AI Compute Spot Pricing Runs 2-4x Above Long-Term Contracts and Hyperscaler Breakeven
TREE NEWS reports: The market for AI compute leasing in the United States is showing a stark “expensive short-term, cheap long-term” structure. One gigawatt of annualized compute revenue is priced at roughly $40-50 billion on short-term spot leases, versus $20-30 billion on long-term contracts — while hyperscaler self-built capacity breaks even at just $12 billion per gigawatt.
SpaceX told investors its new hosting contracts were priced at the high end of a historical $30-50 billion per gigawatt range. Nebius disclosed Q3 short-term pricing of $40-50 billion/GW. CoreWeave disclosed recent short-term contracts around $40 billion/GW. IREN disclosed three-year contracts above $20 billion/GW, with deals under negotiation near $25 billion/GW. Deutsche Bank analyst Edison Yu’s teardown of SpaceX GPU-hour pricing implies Google and Reflection AI contracts at $50-51 million per megawatt, and two undisclosed customers at $60-61 million per megawatt — equivalent to as much as $61 billion per gigawatt, with GB300 chips renting near $14 per hour.
Backwardation: A Commodity Market Signal
Short-term spot prices are roughly double long-term contract prices, with the long-term floor holding near $20 billion/GW. This term structure is known in commodity markets as backwardation — typically a signal that the market views current shortages as temporary. Buyers are willing to pay a premium for “not committing,” suggesting demand is real but even buyers are unsure how durable it is.
For hyperscale cloud operators, Goldman calculates all-in capex at roughly $42 billion per gigawatt. At $40-50 billion in annual spot revenue, construction costs are recovered in about one year; even at the $20 billion long-term floor, just two years. But Goldman flags a critical caveat: most hosting contracts are cancellable by either party within 90 days, so recognizing revenue at current prices through contract expiry — mostly 2029 — carries optimistic risk. A $50 billion annualized contract is legally a quarterly lease.
Hyperscaler ROI Threshold: $11.6 Billion per GW
Goldman’s tech team estimates the six largest US hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, SpaceX) need roughly $1.42 trillion in cumulative AI revenue between 2028 and 2030 — about $11.6 billion per gigawatt annually — to earn a 15% ROIC on 2026-2027 AI compute investment. Even at a 30% ROIC target and maximum capex assumptions, the figure stays under $18.6 billion per GW. Current short-term supercompute cloud rents are 2-4x that threshold.
Google is paying roughly $45 billion per gigawatt to lease SpaceX compute, while Goldman’s framework shows self-built data centers need only about $12 billion per gigawatt in returns. No one pays a 4x premium for a commodity unless they have no choice — Google leases because self-build cannot keep pace. Once self-built capacity catches up, the leasing premium disappears and supercompute cloud revenue shrinks sharply.
Token Economics Cannot Support Spot Prices
JPMorgan analyst Gokul Hariharan offers an optimistic framework: frontier model inference at 60-80% gross margin generates $20-40 billion in annualized token revenue per gigawatt, up from $10 billion in 2025. But if compute costs are 20-40% of token revenue, a lab leasing a gigawatt at $45 billion/year needs to sell $110-225 billion in tokens from that gigawatt to break even — 3-11x JPMorgan’s own optimistic range. The conclusion: no one leases spot compute at $45 billion/GW to serve paying customers profitably. They lease to train the next model, funded by the next equity or debt round — mostly debt. This is the core argument behind Rothschild Redburn analyst Alexander Haissl’s sell ratings on CoreWeave and Nebius: training demand persists only if external capital remains abundant.
Same Money, Counted Multiple Times
Supercompute cloud revenue, hyperscaler revenue, and “AI industry” revenue are largely the same money counted multiple times. A dollar of end-user spend flows to Anthropic (booked at gross including cloud partner share), to the cloud partner (booked as cloud revenue), to SpaceX or CoreWeave (hosting revenue), and to Nvidia (data center revenue). Summing the chain yields a multiple of what end users actually pay.
This week’s AI stock selloff was triggered by a Financial Times report that OpenAI’s annualized revenue through September was about $50 billion, below the roughly $70 billion MRR figure circulating. Goldman TMT specialist Sean Johnstone explained the gap: Anthropic reports closer to gross (including cloud partner spend), OpenAI closer to net (its own share). Reconciling the two produced the higher $40 billion then $70 billion figures. Even the $50 billion number is based on a shorter-period annualized sales forecast — itself an upward stretch.
Key Takeaways for Investors
- Spot compute prices are a scarcity marginal price, not an average price the overall buildout can sustain. Pricing 41 GW of 2026-2027 hyperscaler capex: ~$470 billion/year at Goldman’s 15% ROIC threshold, ~$810 billion at long-term contract floors, ~$1.8 trillion at current spot. Top three AI labs did roughly $100 billion combined annualized in July; investor Brad Gerstner says $180-200 billion by year-end is needed to “keep the AI trade alive”; Goldman’s portfolio strategy team says the market “now needs to see” evidence of ~$300 billion annualized AI revenue.
- Capacity expansion will compress the scarcity premium. SpaceX’s compute footprint is forecast to grow from 1.4 GW in Q2 2026 to ~7 GW by end-2027 and ~10.6 GW by end-2028 — over 7x in four years. Industry supply is rising fast. In shipping and other commodity markets, nothing cures record prices like record prices.
- Watch credit markets. Goldman TMT has flagged Oracle and Broadcom credit spreads as key indicators of debt-driven AI capex risk. Oracle’s five-year CDS closed at 261 basis points Thursday, a record high. The question is whether end-user revenue arrives before credit markets do — or whether chip-collateralized debt instruments priced for spot as a permanent state discover the answer first.
- Returns will come from selecting winners, not from holding anything AI-capex-related.
