OpenAI on Tuesday rolled out a new enterprise offering called Guaranteed Capacity, a program that lets large customers lock in long-term access to the company’s compute resources in exchange for multi-year spending commitments. The move, reported by CNBC, is part business-model evolution and part signal flare to the public markets ahead of what is expected to be one of the largest technology IPOs in history.
The pitch is simple. Enterprise buyers worried about being rate-limited or deprioritized during demand spikes can now sign one, two, or three-year contracts that guarantee access to a defined slice of OpenAI’s compute, with discounts that scale up with the length of the commitment. Customers can draw down their reserved capacity across OpenAI’s product portfolio, meaning a single commitment covers usage of the API, ChatGPT Enterprise, Codex, Sora, and any new products the company ships during the contract term. For CIOs trying to plan AI budgets that have grown from line items into board-level capital allocation decisions, the predictability matters.
Altman’s Capacity Thesis
Sam Altman framed the offering in classic supply-and-demand terms. “Customers are increasingly asking us for certainty on capacity,” he said in announcing the program. “As models get better, we expect that the world will be capacity-constrained for some time.” The new arrangement, Altman added, helps OpenAI plan ahead, which he described as a “big win-win.”
That language is doing a lot of work. The unspoken subtext is that OpenAI’s most demanding customers, the ones building agentic workflows that consume billions of tokens per day, have been quietly grumbling for months that they cannot get the throughput they paid for during peak hours. Anthropic’s Claude business has used that frustration as a wedge, signing several large enterprise accounts on the strength of more reliable inference latency. Guaranteed Capacity is OpenAI’s structural response to that competitive pressure. By giving large customers a written promise of access, OpenAI converts a soft pain point into a hard contract, and in exchange it gets the multi-year revenue visibility it desperately needs to underwrite the next leg of its build-out.
The discount structure also matters. According to people familiar with the program, the longest commitments, three-year deals, can carry double-digit percentage discounts relative to pay-as-you-go pricing, with the steepest concessions reserved for customers who pre-commit to very large dollar floors. The arrangement effectively rewards customers who help OpenAI lock in revenue ahead of its capital expenditure curve, while creating switching costs that bind enterprise spend to the OpenAI stack for years. From a competitive standpoint, the playbook is straight out of the AWS reserved-instance handbook from the early 2010s, adapted to the era of foundation models.
The Numbers Behind the Move
The strategic context is enormous. OpenAI is valued by private investors at more than $850 billion and is preparing for a potential IPO that could value the company well into thirteen-figure territory. To support that valuation, the company has told investors it now plans to spend roughly $600 billion on compute through 2030. That figure is a significant reduction from the $1.4 trillion projection Altman floated last year, but it remains an extraordinary number, larger than the entire annual capital expenditure of every hyperscaler combined just a few years ago.
The path to spending that much money has shifted in important ways. The original Stargate joint venture, announced at the White House in January 2025 with Larry Ellison and Masayoshi Son, was structured as a roughly $500 billion four-year program to build dedicated AI infrastructure. In practice, OpenAI has pivoted toward a rent-rather-than-build strategy, signing massive multi-year deals with Amazon Web Services, Google Cloud, CoreWeave, Oracle, and Microsoft Azure. The AWS arrangement alone has grown from an initial $38 billion agreement into a $100 billion eight-year commitment, with similar order-of-magnitude commitments across the rest of the cloud stack. We covered the broader pattern in our earlier piece on Anthropic’s compute deal with SpaceX, where the same dynamic of locking in long-term capacity is playing out across the industry.
Guaranteed Capacity is the demand-side counterpart to those supply-side deals. Where OpenAI buys long-term capacity from clouds, it now sells long-term capacity to enterprises. The arbitrage between those two contracts, both in terms of margin and in terms of risk, is the operating model the company appears to be moving toward as it scales.
Revenue, Margins, and the IPO Calendar
OpenAI is targeting roughly $280 billion in revenue by 2030, a figure that would make it one of the largest software companies in the world if achieved. To get there, the company needs not only to grow ChatGPT subscriptions but to convert raw model access into recurring enterprise contracts. Guaranteed Capacity is a direct lever on that conversion. A customer who signs a three-year commitment is no longer a transactional buyer of inference. They are an annuity, and Wall Street values annuities very differently than it values usage-based revenue.
The pricing model also previews the conversation OpenAI will need to have with public-market investors. Underwriters and institutional buyers want to see net dollar retention, contract length, and committed revenue. A company that can show a growing book of multi-year guaranteed commitments alongside its API and ChatGPT consumer numbers is much easier to model. By the time OpenAI files an S-1, the percentage of revenue covered by Guaranteed Capacity contracts will be one of the headline metrics the market scrutinizes, alongside gross margin per token, model improvement cadence, and the trajectory of training versus inference spend.
Investors are watching all of this with mixed feelings. The bull case is that OpenAI is building the kind of moat that justifies its valuation, locking in customer spend ahead of competition from Anthropic, Google, xAI, and a wave of Chinese model labs. The bear case is that the company is making outsized capital commitments against revenue that has not yet fully materialized, and that any slowdown in enterprise AI adoption could leave it with stranded capacity and rich fixed-cost obligations. Our analysis of big tech AI spending and investor payouts walked through the same tension across hyperscalers.
What This Means for Customers
For enterprise buyers, Guaranteed Capacity is a useful tool, but it is not without trade-offs. The biggest benefit is predictability, both for engineering teams that need to plan capacity for agentic workloads and for finance teams trying to forecast AI spend that has, in many companies, become the largest non-labor line item in the technology budget. Locking in three years of access at a fixed discount is, in many ways, just rational corporate hedging.
