 about a rocket company.

I cover capital and infrastructure for TensorFeed, and I read this prospectus as two documents stapled together. One is a satellite-internet business going public at a generational valuation. The other is a window into how AI compute actually gets financed and rented right now, complete with a contradiction between what the filing says and what Elon Musk says out loud. Let me take them in that order, because the second one is the reason this IPO belongs on our beat at all.

The business going public: Starlink is the engine

Strip away the Mars narrative and SpaceX is, financially, a connectivity company. Starlink now serves roughly 8 million customers, and per the IPO reporting the connectivity segment generated about $11.4 billion in revenue, $4.4 billion of operating income, and $7.2 billion of segment adjusted EBITDA in 2025. Starlink does not get carved out; it trades as part of SpaceX post-IPO. So a public-market investor buying SPCX is buying a high-margin recurring-revenue subscription business with a launch monopoly bolted to the front of it.

That cash engine matters for the rest of this piece. A company throwing off billions in segment EBITDA from a subscription product can self-fund things capital-starved AI startups cannot, including the kind of multi-gigawatt compute ambitions Musk keeps gesturing at. The $1.8 trillion number is not priced on satellites alone. A meaningful slice of it is the market paying up for optionality on everything Musk's constellation of companies might do with that cash flow next, and AI is at the top of that list.

The disclosure: an S-1 is where the Anthropic-Colossus deal became public

Here is the move that turned a finance event into a compute story. The terms of Anthropic's compute lease were not volunteered by anyone. They did not come out in an Anthropic announcement or a SpaceX press release. They surfaced inside SpaceX's IPO filing. The number is large enough to be its own headline: Anthropic pays $1.25 billion per month for the full output of Colossus 1, the roughly 300 MW datacenter near Memphis with more than 220,000 NVIDIA GPUs across a mix of H100, H200, and GB200 silicon.

One clarification, because the corporate structure is the whole reason this sits in the filing. Colossus was built by xAI, and xAI is now part of SpaceX. SpaceX bought xAI in an all-stock deal that closed in February 2026, the largest merger ever struck, and folded it into a division it calls SpaceXAI in May 2026. So xAI is a wholly owned SpaceX subsidiary, and Colossus 1 is, by the org chart, a SpaceX asset. That is precisely why the lease is in the prospectus: a company going public has to consolidate and disclose its subsidiaries' material contracts, and a $1.25 billion-a-month revenue line clears that bar without trying. This is not a competitor's lease that leaked through SpaceX's paperwork by accident. It is SpaceX disclosing that it rents a block of idle GPUs to Anthropic, an AI rival, for nine figures a month, because the S-1 left it no choice.

We already wrote the deal itself up when the partnership first broke, including what 220,000 accelerators actually buy a compute-constrained lab and the orbital ambition sitting behind it. If you want the booking mechanics, that piece is here: Anthropic Just Booked 220K GPUs on Colossus 1. This article is about the new thing: what it means that the terms came out in the largest IPO ever, and the fact that the two parties cannot agree on what those terms are.

Why the GPUs were idle in the first place

The detail that makes the economics click is utilization. Colossus 1 had been running at roughly 11 percent because xAI moved its own training onto Colossus 2. So the original cluster sat mostly dark, a depreciating 300 MW asset earning almost nothing. Anthropic took the idle capacity, the full output of Colossus 1, while xAI keeps Colossus 2 for its own runs.

That is a cleaner trade than the headline number makes it sound. xAI converts a stranded asset into $1.25 billion a month of contracted revenue. Anthropic, which has been visibly short on compute for its consumer tiers, gets a large block of frontier silicon outside the Google and AWS stack it leans on. Two rivals, one transaction, both rational. The interesting question is not whether the trade makes sense. It is how long it lasts, and that is exactly where the story falls apart.

The dispute: SpaceX says 2029, Musk says 180 days

Read SpaceX's S-1 and you get one picture: a monthly fee running through May 2029. Read that as a three-year arrangement and the contract is worth north of $40 billion to SpaceX over its life. That is a colossal multi-year commitment, the kind of number that reshapes how you think about a lab's cost structure and a datacenter operator's revenue base.

Then Musk opened his mouth. He has publicly described it as a 180-day lease with 90-day mutual cancellation, and added that the short term was "our request, not Anthropic's." Put those two framings side by side and they do not reconcile. One is a roughly $40 billion three-year revenue line in an SEC filing. The other is a six-month rental either side can walk out of on a quarter's notice. Musk's own public statement contradicts his own company's filing, in the middle of the biggest IPO ever priced.

I am not going to pretend to resolve which version is operative, because the public record does not let me, and I am not going to invent a reconciliation. Both framings can be technically true at once: a contract can name an outside date of May 2029 while granting a 90-day cancellation right that makes the practical commitment six months at a time. But you cannot have it both ways when you are pricing a security. A reader of the S-1 sees a $40 billion-plus counterparty relationship. A listener to Musk hears a cancelable short-term rental. Those are very different inputs into a $1.8 trillion valuation, and the gap between them is sitting in a live prospectus.

What a $1.8 trillion AI-adjacent infra IPO actually signals

Step back from the dispute and the structural read is the part that matters for anyone tracking where AI capital goes. Three signals stack up.

First, the cash-engine pattern. The most durable way to fund frontier compute right now is not a venture round; it is a high-margin subscription business throwing off billions, used to bankroll the buildout. Starlink connectivity at $7.2 billion of segment adjusted EBITDA is that engine for Musk's orbit. It is the same shape we keep seeing across the compute map: the companies that can actually finance gigawatts are the ones with a profit machine already running underneath.

Second, idle-GPU monetization is now a real line of business. A 300 MW cluster running at 11 percent is a balance-sheet problem, and the Anthropic lease is the answer: when your own training moves to the next cluster, you rent the old one to a competitor by the month. Expect more of this. Frontier silicon is too expensive to leave dark, and the labs are short enough on capacity to pay a rival's asking price. We track the buildout and the operators behind it on our AI infrastructure tracker, and idle-capacity leasing is becoming one of the load-bearing dynamics there.

Third, the disclosure path. The biggest single data point about AI compute economics this quarter did not come from a lab being transparent. It came out because SpaceX had to file an S-1, and a company going public cannot bury a subsidiary's billion-dollar-a-month contract. That is worth sitting with. As more of this industry approaches the public markets, the disclosure regime is going to keep surfacing terms the labs would never volunteer, $1.25 billion a month being the latest. The IPO calendar is becoming an involuntary transparency engine for the compute layer, and this is the clearest example yet.

My take

The space press will cover SPCX as a valuation milestone, and on those terms it is one: the largest IPO in history, more than twice the prior record, on the back of a satellite-internet business with real margins. Fair enough. But the most important sentence in this filing is the one about a competitor's GPUs. It tells you that AI compute is now financed through subscription cash flow, monetized even when idle, and disclosed mostly by accident, and that the people signing the biggest checks do not always agree on what they signed.

Watch three things after the June 11 pricing. One: whether the prospectus risk factors get amended to square the May 2029 framing with the 180-day characterization, because a live contradiction between an SEC filing and a founder's public statements is the kind of thing that does not survive a roadshow quietly. Two: whether Anthropic confirms either version, since it is the counterparty and has said almost nothing. Three: whether the idle-Colossus-1 trade becomes a template, with other operators renting last-generation clusters to capacity-short labs by the month. If it does, the most interesting number in the SpaceX IPO will not have been the valuation. It will have been $1.25 billion a month, for GPUs that were sitting dark.