Written by: Long Yue
The U.S. AI compute leasing market is showing a clear "expensive short-term, cheap long-term" characteristic.
On October 10, according to a chart by Goldman Sachs internet analyst Eric Sheridan in a SpaceX quarterly preview report, the annualized revenue of one gigawatt of compute is about $40 billion to $50 billion for short-term spot leasing, about $20 billion to $30 billion for long-term contracts, while the break-even threshold for hyperscalers' self-built capacity is only about $12 billion.
But Goldman Sachs analyzed that the fundamental reason hyperscalers such as Google lease compute at high premiums is that self-built capacity cannot keep up. Once self-built capacity catches up, the spot premium will disappear. It predicts that SpaceX's compute scale alone will grow from 1.4 gigawatts in the second quarter of 2026 to about 7 gigawatts by the end of 2027, and about 10.6 gigawatts by the end of 2028—an expansion of more than 7x in four years. Compute supply across the entire industry is increasing rapidly.
Spot price: $50 billion per gigawatt
The Goldman Sachs report shows that SpaceX announced two new compute hosting contracts in the past quarter and told investors that these contracts were "priced at the high end of the historical discussion range of $30 billion to $50 billion per gigawatt."
SpaceX is not an isolated case.
Nebius (NBIS) disclosed in August that short-term compute contract pricing in the third quarter was about $40 billion to $50 billion per gigawatt.
CoreWeave (CRWV) disclosed that recent short-term contract pricing was about $40 billion per gigawatt.
IREN disclosed that recent three-year contract pricing exceeded $20 billion per gigawatt, with contracts under negotiation at about $25 billion per gigawatt.
Nebius long-term contract pricing is also in the range of $20 billion to $25 billion per gigawatt.
Deutsche Bank analyst Edison Yu further broke down the GPU-hour pricing of SpaceX's various contracts in his "Got compute?" report, concluding that the contracts with Google and Reflection AI were about $50 million to $51 million per megawatt, while the other two undisclosed customers were about $60 million to $61 million per megawatt—equivalent to as much as $61 billion per gigawatt, with the hourly rental rate for GB300 chips approaching $14.
"Spot price" is 2x the "long-term contract price"
Short-term spot prices are about twice long-term contract prices. The floor for long-term agreement pricing is currently maintained at about $20 billion/GW.
This term structure inversion has a specialized term in commodity markets: backwardation. It usually means the market views the current shortage as temporary.
In other words, buyers are willing to pay a premium for "not committing," which precisely shows that although demand is real, even buyers themselves are not sure whether this demand can persist.
For supercomputing cloud operators, Goldman Sachs estimates that all-in capex for supercomputing cloud is about $42 billion per gigawatt. Based on spot revenue of $40 billion to $50 billion per year, the entire construction cost can be recovered in about one year; even at the long-term contract floor of $20 billion, it would take only two years.
But there is one key detail that cannot be ignored. Goldman Sachs specifically noted in the report:
Most of these hosting contracts can be canceled by either party within 90 days, so our assumption that revenue is recognized at current prices through contract expiration (mostly 2029) carries a risk of being overly optimistic.
A contract with $50 billion in annualized revenue is legally a quarterly lease.
Goldman Sachs return calculation: cloud providers need about $11.6 billion in revenue per GW
Goldman Sachs' technology analysis team attempted last month to estimate the scale of AI revenue needed for hyperscaler capex, concluding that the six major U.S. hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, SpaceX) need to generate a cumulative total of about $1.42 trillion in AI revenue from 2028 to 2030, equivalent to about $11.6 billion per gigawatt per year, in order to achieve a 15% return on invested capital (ROIC) on AI compute investment in 2026 to 2027.
Even if the ROIC target is raised to 30% and the highest capex assumptions are used, this figure does not exceed $18.6 billion per gigawatt.
Current short-term supercomputing cloud rents are 2x to 4x this threshold.
What does this mean?
Google leases SpaceX compute at about $45 billion per gigawatt (Deutsche Bank estimates $50 billion), while Goldman Sachs' ROIC framework shows that Google only needs a return of about $12 billion per gigawatt for its own self-built data centers.
No one would pay a 4x premium for a commodity unless there is no other choice. The analysis points out that Google leases compute because its self-built speed cannot keep up. Once self-built capacity catches up, the leasing premium will disappear, and supercomputing cloud revenue will shrink substantially as a result.
Even the most optimistic token economics cannot support spot prices
JPMorgan analyst Gokul Hariharan provided a relatively optimistic framework: frontier model companies' inference business has a gross margin of 60% to 80%, and each gigawatt of compute can generate $20 billion to $40 billion in annualized token revenue (compared with $10 billion in 2025).
