The editorial argues that a top-tier frontier lab publicly committing $1.25B to train on infrastructure operated by a direct competitor's parent company is unprecedented. It frames the deal as proof that Anthropic couldn't secure enough megawatts fast enough from AWS or GCP, forcing it to break the standard hyperscaler-lab pairing and take capacity wherever it could find it.
The editorial contends the previous map of frontier compute — Microsoft/OpenAI, Google/DeepMind, AWS+GCP/Anthropic — has been punctured by this contract. Anthropic paying Musk's holding company to train a model that competes directly with Grok signals that competitive boundaries are collapsing under capacity pressure.
Citing the S-1's $18.7B revenue, $2.6B operating loss, and $4.9B net loss against $6.6B adjusted EBITDA, the editorial argues Starlink subscriber ARPU is the cash engine funding Starship, Raptor, and the COLOSSUS II buildout. The rumored >$1T IPO valuation is implicitly underwritten by that single profitable line of business.
By submitting the raw SEC filing rather than a secondary news article, this submitter signaled that the primary document itself deserves direct reader attention. The 386 points and 297 comments validate the view that the filing contains material disclosures — including the Anthropic line item — worth surfacing without editorial mediation.
SpaceX filed its S-1 ahead of a public offering, and buried in the customer-concentration disclosures is a sentence that reframes the entire AI compute market. "In May 2026, we entered into Cloud Services Agreements with Anthropic PBC ... with respect to access to compute capacity across COLOSSUS and COLOSSUS II," the filing reads, with the customer committed to pay at least $1.25 billion.
COLOSSUS is the Memphis-based GPU supercluster originally built by xAI — the same xAI that ships Grok and competes head-on with Anthropic's Claude. Following the xAI/X/SpaceX consolidation, those clusters now sit on SpaceX's balance sheet, which means Anthropic is, in effect, paying Elon Musk's holding company to train the model that competes with Musk's own model.
The rest of the financials are equally instructive. 2025 revenue: $18.7B, up from $14.0B in 2024. Operating loss: $2.6B. Net loss: $4.9B. Adjusted EBITDA: $6.6B. Starlink is doing the heavy lifting on cash generation; everything else — Starship, Raptor, COLOSSUS II — is being funded out of subscriber ARPU and a rumored >$1T IPO valuation.
The Anthropic line item matters more than the headline IPO number. For the first time, a top-tier frontier lab has publicly committed to training on infrastructure operated by a direct competitor's parent company. That's not a rounding error in a procurement spreadsheet — it's a structural admission about how scarce H100/H200/B200-class capacity has become.
Until now, the standard frontier-lab compute story has been: Microsoft buys Nvidia, OpenAI rents from Microsoft; Google builds TPUs, DeepMind uses them; Anthropic splits between AWS Trainium and GCP. The COLOSSUS deal punches a new door into that wall. Anthropic clearly couldn't get the megawatts it wanted from AWS or GCP fast enough, and was willing to take capacity from a Musk-controlled entity to close the gap. The contract is large enough ($1.25B is the floor) to suggest sustained training runs, not burst inference.
For practitioners, the more interesting subtext is what this says about COLOSSUS itself. xAI's cluster has been mocked as a marketing exercise — "the world's largest GPU cluster" claims that didn't always survive scrutiny. An Anthropic contract is the strongest possible third-party validation: Anthropic's infra team is famously conservative about reliability, networking topology, and storage throughput, and they don't sign nine-figure deals for hardware that doesn't deliver.
Community reaction on Hacker News skewed skeptical of the valuation but bullish on the strategic logic. "Crazy this company will IPO for >1B with such bad financials!" one commenter noted before pivoting: "Starlink seems to be a real cash machine, not as good as ads but enough to support AI bets." Another flagged the structural awkwardness: "I was skeptical when xAI and Twitter were rolled into it. The S-1 here makes it even more disappointing." The complaint is fair — SpaceX is now a conglomerate where rocket margins, satellite ARPU, social-media ad revenue, and GPU rental fees all flow into the same P&L.
If you're an infra lead at anything smaller than a frontier lab, the practical signal is grim: the labs are bidding capacity off each other's roadmaps, and the secondary market for training-grade GPU hours just got tighter. Expect on-demand H200 pricing on AWS/GCP/Azure to stay sticky through 2026, and expect reserved-instance negotiations to require longer commits.
For anyone building on Anthropic's API, this is mildly reassuring — it suggests Claude's training and serving capacity is being aggressively expanded, not constrained. The $1.25B floor implies enough silicon to materially move Anthropic's model-training cadence. If you've been planning around Claude rate limits or worrying about whether the next model bump arrives on time, the COLOSSUS deal is a tailwind.
For builders considering xAI's API for Grok, the read is more ambiguous. Selling capacity to Anthropic is rational revenue, but it's also tacit acknowledgment that xAI itself can't fully absorb COLOSSUS II's output. If your roadmap was leaning on Grok-4-class capability becoming dominant, the resource-allocation story just got muddier.
The IPO will price on the SpaceX/Starlink story — rocket cadence, satellite ARPU, Starship cargo economics. But the more durable question is whether Musk has accidentally built the third hyperscaler. If COLOSSUS II ships on time and Anthropic renews, SpaceX becomes a structural seller of training compute, sitting alongside AWS, Azure, and GCP in the lab procurement decks. That's a far weirder business than "rockets plus satellites," and the S-1 is the first document that makes it impossible to ignore.
Crazy this company will IPO for >1B with such bad financials! That said, Starlink seems to be a real cash machine, not as good as ads but enough to support AI bets.2025:- Revenue: $18.7B, up from $14.0B in 2024- Operating loss: -$2.6B- Net loss: -$4.9B- Adjusted EBITDA: $6.6B- Operating cash flow
If any company can put profitable data centers in space, it will be SpaceX. But I doubt that any company can. The difficulties of the physics and engineering of cooling seem like they will always outweigh the advantages of keeping your data center on Earth.I am annoyed by the insistence that the val
It’s surprising just how low the revenue is for SpaceX. There are some 700+ companies with larger revenue figures, and yet just a small handful exceed SpaceX’s proposed valuation.In 2026 one gets the impression that SpaceX is a huge company, among the largest in the world. It’s wild to see that its
SpaceX is incredibly exciting, but I was skeptical when XAI and Twitter were rolled into it. The S-1 here makes it even more disappointing.I did want a piece of SpaceX but the valuation here is pretty eye watering compared to the fundamentals. I don't think I can put my money into this, althoug
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
"in May 2026, we entered into Cloud Services Agreements with Anthropic PBC (“Anthropic”), an AI research and development public benefit corporation, with respect to access to compute capacity across COLOSSUS and COLOSSUS II. Pursuant to these agreements, the customer has agreed to pay us $1.25