How OpenAI makes money: The most expensive free product ever built

OpenAI gives away ChatGPT to 900 million weekly active users for free. It fields 2.5 billion queries per day and processes 15 billion tokens per minute. And yet, the company is on track for $25 billion to $29 billion in annualized revenue in 2026 — while simultaneously losing $14 billion a year.

That paradox sits at the center of a growth trajectory few software companies have matched. In 2022, OpenAI generated just $28 million in revenue. By December 2023, that figure had climbed to roughly $1.6 billion annualized. By mid-2025, it hit $10 billion. By early 2026, it crossed $24 billion. That's a nearly 900x increase in under four years.

In this breakdown, we'll unpack the full mechanics of OpenAI's revenue model — consumer subscriptions, enterprise products, API pricing, advertising, and creative tools — and explain why profitability remains years away despite the staggering top line.

OpenAI was founded in late 2015 as a nonprofit with a singular mission: ensure that artificial general intelligence (AGI) benefits all of humanity. 

The nonprofit structure, however, couldn't attract the billions required for large-scale compute. In 2019, the organization pivoted, creating a capped-profit subsidiary designed to bring in outside capital while preserving mission alignment. That restructuring culminated in October 2025, when OpenAI transformed into a Public Benefit Corporation (PBC) with a two-part structure: the OpenAI Foundation and the OpenAI Group.

The product stack reflects the breadth of that ambition. On the consumer side, ChatGPT is available in four tiers: Free, Go ($8/month for emerging markets), Plus ($20/month), and Pro ($200/month). For businesses, there are Team, Business, Enterprise, and Education plans. Developers access models through the API platform, which supports token-based billing across a wide range of models. Specialized products include DALL-E for image generation, Codex for code, and Operator for agentic tasks. OpenAI's video generation model, Sora, was discontinued in April 2026 after a brief commercial run.

The user journey follows a deliberate funnel. The free tier serves as a massive acquisition channel, pulling in 900 million weekly users. Feature gating — limited model access, usage caps, restricted advanced tools — pushes a fraction of those users toward paid plans. From there, business workspaces and the developer API capture higher-value segments.

OpenAI's differentiators are rooted in scale and timing. It holds first-mover advantage in generative AI, serves 92% of the Fortune 500, and processes 2.5 billion queries daily. The company's strategy resembles a superapp play: bundling chat, code, search, image generation, and agentic workflows into a single platform. Though the discontinuation of Sora suggests that not every modality has proven viable within that bundle.

Underneath all of this sits what might be called an "Intelligence as a Service" model. OpenAI operates a compute-revenue flywheel: a 9.5x increase in compute capacity has corresponded to a roughly 10x increase in revenue. The infrastructure runs across multiple clouds — Azure as the primary partner, plus AWS, Oracle, and Google Cloud — giving the company flexibility to scale inference capacity as demand grows.

The revenue trajectory tells the story more clearly than any single number. OpenAI went from $28 million in 2022 to $25 billion annualized by Q1 2026. That growth is distributed across five distinct revenue streams, each with different economics and growth profiles.

Consumer subscriptions account for roughly 60% of OpenAI's revenue, making them the financial backbone of the business. The tier structure is straightforward: ChatGPT Go at $8/month targets emerging markets, ChatGPT Plus at $20/month is the core offering with approximately 18 million subscribers, and ChatGPT Pro at $200/month serves power users willing to pay for maximum model access and capacity. Pro previously included unlimited Sora video generation.

By February 2026, OpenAI had crossed 50 million consumer subscribers. The conversion rate from free to paid sits at about 7% — modest on the surface, but significant given the sheer size of the free user base. Quarterly retention runs at 89%, dropping to 74% at nine months, suggesting solid but not exceptional stickiness.

A lower-bound estimate puts consumer subscription revenue at $4.8 billion annualized or higher. The math works because OpenAI has built one of the largest freemium funnels in software history: 900 million weekly users flowing into a paid tier with strong average revenue per user across the Go, Plus, and Pro plans.

Enterprise is OpenAI's fastest-growing segment, now accounting for more than 40% of revenue as of Q1 2026. The company reports 9 million paying business users across Team, Business, and Enterprise tiers. Business plans start at $25 per user per month, while Enterprise pricing is custom and typically higher.

A floor estimate based on 9 million users at $20/month yields roughly $2.16 billion annualized. But the top-down math — 40% of $24 billion — suggests the real figure is closer to $9.6 billion to $10 billion or more, implying significant revenue from higher-priced Enterprise contracts.

Adoption among large organizations is striking: 92% of the Fortune 500 and 36.5% of U.S. businesses use OpenAI products. The most common enterprise use cases are data analysis (60%), content generation (51%), and personalization (42%). The company expects enterprise revenue to reach parity with consumer subscriptions by late 2026, a milestone that would mark a fundamental shift in the business mix.

