Oracle books $664B in backlog on $19.3B a quarter

Oracle reported Q1 FY2027 on September 10 with $19.35 billion in revenue and a remaining performance obligation balance of $664 billion. That is roughly seven years of the company's own full-year revenue guidance, already signed. Cloud infrastructure grew 121% to $7.4 billion. The line that matters for anyone renting AI compute is in Oracle's own release: demand for AI training and inference services "continues to grow faster than supply."

What actually happened

Oracle's Q1 press release puts total revenue up 30% year over year, cloud revenue at $11.6 billion (up 62%), IaaS at $7.4 billion (up 121%), and SaaS at $4.2 billion (up 10%). Non-GAAP EPS was $1.92, up 30%.

The backlog is the headline. RPO landed at $664 billion, up $209 billion year over year, on more than $30 billion of new AI cloud contracts booked in the quarter alone. CNBC notes that beat the StreetAccount consensus of $630.6 billion, and shares rose about 7% after hours. Full-year guidance is at least $90 billion in revenue and $8.10 in adjusted EPS.

Oracle also disclosed what it costs to serve that backlog. It added 850 megawatts of datacenter capacity and delivered more than 300,000 GPUs to AI cloud customers — nearly triple the prior quarter. Operating cash flow hit $23 billion, up 184%, but free cash flow was negative $5 billion. Capital expenditure ran $28.5 billion against $8.5 billion a year earlier, and the company now carries roughly $125 billion in debt after a $20 billion at-the-market equity sale.

Why a $664B AI backlog matters for your business

You are not buying from Oracle. You are probably buying tokens from OpenAI or Anthropic, or renting a box from a smaller provider. It still reaches you, in two ways.

First, the "prices always fall" assumption now has a counterweight. Per-token prices have dropped steadily because model efficiency improved faster than compute got scarce. A backlog of this size says the capacity to serve 2028 is being sold today, and that a vendor financing it with $28.5 billion of annual capex and $125 billion of debt has a hard floor under what it can charge. Plan for your unit cost to flatten rather than keep halving.

Second, capacity contracts are a queue, and small buyers are at the back of it. If your roadmap assumes you can 10x inference volume next spring on the same on-demand terms, that assumption is worth testing with your provider before you build against it.

The practical move is boring and it works: make the model a swappable adapter. One interface, provider behind it, a config change to switch. We build systems where the LLM call is a single module with a fallback chain, because a studio that has actually run a business knows the vendor you sign with is not the vendor you'll be on in eighteen months. Measure tokens per completed task, not tokens per call — that is the number that survives a price change.

Sources: Oracle Q1 FY2027 results, CNBC: Oracle Q1 earnings report.