The trade-off is flexibility. The AI model landscape continues to move at a breathtaking pace, and committing to multi-year OpenAI spend may look different in eighteen months if Anthropic, Google’s Gemini line, or an open-source competitor delivers a step-change in capability. That risk is partly mitigated by drawdown rights that span OpenAI’s full product line, which means customers benefit from new product launches without renegotiating, but the structural commitment to OpenAI as a vendor remains.
Buyers also need to scrutinize what guaranteed access actually means under the contract. Reserved capacity in cloud computing comes in many flavors, from genuinely partitioned hardware to priority queueing during congestion. The precise mechanics of OpenAI’s guarantee, the SLAs around latency and availability, and the remedies if those SLAs are missed will determine how meaningful the offering is in practice. Initial reporting suggests the program offers genuine priority access during peak periods rather than physically reserved hardware, which is sensible at the scale OpenAI operates but is a different thing than dedicated infrastructure.
The Competitive Picture
OpenAI is not the only AI provider grappling with capacity. Anthropic has signaled to enterprise customers that it is willing to negotiate multi-year deals for guaranteed access to Claude, and Google has been quietly pitching similar arrangements through Vertex AI. Microsoft, which sits in the awkward position of being both an OpenAI investor and a competitor through its own Azure AI offerings and its evolving relationship with the Stargate program, will be watching closely to see how Guaranteed Capacity is priced relative to its own enterprise contracts. Our coverage of the best AI stocks to buy maps the broader competitive landscape, and the leaderboard continues to shift roughly every quarter.
The chip layer matters too. OpenAI’s ability to deliver on multi-year guarantees depends on the supply of Nvidia accelerators, the maturity of in-house silicon, and the willingness of cloud partners to dedicate the capacity. Nvidia’s earnings, due this week, will give investors a fresh read on whether the AI capex cycle is still accelerating or has begun to plateau. Custom AI chip ventures from challengers like Cerebras, whose IPO we covered in our piece on the $48 billion AI chip race, are part of the same supply-side puzzle.
What to Watch Next
Three things will tell us how Guaranteed Capacity actually performs. First, the rate at which OpenAI converts existing enterprise customers from usage-based contracts to multi-year commitments will indicate whether the offering is creating new demand or simply repackaging existing spend. Second, the discount levels and contract terms that leak into the trade press over the next quarter will tell us how aggressive OpenAI is willing to be in trading present pricing for future visibility. Third, and most importantly, the size and composition of the committed revenue book will be a leading indicator for the timing and structure of the IPO. If the program lands hard, expect an accelerated filing timeline.
For now, the takeaway is straightforward. OpenAI has formalized what was already happening in private contracts and made it a productized offering. The company is telling the market it expects compute to be the scarce factor for years, that it is willing to underwrite that scarcity with multi-year customer commitments, and that the business model that supports an eventual trillion-dollar IPO will look more like infrastructure software with annuity revenue than the consumer-facing chatbot business that made it famous.
Frequently Asked Questions
What is OpenAI's Guaranteed Capacity offering?
Guaranteed Capacity is a new enterprise program that lets customers commit to one, two, or three years of OpenAI spend in exchange for guaranteed access to compute and discounts that scale with the length of the commitment. Customers can draw down their reserved capacity across OpenAI’s full portfolio of products, including the API, ChatGPT Enterprise, Codex, Sora, and any new products OpenAI releases during the contract term.
Why is OpenAI doing this now?
OpenAI is preparing for a potential public offering and needs to show institutional investors a book of committed, predictable revenue rather than purely transactional API usage. Sam Altman has also said that he expects compute to remain capacity-constrained for years as model capabilities improve and demand from agentic workflows scales, which makes long-term commitments attractive to both sides of the contract.
How much is OpenAI spending on compute?
OpenAI has told investors it expects to spend roughly $600 billion on compute through 2030, a reduction from earlier $1.4 trillion projections but still an extraordinary figure. That spend is split across multi-year deals with Amazon Web Services, Google Cloud, CoreWeave, Oracle, and Microsoft Azure, alongside the original Stargate program. The AWS commitment alone has grown to $100 billion over eight years.
How does this affect competitors like Anthropic and Google?
Anthropic and Google offer broadly similar enterprise guarantees through bespoke contracts and through Vertex AI respectively, but OpenAI’s productized version of the program raises the bar for the industry. Expect both Anthropic and Google to formalize their own reserved-capacity tiers in coming months, and expect enterprise buyers to use OpenAI’s pricing as a negotiating reference point in deals with all three vendors.
What are the risks for enterprise buyers signing multi-year deals?
The main risk is flexibility. The AI model landscape continues to move quickly, and a buyer that locks in three years of OpenAI spend may face a competitor with a meaningfully better model partway through the contract. That risk is partially offset by drawdown rights that span all OpenAI products, so customers can shift consumption to new launches without renegotiating, but the vendor lock remains. Buyers also need to scrutinize the precise SLAs around latency and availability that define what guaranteed actually means.
How does Guaranteed Capacity relate to OpenAI's IPO?
Multi-year committed revenue is exactly the kind of disclosure that institutional investors want to see when evaluating a public offering. By the time OpenAI files an S-1, the percentage of revenue covered by Guaranteed Capacity contracts will be a headline metric alongside gross margin per token, model release cadence, and the ratio of training to inference spend. A strong early uptake of the program could accelerate the filing timeline.