The API contributes approximately 15% of revenue and serves a different customer profile: developers and companies building AI-powered applications on top of OpenAI's models. Pricing follows a per-token billing model with a 150x price range across model tiers.

At the top end, GPT-5.5 Pro costs $30 per million input tokens and $180 per million output tokens. At the bottom, lighter models like Nano run at $0.20/$1.25. OpenAI offers aggressive volume discounts: Cached Input pricing provides up to a 90% discount on repeated prompts, and the Batch API delivers a 50% discount for non-real-time workloads.

The platform processes 15 billion tokens per minute, a figure that reflects both the scale of developer adoption and the compute intensity of the underlying infrastructure. Web search calls are billed separately at $10 per 1,000 calls. The API is a strategically important segment because it embeds OpenAI's models into third-party products, creating switching costs and long-term revenue stickiness.

OpenAI launched advertising in February 2026, initially limited to users on the Free and Go tiers. The ad formats are contextual — sidebar placements, in-conversation suggestions, and post-response units — rather than the behavioral targeting model used by Google or Meta.

Early CPMs are running at approximately $60, a premium rate that reflects the high-intent nature of ChatGPT queries. Internal projections target $2.5 billion in ad revenue for 2026, with advertising and commerce expected to grow to roughly 20% of total revenue over time.

The advertising model is still nascent, but the underlying logic is compelling. With 94.5% of ChatGPT users on the free tier, OpenAI has an enormous audience that it currently monetizes at zero. Even modest per-user ad revenue — estimates suggest around $5 per month per free user — could generate $10 billion or more annually if applied across the full free user base.

OpenAI's creative tools generate revenue through both direct pricing and strategic partnerships. The DALL-E API charges between $0.04 and $0.12 per image depending on resolution and model. Sora, the video generation model, briefly contributed revenue through Pro subscriptions and enterprise licensing before OpenAI discontinued it in April 2026 — a notable retreat that underscored the difficulty of monetizing compute-intensive generative video at scale.

The most notable deal in this category is the $1 billion partnership with Disney, signed in December 2025, which provided Disney with custom model access for content creation. The future of that arrangement post-Sora remains unclear.

Government sales represent another emerging channel: a March 2026 deal routes OpenAI products to defense and government agencies through Amazon's cloud infrastructure.

These revenue streams are small relative to subscriptions and enterprise, but they signal OpenAI's ambition to monetize beyond text-based AI — expanding into image generation, sector-specific licensing, and government contracts.

OpenAI's revenue growth is dramatic, but it remains outpaced by its costs. The company projects a $14 billion net loss in 2026 and doesn't expect to reach cash-flow positive status until 2029. The gap between revenue and profitability reveals the true economics of running a frontier AI company.

Inference costs — the expense of running models for every user query — are OpenAI's largest and fastest-growing cost center. In 2025, inference spending reached approximately $8.4 billion. For 2026, that figure is projected to climb to $14.1 billion, driven by growing user demand and the computational intensity of newer models.

The scale of infrastructure investment is staggering. Stargate, the joint venture with SoftBank and others, is expected to deploy between $100 billion and $500 billion over four years to build AI-focused data centers. A partnership with NVIDIA involves 10 gigawatts of power capacity and millions of GPUs, with potential investment reaching $100 billion. OpenAI's compute footprint has grown from 0.2 gigawatts in 2023 to 0.6 GW in 2024 to 1.9 GW in 2025.

Cash burn tells the story plainly: roughly $9 billion in 2025 and a projected $17 billion in 2026. The fact that 94.5% of users are on the free tier means the vast majority of inference costs are subsidized by paying customers and investors.

Microsoft, OpenAI's largest investor and cloud partner, receives 20% of OpenAI's revenue through 2030. At current revenue levels, that translates to payments of $13 billion or more across 2026 and 2027. OpenAI has also committed to spending $250 billion on Azure infrastructure by 2032, locking it into a long-term financial relationship with Microsoft even as it diversifies to other cloud providers.

These obligations are a direct consequence of the capital-intensive partnerships that enabled OpenAI's rapid scaling. They are substantial, but they decline over time as revenue grows and the Microsoft revenue-share cap takes effect.

Training frontier AI models costs billions. OpenAI's 2025 training budget exceeds $25 billion, and projections suggest cumulative training spend could reach $121 billion by 2028. In 2024 alone, the company reported losses of approximately $5 billion, much of it attributable to research and training costs.

Unlike traditional R&D, OpenAI's research spending is fused with its product roadmap. Every model improvement — from GPT-5.4 to GPT-5.5 — directly impacts the performance of consumer and enterprise products, API throughput, and competitive positioning. The research function is not a cost center in the traditional sense; it is the engine that drives the entire product line.

OpenAI employed roughly 3,000 people in 2025 and is planning to scale to approximately 8,000 by the end of 2026. Total compensation already exceeds $2.5 billion per year, reflecting the premium required to attract and retain top AI researchers and engineers in a brutally competitive talent market.

