Anatomy of a $45 billion deal: how SpaceX's Colossus rental to Anthropic turns the S-1 into a hybrid Starlink-plus-neocloud IPO
On 5 June 2026, the hosts of the allin podcast walked through a draft registration statement that, if it prices where the filing contemplates, would not merely be the largest IPO in market history. It would be a reclassification of what an aerospace listing is supposed to be. Jason Calacanis read from the SpaceX S-1 a target valuation of $1.75 trillion and a target raise of $75 billion — more than two and a half times Saudi Aramco's 2019 debut of $29 billion. A Polymarket contract cited on the same episode put the implied probability of a first-day close above $2 trillion at 71%. The headlines wrote themselves: SpaceX was going public.
The single most consequential line in the filing, though, is not the valuation. It is the disclosure that Anthropic is paying SpaceX $1.25 billion per month to rent Colossus 1 and parts of Colossus 2 — a three-year, $45 billion contract cancellable by either side on 90 days' notice. That number, paired with a separate SpaceX AI compute segment of $3.2 billion in revenue growing at roughly 2x year-over-year against $6.4 billion in operating losses, restructures the deal. This is not an aerospace IPO with a side hustle in machine learning. It is a Starlink-plus-neocloud hybrid masquerading as one.
The four-segment stack
The S-1, as read aloud on the 5 June episode, breaks SpaceX into four operating segments. Starlink delivered $11.4 billion in 2024 revenue at 50% growth and $4.4 billion in operating income, on a base of more than 10 million subscribers. The launch business posted $4 billion at 17% growth against $650 million in operating losses. The AI compute segment ran $3.2 billion at 2x year-over-year growth with $6.4 billion in operating losses. A fourth constellation of adjacent revenue — Cursor, the AI coding tool that Chamath Palihapitiya said SpaceX is in the process of buying, contributing an estimated $2-3 billion in additional top line growing at roughly 2x year-over-year — is not yet in the filing.
Capex in 2024 ran $20 billion, with 60% directed at AI compute. The data-center build velocity disclosed in the S-1 is the most striking operational metric: Colossus 1 took 122 days to deploy, Colossus 2 took 91, and Colossus 3 is targeted at 66. That is a near-halving of build time across three consecutive sites, and it explains the operating losses in the segment. Colossus is not yet a margin business; it is a capacity business being constructed on the assumption that a tenant pipeline exists to absorb it. The Anthropic contract, in that reading, is the underwriting.
The Anthropic anchor and what it prices
The $1.25 billion-per-month, $15 billion-per-year commitment is the largest single compute rental on public record. Gavin Baker, the founder of Atreides Management who appeared as a guest on the episode, framed the broader market context this way: OpenAI and Anthropic are at roughly $100 billion of ARR combined, with inference gross margins near 80%, and the end-of-year ARR across the major LLM players could reach $200-400 billion. Against that backdrop, a $15 billion annual commitment from one of the two anchor tenants is large but not absurd — it is closer to a wholesale-of-capacity deal than a strategic anomaly.
The arrangement also points to a feature of frontier AI economics that rarely makes it into press coverage. Baker noted on the episode that Cursor's Composer 2.5 model achieved Pareto dominance in three to four weeks of reinforcement learning on Colossus 2, using the same base model as the prior Composer 2 release, and that Cursor allegedly holds more tokens of coding-specific data than exist on the public internet. If the claim holds, it suggests that the binding constraint on next-generation model performance is no longer pre-training data volume but proprietary reinforcement-learning corpora — and that the value of the underlying compute is partly a function of which tenant is using it. SpaceX is not selling commodity megawatt-hours. It is selling time on a platform whose productivity varies by customer.
The Nvidia anchor and the death of the Burry thesis
Nvidia's most recent quarter, also discussed on 5 June, complicates the neocloud economics in SpaceX's favour. The company reported $81.6 billion in revenue, up 85% year-over-year and 20% quarter-over-quarter, on $58 billion of net income, $48 billion of free cash flow, 75% gross margins, and a $5.3 trillion market capitalisation, with the stock up 16% year-to-date. The capital return programme added $80 billion in buybacks and lifted the quarterly dividend 25x, from 1 cent to 25 cents, with 50% of free cash flow committed to shareholders. A separate projection baked into the quarter — a $20 billion annual CPU business that would make Nvidia one of the largest CPU manufacturers overnight — points to a co-design pivot rather than a one-product company.
