According to SemiAnalysis, Anthropic's inference business could achieve a margin of up to 88% considering computing costs. Approximately 75–85% of annual recurring revenue comes from APIs, whose margins exceed 80%.
Subscriptions account for about 10% of revenue but consume more than 40% of computing resources, with a margin of approximately 50%. Subscribers receive four times more computing power per dollar than API buyers, and on certain workloads, up to five times more value than OpenAI's API. Subscriptions reduce revenue per megawatt by $36 million but do not lead to losses. Anthropic's gross margin increased from minus 94% in 2024 to the mid-60s by mid-2026. Previous forecasts estimated reaching $1 billion in quarterly operating profit by the third quarter of 2026. Note: 88% is an estimate only for inference, as model training remains extremely expensive.
Goldman Sachs believes that the cost of computing rentals is significantly higher than the revenue needed for profitability in the AI industry, meaning clients pay a premium for scarcity. SpaceX's AI segment generated $2.56 billion in the second quarter of 2026, up 247% year-over-year, and showed its first positive adjusted earnings before interest, taxes, and amortization of $1.1 billion. Anthropic pays about $1.25 billion per month until May 2029, Google $920 million per month, and a third client $1.11 billion per month from December 2026. Total annual commitments exceed $40 billion. Goldman estimates a gap of $230 billion for hyperscalers and forecasts SpaceX's AI revenue growth from $3.2 billion in 2025 to $322 billion by 2030. Capacities: 1.4 GW in June 2026, more than 2 GW by the end of the year, 5–10 GW in 2027. Risks: own capacity growth undermines the premium, and high client concentration. SpaceX is traded on Nasdaq under the ticker SPCX.
This material is prepared solely for informational purposes and does not constitute financial advice or a recommendation.