But do a simple math check:
If compute cost accounts for 20% to 40% of token revenue (that is, 60% to 80% gross margin), then a lab leasing one gigawatt of compute at $45 billion per year would need to sell about $110 billion to $225 billion of tokens from that gigawatt to cover costs and make a profit.
That is 3x to 11x JPMorgan's own optimistic forecast range.
There is only one conclusion: no one leases spot compute at $45 billion per gigawatt to serve paying customers and make a profit. They lease compute to train the next model, funded by the next round of equity or debt financing—mainly debt.
This is also the core argument when Rothschild Redburn analyst Alexander Haissl issued sell ratings on CoreWeave and Nebius last month. He directly pointed out:
Training demand can continue only if external capital remains abundant.
AI revenue: the same money counted several times
Supercomputing cloud revenue, hyperscaler revenue, and the AI "industry" are to a large extent the same money, but counted multiple times.
Track where one dollar of end-user spending flows:
An enterprise pays Anthropic, and Anthropic books it at gross revenue (including cloud partner revenue sharing);
The cloud partner (AWS, Google) counts its share in cloud revenue;
If that partner lacks sufficient compute, it leases one gigawatt from SpaceX or CoreWeave, and the latter counts hosting revenue;
SpaceX and CoreWeave pay Nvidia, and Nvidia counts data center revenue.
Adding up "AI revenue" across the entire industry chain produces a number that is several times the amount actually paid by end users.
And this week's sharp decline in AI stocks was triggered by a Financial Times report that OpenAI's annualized revenue as of the end of September was about $50 billion, lower than the previously circulated figure of about $70 billion in monthly recurring revenue (MRR).
Goldman Sachs TMT expert Sean Johnstone explained the source of the gap:
Anthropic is closer to a gross revenue basis (including all customer spending of cloud partners, with partner revenue sharing counted as cost). OpenAI is closer to a net revenue basis (mainly its own share). When some investors tried to align the two bases and gross up OpenAI's data, higher figures emerged—about $40 billion in August and then about $70 billion... This event highlights the market's high sensitivity to the revenue trajectories of two private companies using different reporting bases.
According to Bloomberg, even the $50 billion figure is "a forecast of annualized sales based on a shorter period," which is itself an upward-stretched calculation method.
Spot prices cannot support the overall scale of construction
In the end, the math itself explains the problem.
Dividing Goldman Sachs' $1.42 trillion revenue requirement by $11.6 billion per gigawatt per year implies that the six major hyperscalers' 2026 to 2027 capex corresponds to about 41 gigawatts of AI compute.
Price these 41 gigawatts at different prices:
At Goldman Sachs' 15% ROIC threshold (about $11.6 billion/GW): about $470 billion in AI revenue per year
At the long-term hosting contract floor ($20 billion/GW): about $810 billion per year
At the current short-term supercomputing cloud spot price (about $45 billion/GW): about $1.8 trillion per yearCompare with reality: the top three AI labs had combined annualized revenue of about $100 billion in July; investor Brad Gerstner believes it needs to reach $180 billion to $200 billion by year-end to "sustain the AI trade logic"; Goldman Sachs' portfolio strategy team says the market "now needs to see" evidence of about $300 billion in annualized AI revenue.
Current supercomputing cloud spot prices are the marginal price of scarce compute, not the average price achievable at the overall buildout scale.
Summary from Goldman Sachs' TMT team:
Future returns should increasingly come from picking winners... rather than holding anything related to AI capex.
Capacity expansion will compress the scarcity premium.
Goldman Sachs forecasts that SpaceX alone will grow its compute scale from 1.4 gigawatts in Q2 2026 to about 7 gigawatts by the end of 2027, and about 10.6 gigawatts by the end of 2028—an expansion of more than 7x in four years.
Compute supply across the entire industry is increasing rapidly.
The analysis points out that in shipping and other commodity markets, "nothing cures record prices like record prices." The term curve is already telling the market where AI compute is headed: long-term contract prices are only half of spot prices.
The only question is: can end-user revenue arrive before the credit market does—or will chip-collateralized debt instruments priced on the assumption that spot prices are a permanent state discover the answer on their own first.
It is worth noting that Goldman Sachs' TMT team has listed Oracle and Broadcom credit spreads as key indicators for monitoring debt-driven AI capex risk. Oracle's five-year credit default swap (CDS) closed at 261 basis points on Thursday, a record high.