OpenAI on May 19 announced Guaranteed Capacity, a commit-based offering that lets enterprise customers reserve long-term access to OpenAI compute for production systems, customer-facing applications, and AI agents. Customers can choose one, two, or three-year commitments, with discounts that increase at higher annual spend levels, according to OpenAI’s product page.
“Customers are increasingly asking us for certainty on capacity. As models get better, we expect that the world will be capacity-constrained for some time,” CEO Sam Altman wrote on X. He called the arrangement a “big win-win” that helps OpenAI plan its infrastructure buildout while giving buyers predictable access.
How the Offering Works
Guaranteed Capacity commitments can be drawn down across OpenAI’s full product portfolio and across supported cloud providers and model families, according to CNBC. OpenAI will work with customers to model capacity against forecasted demand, product expansion, and multi-year AI adoption plans.
There is no public per-token pricing, no published list of qualifying products, and no disclosed minimum spend. As Kingy AI noted, “this isn’t being aimed at the self-serve developer signing up with a credit card. It’s aimed at the buyer whose AI bill is now a line item in the board pack.”
Altman said OpenAI will reserve enough capacity for its own products, including ChatGPT and Codex, but will offer Guaranteed Capacity to external customers until the current allocation sells out. The company plans to revive the offering periodically as infrastructure expands, per CNBC.
Compute Scarcity as the Bottleneck
The launch reinforces a pattern forming across the industry: compute access, not model capability, is becoming the primary constraint on enterprise AI adoption. OpenAI has told investors it is targeting roughly $600 billion in total compute spend by 2030, according to CNBC. Its Stargate infrastructure program crossed 10 gigawatts of contracted U.S. AI infrastructure in April, hitting a milestone originally planned for 2029.
The timing is deliberate. Salesforce committed $300 million to Anthropic tokens in 2026. Google announced Gemini Spark at I/O 2026 requiring dedicated cloud infrastructure. Enterprises deploying agent fleets at scale are reporting queue bottlenecks that limit throughput. Reserved compute is no longer a premium feature. It is becoming table stakes.
The AWS Parallel
The structure mirrors Amazon Web Services’ Reserved Instances model, which transformed cloud computing from pay-as-you-go into long-term infrastructure contracts. Kingy AI’s analysis called it “the moment OpenAI stopped behaving like a vendor and started behaving like infrastructure.”
For smaller startups and independent builders, the move raises a familiar concern: enterprises locking up capacity through multi-year contracts could squeeze shared pools, creating access tiers where the largest buyers get priority and everyone else competes for what remains. OpenAI has not addressed how Guaranteed Capacity affects availability for pay-as-you-go API customers.