Retention is a visible challenge. OpenAI engineers are reportedly 8x more likely to leave for Anthropic than to stay, a metric that underscores the intensity of the talent war in frontier AI. Compensation costs will continue to rise as the company scales headcount and competes with deep-pocketed rivals for a limited pool of specialized talent.

Cumulative losses from 2023 through 2028 are expected to reach $44 billion. Some of the most dramatic cost items are one-off in nature: Sora, the now-discontinued video model, reportedly cost between $1 million and $15 million per day to operate. Total fundraising across all rounds has reached $140 billion, providing the runway to sustain these losses while the business scales toward profitability.

OpenAI's competitive landscape extends well beyond chatbots. The real battle plays out across the same surfaces where OpenAI generates revenue: consumer subscriptions, enterprise copilots, code generation, API services, and developer ecosystems.

Anthropic is no longer just gaining ground, but it has overtaken OpenAI in revenue. As of April 2026, Anthropic reached a $30 billion annualized run rate, surpassing OpenAI's estimated $24–$25 billion. That trajectory is striking: Anthropic was at roughly $1 billion in early 2025, meaning it scaled 30x in little over a year. The surge has been driven primarily by enterprise adoption and the breakout success of Claude Code, which commands 40% to 54% of the code generation market versus OpenAI's 21%.

In enterprise AI spending more broadly, Anthropic now captures 40% of the market compared to OpenAI's 27% — a dramatic reversal from earlier splits. Claude Pro is priced at $20/month, matching ChatGPT Plus, while Sonnet 4.6 is competitively positioned at $3/$15 per million tokens on the API side. The company is expected to reach cash-flow positive by 2027 — two years ahead of OpenAI — in part because its cost structure is leaner and its enterprise focus is more concentrated.

Google brings a level of vertical integration that no other competitor can match: it builds its own models, designs its own chips (TPUs), operates its own data centers, runs a major cloud platform, and controls a productivity suite used by 3 billion Workspace users. Google accounts for roughly 21% of enterprise LLM spending.

Gemini 3.1 is now benchmarked as tied with GPT-5.4, eliminating the model-quality gap that once gave OpenAI a clear edge. Google's 2026 capital expenditure for AI infrastructure is projected at $175 billion to $185 billion, dwarfing OpenAI's spending. The risk for OpenAI is that Google can subsidize AI losses across its profitable ad business indefinitely, while OpenAI must rely on external funding and eventual profitability.

Meta's Llama 4 represents a structural threat to OpenAI's API revenue. As an open-weight model with a 10-million-token context window, Llama 4 gives developers a free alternative to OpenAI's paid API. The broader open-source ecosystem is accelerating commoditization of intelligence, putting downward pressure on OpenAI's API pricing and forcing repeated price cuts to remain competitive.

The long-term risk is clear: if open-source models reach "good enough" quality for most use cases, OpenAI's API pricing power erodes significantly. The company's response has been to differentiate on model quality, tooling, and enterprise features — but the gap between open and closed models continues to narrow.

Several smaller players are carving out positions in specific segments. xAI offers Grok with aggressive pricing — $2/$6 per million tokens compared to OpenAI's $5/$30 for comparable models — leveraging data from the X platform.

DeepSeek has produced a model competitive with GPT-5 for just $5.3 million in R&D spending, a fraction of OpenAI's $5.7 billion in the first half of 2025 alone.

Perplexity holds 8% of the chatbot market and 5.5% of AI search, positioned as a search-first alternative to ChatGPT.

These competitors individually pose limited threat, but collectively they fragment the market and put pressure on OpenAI's pricing, distribution, and differentiation.

OpenAI is widely expected to pursue an IPO in the second half of 2026, targeting a valuation of $1 trillion and raising approximately $60 billion in new funding. The revenue targets are ambitious: $29.4 billion for 2026, $125 billion by 2029, and $213 billion by 2030. Cash-flow positive is projected for 2029, when the company expects to generate roughly $2 billion in positive cash flow.

The strategic vision is a pivot from chatbot to "Agentic Superapp" — a platform where AI agents handle complex, multi-step tasks across work, commerce, and daily life. This aligns with product moves already underway: Codex for autonomous coding, Operator for task execution, and deeper enterprise integrations that embed OpenAI into business workflows.

Several key milestones will determine whether OpenAI can deliver on this vision:

The central question for OpenAI is whether the compute-revenue flywheel can outrun two forces simultaneously: intensifying competition and the escalating costs of building and maintaining frontier AI.

The company has demonstrated a remarkable ability to scale revenue, but it's no longer the industry's top line leader. What it hasn't yet demonstrated is the ability to turn that revenue into profit. The next three years will determine whether OpenAI becomes the defining technology company of the AI era — or a cautionary tale about the economics of building frontier AI.