How expensive is AI computing power rental in the US? "Spot price" is twice that of long-term contracts, and four times the return threshold for cloud service providers.

The AI computing power leasing market in the United States is showing a clear “expensive short-term, cheap long-term” characteristic.

On October 10, according to a chart from Goldman Sachs internet analyst Eric Sheridan in SpaceX’s quarterly preview report, the annualized revenue for 1 gigawatt of computing power via short-term spot leasing reaches $40–50 billion, while long-term contracts are around $20–30 billion, and hyperscalers’ (hyperscaler) self-built capacity has a break-even threshold as low as $12 billion.

However, Goldman Sachs notes the root cause for hyperscalers such as Google to rent compute power at steep premiums is that their self-built capacity can’t keep up. Once self-built capacity catches up, the spot price premium will disappear. Their forecast: SpaceX’s computing power scale will grow from 1.4 GW in Q2 2026 to about 7 GW by the end of 2027, and roughly 10.6 GW by the end of 2028—a more than sevenfold expansion in four years. The entire industry’s supply of computing power is rapidly increasing.

Spot Price: $50 Billion per GW

According to the Goldman Sachs report, SpaceX announced in the past quarter two new compute colocation contracts, informing investors that these contracts were “at the high end of the historical discussion range $30–50 billion/GW.”

SpaceX is not alone.

• Nebius (NBIS) disclosed in August that its short-term compute contract pricing for Q3 was around $40–50 billion/GW.

• CoreWeave (CRWV) recently disclosed short-term contract pricing at about $40 billion/GW.

• IREN revealed recent three-year agreement pricing above $20 billion/GW, with contracts in negotiation around $25 billion/GW.

• Nebius’ long-term contract pricing is also in the $20–25 billion/GW range.

Deutsche Bank analyst Edison Yu further broke down SpaceX’s GPU hour pricing in his “Got compute?” report, concluding that Google and Reflection AI’s contracts are approximately $50–51 million per megawatt, while two undisclosed clients are at about $60–61 million per megawatt—which equates to as much as $61 billion per GW, with GB300 chip hourly rental rates approaching $14.

Spot Price Is Twice the Long-Term Contract Price

The short-term spot price is about twice the long-term contract price. The lower bound of long-term contract pricing is currently about $20 billion/GW.

This term structure inversion is described in commodity markets as backwardation, which generally means the market considers the current shortage as temporary.

In other words, buyers’ willingness to pay a premium for “no commitment” highlights that, while demand is real, even the buyers are unsure whether this demand can persist.

For hyperscale cloud operators, Goldman Sachs estimates all-in capex for supercomputing cloud is about $42 billion per GW. With spot revenue at $40–50 billion per year, construction costs could be recovered in a year; even at the long-term contract floor price of $20 billion, it would take only two years.

But there’s a key detail that can’t be ignored. Goldman Sachs specifically warns in the report:

Most of these colocation contracts are cancellable by either party within 90 days, meaning the practice of recognizing revenues at current pricing through contract expiry (mostly 2029) carries risks of being overly optimistic.

A revenue contract with an annualized value of $50 billion is, legally, a quarterly lease.

Goldman’s ROI Estimates: Cloud Firms Need About $11.6 Billion per GW in Revenue

Last month, Goldman Sachs’ tech analysis team attempted to estimate the AI income required to offset hyperscaler capex. Their finding: The six top US hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, SpaceX) will need to generate a cumulative $1.42 trillion in AI revenue from 2028 to 2030—translating to about $11.6 billion per GW per year to achieve a 15% ROIC (return on invested capital) on AI computing power invested in 2026–2027.

Even if the ROIC target is raised to 30% and the highest capex assumption is used, the number does not exceed $18.6 billion per GW.

Currently, short-term supercomputing cloud rental rates are2 to 4 times this threshold.

What does this mean?

Google rents SpaceX’s computing power for about $45 billion per GW (Deutsche Bank’s estimate: $50 billion), while Goldman Sachs’ ROIC framework shows Google’s self-built data centers would need only about $12 billion per GW in returns.

No one pays a 4x premium for a commodity unless there’s no choice. Analysis points out Google is renting compute because their self-build cannot keep pace. Once self-built capacity catches up, the rental premium will vanish and supercomputing cloud revenue will drop significantly.

Even the Most Optimistic Token Economics Cannot Sustain Spot Prices

JP Morgan analyst Gokul Hariharan provides an optimistic framework: frontier model companies have inference business gross margins of 60–80%, and each GW of compute could generate $20–40 billion in annualized Token revenue (compared to $10 billion in 2025).

