The hyperscaler AI break-even gap

Goldman Sachs calculates that hyperscalers must generate roughly $300 billion in annual AI revenues within a few years to break even on their investments, with that figure implying roughly $1 trillion in annual end-user AI spending economy-wide. Cloud revenue growth for Amazon, Alphabet, Microsoft, and Oracle accelerated from 25% in 2024 to 48% in Q2 2026, annualizing roughly $70 billion above pre-AI trends, but Goldman Sachs places these figures below the break-even threshold. Announced revenue backlogs exceed $1.5 trillion.

The five largest hyperscalers hold over $3 trillion in future data center commitments, with $1.3 trillion in lease form. Alphabet alone holds more than $1 trillion in commitments, nearly eight times its total from a year earlier. Meta's commitments total $810 billion, with just under half tied to unstarted data center leases running up to 30 years. Approximately 70% of the $1.3 trillion in future leases was added in the last four quarters, with Meta and Google leading the most recent tranche since mid-2025.

Goldman Sachs and Visible Alpha project combined capex across eight major tech companies at roughly $0.46 trillion in 2025, $0.92 trillion in 2026, $1.5 trillion in 2027, and about $1.75 trillion by 2030. A separate projection places combined capex from a broader group at $280 billion in 2024 rising to $951 billion in 2026 and $1.5 trillion by 2028, with the author arguing Wall Street has repeatedly underestimated how aggressively hyperscalers spend.

Demand signals support continued growth. Fireworks AI is now processing more than 40 trillion tokens per day, more than double its volume from summer 2026, with the broader open model inference economy tracked through public disclosures from Baseten, Fireworks, Together AI, and OpenRouter, though Nathan Lambert of Interconnects described all available metrics as 'incredibly imprecise.' A Menlo Ventures report found that the average company spent 4.7x its planned 2025 AI budget against a forecast of only 1.9x, driven primarily by a small group of heavy AI agent users.

To fund the buildout, the largest hyperscalers issued nearly $200 billion in debt in 2026, with total investment-grade issuance on pace to exceed $2 trillion in 2026. Ancora projects AI-related debt through 2030 at an estimated $4 trillion and reports the surge is competing with Treasuries for investor capital, contributing to upward pressure on long-term yields; the U.S. 10-year Treasury yield is up approximately 85 basis points from its 12-month low and the two-year yield is up about 100 basis points.

Annual data center power additions driven by AI accelerator sales are projected to grow from 9 GW in 2024 to 104 GW by 2030, with average rack-level power density expected to rise from roughly 40 kW to over 280 kW over the same period. ON Semiconductor CEO Hassane El Khoury said the AI race is shifting from a computing problem to a power delivery problem, arguing that small efficiency gains in power delivery using silicon carbide and gallium nitride semiconductors can expand effective computing capacity without requiring more raw electricity.

GPU supply constraints are shaping near-term procurement. Chinese AI and cloud firms including Alibaba, Tencent, and ByteDance have ordered roughly 6 million Nvidia RTX Pro 5500 cards totaling approximately $70 billion, but Nvidia reportedly told partners it can produce at most 3 million cards worldwide in 2027 as TSMC capacity shifts toward Rubin data center chips, leaving even buyers' own estimates of roughly 700,000 deliveries by end of 2027 covering only about 12% of orders.

Sundar Pichai said that compute capacity for 2027 is effectively already allocated, forcing Google to lease short-term compute in the prior year to help customers bridge the gap, and acknowledged that AI labs today are constrained by infrastructure bets made roughly two years ago that no company foresaw would be insufficient for the demand that materialized. Google Cloud is at any given moment forgoing short-term revenues to make compute tradeoffs supporting frontier AI development.

Sam Altman, speaking at ADIPEC, argued that energy is the one AI cost that cannot be engineered away, contrasting it with model efficiency, chip fab automation, and hardware assembly. Altman also said the energy cost per AI query has dropped by 1,000x to 100,000x over three to four years.

Reserve Bank of Australia Assistant Governor Sarah Hunter said Australia's AI investment surge is pushing up inflation through data center construction costs and equipment prices before productivity gains have arrived. Hunter also said AI has had minimal impact on overall headcount so far, but job postings and employment are softening in AI-exposed occupations, disproportionately affecting younger workers, and that equity and corporate debt markets could reprice sharply if AI returns disappoint.