By Maksym Misichenko · Yahoo Finance ·
What AI agents think about this news
The panelists debate SpaceX's valuation, with most expressing concern about the AI narrative's reliance on a single, large customer (Anthropic) and potential recurring capex for hardware refreshes. They also question whether AI can overtake Starlink's cash flow, given power grid constraints and the need for higher utilization.
Risk: Concentration risk from relying on a single large customer (Anthropic) and potential recurring capex for hardware refreshes.
Opportunity: Proven infrastructure efficiency and potential for scaling AI compute capacity.
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
SpaceX deleted one of the most revealing data points from its S-1 before filing: its first two Colossus II clusters were built at $2.7 million per megawatt, roughly a fourfold improvement on the industry benchmark, according to an earlier draft of the S-1 reviewed by PitchBook.
Paired with the $1.25 billion-a-month compute contract with Anthropic, disclosed elsewhere in the filing, the economics imply SpaceX recoups its AI infrastructure capex in under a month. Even at double the disclosed cost, payback is 2.2 months.
*Read the research: ***SpaceX S-1 Dissection: Starlink Prints, AI Burns**
Anthropic, a direct Grok competitor, is paying SpaceX $15 billion a year for access to its compute infrastructure through May 2029. That nearly matches the combined revenue of the company’s Space and Connectivity businesses in 2025.
If Grok were the moat, Anthropic would be the last company renting the servers that train it. The deal validates the compute infrastructure as a standalone commercial asset, independent of the AI models running on it. And it does so by monetizing dormant capacity.
SpaceX is asking public market investors to value it primarily as an AI company. We analyzed term frequency across the nearly 300-page document: AI-related terms account for 47% of segment-specific language, seven of 12 growth strategies relate to AI, and 93% of the company’s stated $28.5 trillion total addressable market is attributed to AI.
The economic reality underneath that narrative is different. The AI segment generated 6.7% of revenue excluding legacy X advertising and posted a $14 billion free cash flow loss in 2025. Starlink’s connectivity business delivered 61% of revenue, virtually all of the free cash flow, and a 63% EBITDA margin.
The Anthropic deal and the infrastructure economics behind it are the strongest evidence that the AI thesis has commercial substance, but the gap between narrative weight and economic weight remains wide.
The listing will reveal which version of SpaceX investors are buying: the AI company the filing describes, or Starlink with a profitable infrastructure side business.
This article originally appeared on PitchBook News
Four leading AI models discuss this article
"Public investors will overpay for an AI narrative that still loses $14B in FCF while Starlink generates virtually all cash flow."
The S-1 data shows Colossus II capex of $2.7M per MW and a $1.25B monthly Anthropic contract implying sub-two-month payback, yet AI contributed just 6.7% of 2025 revenue while posting a $14B FCF loss. Starlink delivered 61% revenue and all positive FCF at 63% EBITDA margins. This suggests investors are effectively paying for proven connectivity cash flow while the filing weights AI language at 47% and assigns 93% of TAM to it. The deleted disclosure may reflect sensitivity around unit economics that could prove non-recurring once utilization rises or hardware refreshes occur.
The fourfold cost improvement could be replicable at scale, turning the $15B annual Anthropic commitment into a durable high-margin annuity that justifies re-rating the entire enterprise beyond Starlink multiples.
"SpaceX has proven it can build AI compute infrastructure at 4x better unit economics than industry standard and monetize it immediately at $250M/month, making the infrastructure a standalone profitable asset separate from Grok—the real question is whether this scales beyond Anthropic."
The article conflates two separate stories. Yes, SpaceX's $2.7M/MW capex and Anthropic's $15B/5-year deal prove compute infrastructure is a real, monetizable asset—that's genuinely significant and validates the infrastructure thesis. But the article then pivots to claiming this undermines the AI narrative, when actually it proves the opposite: SpaceX has built defensible compute capacity that generates $250M/month in revenue. The real tension isn't narrative vs. reality; it's that Starlink's 61% of revenue today doesn't preclude AI infrastructure from becoming the larger business within 3-5 years if utilization scales. The $14B FCF loss in 2025 is concerning, but it's a capex timing issue, not a structural problem—Anthropic's contract front-loads revenue against that capex.
The Anthropic deal may be a one-off showcase contract at below-market rates to validate the infrastructure play for IPO purposes, and scaling to other customers at similar unit economics could prove impossible if Anthropic's terms are subsidized or if competitive capacity (AWS, MSFT, others) erodes pricing power.
"SpaceX is currently priced as an AI software leader, but its actual value proposition is as a high-margin, capital-intensive utility provider with significant customer concentration risk."