But let’s do some simple math:

If compute costs are 20–40% of Token revenue (that is, 60–80% gross margin), then a lab renting 1 GW for $4.5 billion per year would need to sell approximately $110–225 billion in Tokens annually from that GW just to cover costs and make a profit.

That’s 3 to 11 times the optimistic projection bracket from JP Morgan.

There’s only one conclusion: Nobody rents spot compute at $4.5 billion per GW to serve paying customers and make a profit. They rent compute to train the next model, and their funding source is the next round of equity or debt financing—mainly debt.

This is also the core argument Rothschild Redburn analyst Alexander Haissl cited last month when issuing a sell rating on CoreWeave and Nebius. He straightforwardly noted:

Training demand can continue provided external capital remains ample.

AI Revenue: The Same Dollar Gets Counted Multiple Times

Supercomputing cloud income, hyperscaler revenue, and “industry-wide” AI revenue often amount to the same dollar being counted multiple times.

Follow the path of a dollar from the end-user:

1. An enterprise pays Anthropic, and Anthropic books total revenue (including share for cloud partners);

2. Cloud partners (AWS, Google) count their portion as cloud revenue;

3. If the partner is short on compute, they rent 1 GW from SpaceX or CoreWeave, who then count it as colocation income;

4. SpaceX and CoreWeave pay Nvidia, who books it as data center revenue.

When adding up “AI revenue” throughout the value chain, the resulting number is several times what the end-user actually paid.

This week’s sharp decline in AI stocks was triggered when the Financial Times reported OpenAI's annualized revenue at the end of September was about $5 billion, lower than earlier rumors of $7 billion in monthly recurring revenue (MRR).

Goldman’s TMT expert Sean Johnstone explained the discrepancy:

Anthropic is closer to gross revenue (including total client spend for cloud partners, where partner share is treated as cost). OpenAI is closer to net revenue (mainly its own share). When some investors try to unify both and restate OpenAI data on a gross revenue basis, you get the higher numbers—about $4 billion in August, around $7 billion later... This event highlights market sensitivity to the revenue trajectories of two private companies with different reporting standards.

According to Bloomberg, even the $5 billion figure is “based on annualized projections from a shorter time frame,” which itself is an upward-stretched calculation method.

Spot Price Can’t Support Overall Construction Scale

In the end, the math explains everything.

If you divide Goldman’s $1.42 trillion revenue requirement by $11.6 billion per GW per year, it calculates that the six hyperscalers’ capex in 2026–2027 corresponds to about 41 GW of AI compute.

Pricing out these 41 GW by different benchmarks:

• At Goldman’s 15% ROIC threshold (about $11.6 billion/GW): About $470 billion in AI revenue per year

• At the long-term colocation contract floor price ($20 billion/GW): About $810 billion per year

• At today’s short-term supercomputing cloud spot price (about $45 billion/GW): About $1.8 trillion per year

Reality check: In July, the top three AI labs had a combined annualized revenue of around $100 billion; investor Brad Gerstner believes $180–200 billion is needed by year-end to “maintain AI deal logic”; Goldman’s portfolio strategy team says the market “now needs to see” evidence of about $300 billion in annualized AI revenue.

The current supercomputing cloud spot price is the marginal price of scarce computing power, not the average price realizable across the total build-out scale.

Goldman’s TMT team summarizes:

Future returns should increasingly come from picking select winners...not from holding anything merely tied to AI capex growth.

Capacity Expansion Set to Squeeze Scarcity Premium

Goldman predicts SpaceX’s computing power scale will grow from 1.4 GW in Q2 2026 to about 7 GW by end 2027 and about 10.6 GW by end 2028—a more than sevenfold expansion in four years.

The entire industry's computing power supply is rapidly increasing.

Analysis points out that in shipping and other commodity markets, “nothing cures record prices like record prices.” The term structure is signaling the direction for AI computing power: long-term contract prices are only half the spot rates.

The only question: can end-user revenue arrive ahead of the credit markets—or, will the chip-collateralized debt instruments (priced as if spot rates are permanent) discover the answer first.

It’s worth noting that Goldman’s TMT team now tracks the credit spreads of Oracle and Broadcom as key risk indicators for debt-driven AI capex. Oracle’s five-year credit default swap (CDS) closed Thursday at 261 basis points, an all-time high.