The most important read-through for SpaceX's S-1 is the second-order effect on the neocloud business model. Baker argued on the episode that the Michael Burry "short useful life of GPUs" thesis is wrong: disaggregated inference, in which a Groq or Cerebras decoder sits in front of older Nvidia GPUs, extends GPU useful life to 10-15 years. That longer depreciation schedule is what permits asset-backed financing in the 6% range at names like CoreWeave, and it is what permits SpaceX to put a 36-month, cancellable-on-90-days contract with Anthropic at the centre of a $20 billion annual capex programme without terrifying the credit committee. The Colossus economics depend on the assumption that the silicon inside does not become obsolete on a two-year cycle. Nvidia's quarter, read carefully, supports that assumption.
There is a counter-narrative worth taking seriously. The same episode highlighted that Nvidia's growth within the western hyperscaler and AI cloud segment is now outpacing Broadcom's, undercutting the once-fashionable thesis that custom ASICs were going to eat the GPU market. If true, the moat around Nvidia extends, but so does the concentration risk for anyone whose compute business is built on Hopper-and-Blackwell silicon purchased at peak prices. The Colossus anchor pricing assumes a world in which the underlying hardware retains residual value; the Burry thesis assumes one in which it does not. Public-market investors in SpaceX are being asked to underwrite the first assumption and ignore the second.
Macro overlay: rates, energy, and a pulled executive order
Two macro factors frame the S-1's reception. On rates, Calacanis read yields on the same episode: the US 10-year at 4.6%, the Japanese 30-year at a record 5.1%, UK gilts at their highest since the global financial crisis, and German bunds at their highest since 2011. May CPI, per Polymarket cited in the broadcast, was trading at a 99% implied probability of printing above 4.2%, with a Survey of Professional Forecasters projection of 6% — a major upward revision from the 3% handle that anchored rate-cut expectations at the start of the year. Friedberg put aggregate global debt-to-GDP at 310%. In that backdrop, a $75 billion equity raise is not just a financing event; it is a withdrawal of future financing pressure from a credit market that is already straining.
The second overlay is geopolitical. A Trump executive order regulating frontier AI models, scheduled for announcement and attended on paper by every major neocloud and hyperscaler CEO, was pulled the morning of release, with the sticking point reportedly federal government review of frontier model deployments. The episode framed the move as a lobbying win for the labs; the more interesting structural read is that the administration chose, in effect, to let the industry self-regulate rather than impose a supervisory framework during an election-sensitive window. Palihapitiya, on the same broadcast, floated a "KYC for AI" alternative in which frontier deployments are gated by identity rather than by pre-release approval. Whether the EO resurfaces in modified form is the open variable. The fact of its near-existence is the signal: the regulatory perimeter around frontier AI in the United States is, for the moment, permeable.
The hybrid read
The cleanest frame for the S-1 is that SpaceX is selling investors a four-engine machine. Starlink is the cash-generative anchor, with the operating-income profile of a mature telco. Launch is the strategic moat, with the operating losses of a business optimised for share rather than margin. The AI compute segment is the growth engine, with the operating losses of a business in mid-build. And the forthcoming Cursor acquisition, if it closes, layers a software-vertical revenue stream on top of the hardware stack. Each segment justifies a different multiple. The S-1's task is to make the sum feel coherent.
The Anthropic contract is the hinge. Without it, the AI compute segment is a $3.2 billion revenue line subsidising $6.4 billion of operating losses on $20 billion of annual capex, with the implicit promise that future tenants will arrive. With it, the segment has an offtake agreement from a counterparty whose own ARR trajectory, per Baker's read, supports a multi-year, multi-billion-dollar compute spend. The 90-day cancellation clause cuts both ways: it means SpaceX cannot lock the revenue in, but it also means Anthropic cannot lock SpaceX into a price that does not reflect falling inference costs. The structure is a bet that compute, like bandwidth, will be priced as a near-commodity within the contract window.
Investors who treat this as an aerospace IPO will misread the risk. Investors who treat it as a neocloud IPO will misread the moat. The honest read is that SpaceX has assembled the only public-market vehicle that offers simultaneous exposure to orbital launch, satellite broadband consumer revenue, frontier AI compute capacity, and — if the Cursor deal closes — a vertically integrated AI coding product. The S-1 is not pricing those four businesses as a sum-of-parts. It is pricing them as a single bet on the assumption that the demand curve for inference will continue to outrun the supply curve for at least the next 36 months. Polymarket's 71% probability of a first-day close above $2 trillion is, in that sense, a wager on that exact assumption — and a useful proxy for how the public market will read the S-1 when it prices.
Wire provenance
This editorial synthesis draws on the following public wire/social posts:
- https://www.youtube.com/watch?v=HGbA6ze0_3M