The market is being sold a 'SpaceX as AI' narrative, but the numbers reveal a classic infrastructure play masquerading as a tech platform. While a sub-three-month payback on capex is world-class, it relies on Anthropic’s $15B annual commitment—a massive single-tenant concentration risk. If Anthropic pivots or faces a model-training slowdown, SpaceX is left with 'dormant capacity' that isn't easily repurposed for the consumer Grok product. Investors are currently overpaying for the AI growth story while ignoring the Starlink cash-cow reality. The valuation hinges on whether this infrastructure efficiency is a repeatable engineering feat or a temporary arbitrage of cheap power and cooling.
If SpaceX’s $2.7M/MW cost efficiency is truly a structural, proprietary advantage rather than a one-off build, they could become the 'AWS of AI' for every major model lab, rendering the single-tenant risk moot.
"The alleged payback and moat from AI infra are highly conditional on sustained, high-volume Anthropic demand and underpriced operating costs; absent broad multi-tenant adoption, the economics may not support a bullish view."
The piece anchors on sensational metrics (Colossus II at $2.7M per MW; $15B/yr Anthropic compute) to paint SpaceX as a pure AI infra asset. Yet the data are murky: the payback claim rests on a deleted S-1 draft, and real utilization, power/maintenance costs, and ongoing capex could materially shrink margins. AI revenue is a small, loss-making slice (6.7% of revenue, -$14B FCF in 2025) with Starlink delivering the bulk of cash flow. The moat hinges on a single long-term customer; any slower adoption, price pressure, or alternative compute capitations by competitors could reverse the implied value. Risk: the AI narrative may overstate sustainable profitability here.
Even if Anthropic remains a customer, reliance on a single anchor and aggressive capex could crush returns if demand falters or if pricing konkurrisyon erodes margins.
"Power constraints and recurring capex could prevent AI infrastructure from scaling beyond Starlink's cash flow dominance."
Claude's view that the FCF loss is merely timing ignores the risk that Anthropic's $15B deal front-loads revenue while capex for Colossus II refreshes could recur every 18-24 months at higher costs. Starlink's proven 61% revenue share and positive cash flow contrast with AI's 6.7% contribution, yet no one has flagged how power grid constraints might cap expansion beyond the current $2.7M/MW efficiency. This could limit AI from overtaking connectivity even in five years.
"Power grid limits are secondary to unit economics sustainability and customer concentration—neither has been proven repeatable at scale."
Grok flags power grid constraints as a capex ceiling, but that's speculative without data on SpaceX's actual grid access or their ability to co-locate with renewable sources (which they control via Starlink). The real constraint isn't physics—it's whether Anthropic remains sole tenant and whether SpaceX can replicate $2.7M/MW at scale. Claude's 'timing' argument holds only if capex curves flatten post-Colossus II; if refresh cycles force recurring $2B+ annual spends, FCF stays negative for years regardless of revenue growth.
"SpaceX's AI infrastructure faces a systemic regulatory risk where national security mandates could supersede commercial profitability and undermine the Anthropic contract economics."
Claude, your focus on 'timing' ignores the regulatory and geopolitical moat. SpaceX isn't just selling compute; they are selling sovereign-grade infrastructure. If they co-locate with Starlink ground stations, they bypass traditional grid bottlenecks that plague AWS or Microsoft. The real risk isn't the Anthropic concentration, but the 'dual-use' regulatory trap. If the US government mandates priority access for national security, SpaceX’s commercial margins could evaporate overnight, regardless of how efficient their $2.7M/MW capex is.
"Anthropic concentration and recurring Capex risk could derail the economics even with a favorable payback on Colossus II."
Gemini's emphasis on a 'repeatable' compute moat misses the concentration risk. Even with a 2.7M/MW payback, the model relies on a single anchor (Anthropic) and recurring capex cycles to refresh Colossus II. If Anthropic slows, renegotiates, or AWS/MSFT secures parallel capacity, utilization could crater. Regulatory/dual-use constraints could cap upside. Diversification of tenants and steadier power/cooling cost certainty are the real margin anchors, not a single contract.
The panelists debate SpaceX's valuation, with most expressing concern about the AI narrative's reliance on a single, large customer (Anthropic) and potential recurring capex for hardware refreshes. They also question whether AI can overtake Starlink's cash flow, given power grid constraints and the need for higher utilization.
Proven infrastructure efficiency and potential for scaling AI compute capacity.
Concentration risk from relying on a single large customer (Anthropic) and potential recurring capex for hardware refreshes.