The US AI compute rental market is showing a clear pattern of "expensive short-term, cheap long-term."

On October 10, according to charts from Goldman Sachs internet analyst Eric Sheridan in a quarterly preview report on SpaceX, the annualized revenue per gigawatt of compute is approximately $40 billion to $50 billion for short-term spot rentals, about $20 billion to $30 billion for long-term contracts, while the breakeven threshold for hyperscaler self-built capacity is only about $12 billion.

However, Goldman Sachs analyzed that the fundamental reason hyperscalers like Google rent compute at such high premiums is that their self-built capacity cannot keep up. Once self-built capacity catches up, the spot premium will disappear. The firm predicts that SpaceX's compute scale alone will grow from 1.4 gigawatts in the second quarter of 2026 to approximately 7 gigawatts by the end of 2027 and about 10.6 gigawatts by the end of 2028—an expansion of more than sevenfold within four years. Compute supply across the entire industry is rapidly increasing.

Spot Prices: $50 Billion Per Gigawatt

Goldman Sachs' report shows that SpaceX announced two new compute hosting contracts in the past quarter and told investors these contracts were "priced at the high end of the historical discussion range of $30 billion to $50 billion per gigawatt."

SpaceX is not alone.

Nebius (NBIS) disclosed in August that short-term compute contract pricing in the third quarter was approximately $40 billion to $50 billion per gigawatt.

CoreWeave (CRWV) disclosed that recent short-term contract pricing was approximately $40 billion per gigawatt.

IREN disclosed that recent three-year contract pricing exceeded $20 billion per gigawatt, with contracts under negotiation at approximately $25 billion per gigawatt.

Nebius' long-term contract pricing is also in the $20 billion to $25 billion per gigawatt range.

Deutsche Bank analyst Edison Yu further broke down the GPU-hour pricing of SpaceX's various contracts in his "Got compute?" report, concluding that contracts with Google and Reflection AI were approximately $50 million to $51 million per megawatt, while two other undisclosed customers were approximately $60 million to $61 million per megawatt—equivalent to as much as $61 billion per gigawatt, with GB300 chip hourly rental approaching $14.

Spot Prices Are Twice Long-Term Contract Prices

Short-term spot prices are approximately twice long-term contract prices. The floor for long-term agreement pricing is currently maintained at about $20 billion per gigawatt.

This term structure inversion has a specific name in commodity markets: backwardation. It typically means the market views the current shortage as temporary.

In other words, buyers are willing to pay a premium for "not committing," which precisely indicates that while demand is real, even buyers themselves are uncertain whether this demand can persist.

For supercomputing cloud operators, Goldman Sachs estimates that all-in capital expenditure for supercomputing cloud is approximately $42 billion per gigawatt. At spot revenue of $40 billion to $50 billion annually, the entire construction cost can be recovered in about one year; even at the long-term contract floor of $20 billion, it would only take two years.

But there is a critical detail that cannot be overlooked. Goldman Sachs specifically noted in its report:

Most of these hosting contracts can be cancelled by either party within 90 days, so our assumption that revenue is recognized at current prices through contract expiration (mostly 2029) carries an overly optimistic risk.

A revenue contract worth $50 billion annualized is legally a quarterly lease.

Goldman Sachs Return Calculation: Cloud Providers Need About $11.6 Billion Revenue Per GW

Goldman Sachs' technology analysis team attempted last month to calculate the scale of AI revenue needed to support hyperscaler capital expenditure, concluding that the six major US hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, SpaceX) need to collectively generate approximately $1.42 trillion in AI revenue from 2028 to 2030, equivalent to about $11.6 billion per gigawatt per year, to achieve a 15% return on invested capital (ROIC) on AI compute investment in 2026 to 2027.

Even raising the ROIC target to 30% and using the highest capital expenditure assumptions, this figure does not exceed $18.6 billion per gigawatt.

Current short-term supercomputing cloud rental rates are two to four times this threshold.

What does this mean?

Google is renting SpaceX compute at approximately $45 billion per gigawatt (Deutsche Bank estimates $50 billion), while Goldman Sachs' ROIC framework shows that Google only needs about $12 billion per gigawatt in returns from building its own data centers.

No one would pay a fourfold premium for a commodity unless they have no choice. Analysis indicates that Google rents compute because its self-built speed cannot keep up. Once self-built capacity catches up, the rental premium will disappear, and supercomputing cloud revenue will shrink significantly.

Even the Most Optimistic Token Economics Cannot Support Spot Prices

JPMorgan analyst Gokul Hariharan provided a relatively optimistic framework: frontier model companies have inference business gross margins of 60% to 80%, and each gigawatt of compute can generate $20 billion to $40 billion in annualized token revenue (compared to $10 billion in 2025).

