According to Goldman Sachs internet analyst Eric Sheridan’s SpaceX quarterly preview report dated Oct. 10, the annualized short-term Spot rental rate for AI Hashrate is approximately $40–50 billion per GW. The six major cloud providers need to cumulatively generate approximately $1.42 trillion in AI revenue from 2028 to 2030 to achieve a 15% return on their AI Hashrate investments in 2026–27.
Spot and Long-Term Contract Pricing
According to Goldman Sachs’ Oct. 10 report and company disclosures, contract pricing among companies in the AI Hashrate market is as follows:
Nebius (NBIS): Q3 short-term Hashrate contracts at $40–50 billion per GW; long-term contracts at $20–25 billion per GW
CoreWeave (CRWV): Recent short-term contracts at approximately $40 billion per GW
IREN: Recent three-year contracts at more than $20 billion per GW, with contracts under negotiation at approximately $25 billion per GW
SpaceX: Signed contracts at “high-end pricing within the historical discussion range of $30–50 billion per GW,” according to the company’s statements to investors
Edison Yu of Deutsche Bank calculated that Google’s and Reflection AI’s SpaceX contracts are approximately $50–51 million per MW, equivalent to $61 billion per GW, with GB300 chip hourly rental rates near $14.
Goldman Sachs noted that this “short-term expensive, long-term cheap” structure, known as backwardation in the Commodities market, usually indicates that the market views the current shortage as temporary. Another key detail is that most of these custody contracts can be canceled by either party within 90 days, and Goldman Sachs warned that assuming current prices will remain in effect through contract expiry carries the risk of being overly optimistic.
Goldman Sachs ROIC Framework: Six Major Cloud Providers Need $1.42 Trillion in Cumulative AI Revenue from 2028–2030
According to calculations by Goldman Sachs’ technology analysis team, for the six major U.S. hyperscale cloud providers—Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX—to achieve a 15% ROIC on AI Hashrate investments in 2026–2027, they need to cumulatively generate approximately $1.42 trillion in AI revenue from 2028 to 2030, equivalent to approximately $11.6 billion per GW per year. Even if the ROIC target is raised to 30%, the figure would not exceed $18.6 billion per GW. Google, however, is renting SpaceX Hashrate at approximately $45 billion per GW, estimated by Deutsche Bank at approximately $50 billion per GW—two to four times the threshold in Goldman Sachs’ ROIC framework. Goldman Sachs noted that Google is renting Hashrate because its own construction cannot keep pace; once its self-built capacity catches up, the rental premium will disappear.
JPMorgan analyst Gokul Hariharan’s framework shows that if laboratories operate at an annual Spot Hashrate cost of $45 billion per GW, they need to generate $110–225 billion in Token revenue per GW to be profitable—three to 11 times JPMorgan’s own optimistic forecast range.
AI Revenue Calculation: The Same Money Is Counted Multiple Times Across the Industry Chain
According to analysis by Goldman Sachs TMT expert Sean Johnstone, AI industry revenue faces a double-counting problem: enterprises pay Anthropic, which records total revenue, including the share allocated to cloud partners; cloud partners such as AWS and Google include their shares in cloud revenue; if they lack sufficient Hashrate, they rent from SpaceX or CoreWeave, which record hosting revenue; SpaceX and CoreWeave then pay NVIDIA, which records data center revenue. As a result, the “AI revenue” total across the entire industry chain is several times the amount actually paid by end users.
Regarding OpenAI’s revenue methodology, the Financial Times reported that its annualized revenue as of the end of September was approximately $50 billion, below the previously circulated MRR of approximately $70 billion. Johnstone explained that Anthropic uses a gross revenue methodology, including all customer spending through partners, while OpenAI uses a net revenue methodology, creating confusion over the figures. Bloomberg also noted that the $50 billion figure itself was “an annualized sales forecast based on a shorter period.”
Oracle CDS Hits Historical High of 261 Basis Points
According to various data sources, Oracle’s five-year credit default swap (CDS) closed at 261 basis points on Thursday, a Historical High. Goldman Sachs’ TMT team has listed the credit spreads of Oracle and Broadcom as key indicators for monitoring the risk of debt-driven AI capital expenditures. On the supply side, Goldman Sachs forecasts SpaceX’s Hashrate capacity at 1.4 GW in Q2 2026 → approximately 7 GW by the end of 2027 → approximately 10.6 GW by the end of 2028, representing more than sevenfold growth in four years.
On the demand side, the three leading AI laboratories had combined annualized revenue of approximately $100 billion in July. Investor Brad Gerstner said that revenue needs to reach $180–200 billion by year-end to “maintain the AI trade thesis.” Goldman Sachs’ portfolio strategy team said the market “now needs to see” evidence of approximately $300 billion in annualized AI revenue.
Last month, Rothschild Redburn analyst Alexander Haissl issued sell ratings on CoreWeave and Nebius, with the core argument being that the sustainability of training demand depends on external capital remaining abundant.
Frequently Asked Questions
What are the Spot and long-term contract prices for AI Hashrate, respectively, and how large is the gap?
According to Goldman Sachs’ Oct. 10 report, annualized short-term Spot rental rates for AI Hashrate are approximately $40–50 billion per GW, long-term contracts are approximately $20–30 billion per GW, and the breakeven threshold for hyperscale cloud providers building their own capacity is approximately $12 billion per GW. Spot prices are approximately twice those of long-term contracts.
How much AI revenue does Goldman Sachs estimate the six hyperscale cloud providers need to achieve a 15% ROIC?
According to Goldman Sachs’ technology analysis team, the six hyperscale cloud providers—Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX—need to cumulatively generate approximately $1.42 trillion in AI revenue from 2028 to 2030, equivalent to approximately $11.6 billion per GW per year, to achieve a 15% ROIC on AI Hashrate investments in 2026–2027.
What is the latest status of Oracle’s CDS, and why has Goldman Sachs listed it as a monitoring indicator?
According to market data, Oracle’s five-year credit default swap (CDS) closed at 261 basis points on Thursday, a Historical High. Goldman Sachs’ TMT team has listed the credit spreads of Oracle and Broadcom as key indicators for monitoring the risk of debt-driven AI capital expenditures.
How can the gap between OpenAI’s $50 billion in annualized revenue and $70 billion be explained?
According to Goldman Sachs’ Sean Johnstone, Anthropic uses a gross revenue methodology, including all customer spending through cloud partners, while OpenAI uses a net revenue methodology, primarily reflecting its own share. The difference in methodologies resulted in higher figures such as $70 billion; the OpenAI annualized revenue reported by the Financial Times was approximately $50 billion.