But a simple mathematical verification:

If compute costs account for 20% to 40% of token revenue (i.e., 60% to 80% gross margin), then a lab renting one gigawatt at $45 billion per year would need to sell approximately $110 billion to $225 billion in tokens from that gigawatt to cover costs and be profitable.

This is three to eleven times JPMorgan's own optimistic forecast range.

The only conclusion is: no one rents spot compute at $45 billion per gigawatt to serve paying customers and be profitable. They rent compute to train the next model, funded by the next round of equity or debt financing—primarily debt.

This is precisely the core argument Rothschild Redburn analyst Alexander Haissl made last month when issuing sell ratings on CoreWeave and Nebius. He directly stated:

Training demand can only persist if external capital remains abundant.

AI Revenue: The Same Money Counted Multiple Times

Supercomputing cloud revenue, hyperscaler revenue, and AI "industry" revenue are largely the same money, but counted multiple times.

Tracing the flow of one dollar of end-user spending:

Enterprises pay Anthropic, and Anthropic records it at gross revenue (including cloud partner revenue sharing);

The cloud partner (AWS, Google) records its share as cloud revenue;

If the partner lacks sufficient compute, it rents a gigawatt from SpaceX or CoreWeave, which record it as hosting revenue;

SpaceX and CoreWeave pay Nvidia, and Nvidia records it as data center revenue.

Adding up "AI revenue" across the entire supply chain yields a figure several times the amount end users actually pay.

The trigger for this week's AI stock selloff was a Financial Times report that OpenAI's annualized revenue through the end of September was approximately $50 billion, below the previously circulating figure of about $70 billion in monthly recurring revenue (MRR).

Goldman Sachs TMT expert Sean Johnstone explained the source of the discrepancy:

Anthropic is closer to a gross revenue basis (including all customer spending through cloud partners, with partner revenue sharing counted as cost). OpenAI is closer to a net revenue basis (primarily its own share). When some investors tried to align the two and restate OpenAI's data to a gross revenue basis, higher figures of about $40 billion in August and then about $70 billion emerged... This event highlights the market's heightened sensitivity to the revenue trajectories of two private companies using different reporting bases.

According to Bloomberg, even the $50 billion figure is "an annualized sales projection based on a shorter time period," itself an upward-stretched calculation method.

Spot Prices Cannot Support Overall Buildout Scale

Ultimately, the math speaks for itself.

Dividing Goldman Sachs' $1.42 trillion revenue requirement by $11.6 billion per gigawatt per year, the six major hyperscalers' 2026-2027 capital expenditure corresponds to approximately 41 gigawatts of AI compute.

Pricing these 41 gigawatts at different rates:

At Goldman Sachs' 15% ROIC threshold (about $11.6 billion per gigawatt): approximately $470 billion in AI revenue annually

At the long-term hosting contract floor ($20 billion per gigawatt): approximately $810 billion annually

At current short-term supercomputing cloud spot prices (about $45 billion per gigawatt): approximately $1.8 trillion annually

Compare with reality: the top three AI labs had combined annualized revenue of approximately $100 billion in July; investor Brad Gerstner believes $180 billion to $200 billion is needed by year-end to "sustain the AI trade thesis"; Goldman Sachs' portfolio strategy team says the market "now needs to see" evidence of approximately $300 billion in annualized AI revenue.

Current supercomputing cloud spot prices represent the marginal price of scarce compute, not the average price achievable across the overall buildout scale.

Goldman Sachs TMT team's summary:

Future returns should increasingly come from carefully selecting winners... rather than holding any AI capital expenditure-related names.

Capacity Expansion Will Compress Scarcity Premiums

Goldman Sachs predicts that SpaceX's compute scale alone will grow from 1.4 gigawatts in the second quarter of 2026 to approximately 7 gigawatts by the end of 2027 and about 10.6 gigawatts by the end of 2028—an expansion of more than sevenfold within four years.

Compute supply across the entire industry is rapidly increasing.

Analysis notes that in shipping and other commodity markets, "nothing cures high prices like high prices." The term curve is already telling the market where AI compute is heading: long-term contract prices are only half of spot prices.

The only question is: can end-user revenue arrive before the credit markets do—or will chip-backed debt instruments priced on the assumption that spot prices are permanent discover the answer on their own first.

It is worth noting that Goldman Sachs' TMT team has already listed Oracle and Broadcom's credit spreads as key indicators for monitoring risks from debt-driven AI capital expenditure. Oracle's five-year credit default swap (CDS) closed at 261 basis points on Thursday, a record high.