The most consequential thing OpenAI built before ChatGPT was a spreadsheet. In 2017, the young nonprofit ran the numbers on frontier AI and reached an awkward conclusion: digital intelligence had acquired a very physical appetite. Internal planning contemplated expanding from roughly 600 GPUs to 5,000, with about $12 million in capital expense and another $5 million to $6 million in annual operating costs just for that step. The curve pointed upward. Donations did not.
That realization explains the company more cleanly than any demo. OpenAI began in December 2015 as a nonprofit research institution, surrounded by a $1 billion funding pledge and a mission to ensure artificial general intelligence benefits all of humanity. By 2019 it had created a nonprofit-controlled for-profit subsidiary and accepted a $1 billion investment from Microsoft. In 2025, the operating company became OpenAI Group PBC, a public benefit corporation still controlled by the OpenAI Foundation. In March 2026, OpenAI closed $122 billion in committed capital at an $852 billion post-money valuation.
The first thing that failed was the financing assumption
OpenAI's original model was elegant: remove the pressure to generate a financial return, publish research and prioritize humanity over shareholders. The science did not make that promise impossible. The cost of doing the science did. Large neural networks improved as researchers combined better methods, more data and more computation. An academic-style lab could no longer assume it would keep pace with Google DeepMind and other well-financed rivals using philanthropy alone.
This changed the founders' minds about structure. Talks with co-chair Elon Musk over a for-profit path ended in 2018 after disagreements about control; Musk left. The remaining team created OpenAI LP the next year, with the nonprofit above it. Microsoft supplied money, Azure supercomputing and commercial reach. The arrangement solved the immediate problem, but planted the question OpenAI still carries: if safety requires enormous capital, what happens when the people supplying that capital expect speed, products and returns?
“Compute is the scarcest resource in AI.”Sarah Friar, OpenAI CFO
A free preview found the business
ChatGPT arrived on November 30, 2022, presented as a free research preview. It could answer follow-up questions, challenge a premise and admit mistakes. It could also produce confident nonsense, stumble over simple logic and reflect biases in its training data. The imperfections were obvious. So was the interface. Instead of requiring developers to call an API, ChatGPT gave ordinary people a blank box and permission to ask.
Users supplied the product map. They used it for homework, trip planning, budgets, drafts and coding. Then they brought it to work for customer emails, spreadsheet checks, campaign ideas and meeting preparation. OpenAI followed them with paid consumer plans, ChatGPT Business, Enterprise and Education. Familiarity at home lowered the training cost at the office. By early 2026, OpenAI said ChatGPT had more than 900 million weekly active users, over 50 million consumer subscribers and more than 9 million paying business users.
The company sells the same underlying capability through several doors. Individuals pay subscriptions for more access and features. Teams buy managed workspaces. Enterprises negotiate security, retention, support and deployment terms. Developers pay by usage through the API. In 2026, the company added the OpenAI Deployment Company, meant to put engineers inside organizations and turn promising demos into production workflows. OpenAI increasingly earns not only when somebody chats, but when a model becomes part of how a business operates.
The chatbot is the shop window. The warehouse is enormous.
OpenAI looks like a software company from the browser and an infrastructure company from the balance sheet. Its product surface spans GPT models, ChatGPT, Codex for software work, Sora for video, image generation, speech systems and APIs. Beneath them sits a portfolio of clouds, chips and data centers. Microsoft remains a central partner and Azure the exclusive cloud for stateless OpenAI APIs under the companies' February 2026 statement. OpenAI has also expanded across Oracle, AWS, CoreWeave and Google Cloud, while working with NVIDIA, AMD, Broadcom and others on silicon.
Stargate makes the ambition visible. Announced in January 2025 by OpenAI, SoftBank, Oracle and MGX, the project initially intended to invest $500 billion over four years in U.S. AI infrastructure. OpenAI had operational responsibility; SoftBank had financial responsibility. Oracle later agreed to develop another 4.5 gigawatts of capacity. The practical aim is not a prettier data center. It is enough reliable compute to train frontier models, serve rising demand and lower the delivered cost of each unit of useful intelligence.
A general assistant for learning, writing, images, voice, research and planning. They buy convenience and capability.
Models and tools delivered through APIs. They buy speed to market without training a frontier model.
Managed workspaces, agents, controls and integration. They buy productivity plus governance.
Deployment, localization and infrastructure relationships. They buy capacity, expertise and strategic access.
Its moat is a loop, not a single model
Anthropic, Google, Meta, xAI, Mistral and open-source communities can all produce capable alternatives. Some customers prefer Claude's behavior, Gemini's Google integration, Copilot's Microsoft distribution or open models they can inspect and run themselves. Benchmarks move too quickly for any permanent crown. OpenAI's difference is the bundle: a consumer product with global reach, a developer platform, enterprise sales and deployment, multimodal research, brand recognition and booked access to scarce computation.
Those pieces reinforce one another. ChatGPT teaches the market what a model can do. Developers extend it into thousands of products. Workplace adoption creates demand for controls and connectors. Revenue finances models and capacity. Better models return to every surface at once. The loop is difficult to reproduce because each part is expensive, and because distribution supplies feedback that a research benchmark cannot.
The loop is not invulnerable. It weakens if inference costs stay above what customers will pay, if a rival offers similar capability more cheaply, if outages make agents unreliable, or if copyright, privacy and safety failures erode trust. It also fits poorly where a hospital, government or bank requires on-premises deployment, open weights or tight vendor independence. OpenAI can be the broad platform and still lose specialized workloads.
The clever part was not predicting every use. It was giving millions of people a useful object and watching which habits survived novelty.
What builders can copy, minus the $122 billion
The transferable playbook begins with iterative deployment. OpenAI released ChatGPT with visible limitations because real use would expose strengths, failures and demand faster than a closed lab. A smaller company can do the same if the product is safe enough to release, feedback arrives quickly and the team can improve what users actually repeat. Start with one magical interaction. Measure retention, not applause. Follow the behavior from curiosity into a workflow where the result saves time, makes money or reduces risk.
Second, price along the path of value. OpenAI uses free access for reach, subscriptions for frequent individuals, seats for teams and metered APIs for production workloads. That ladder lets customers deepen commitment without changing vendors. Third, treat deployment as product work. An impressive model is not a finished enterprise system; data access, permissions, evaluations, change management and accountability decide whether it survives Monday morning.
The conditions matter. This approach will not work when the preview is merely a demo, when users cannot verify important outputs, when variable costs swallow gross margin, or when each customer needs a bespoke rebuild. It also fails when capital is booked far ahead of unproven demand. OpenAI can spread infrastructure commitments across hundreds of millions of users and many products. A startup cannot assume the same escape velocity.
The mission now has shareholders for neighbors
OpenAI's 2025 structure tries to turn the central conflict into an alignment mechanism. The OpenAI Foundation controls OpenAI Group PBC and held a 26 percent stake valued around $130 billion when the recapitalization closed. Microsoft held roughly 27 percent; employees and other investors held the remainder. If the operating company gains value, the Foundation gains resources for philanthropy and mission work. Commercial success is designed to feed public benefit.
That is the theory. The test arrives when safety, competitive speed and investor economics point in different directions. OpenAI's brief 2023 firing and reinstatement of Sam Altman showed how quickly formal governance can collide with employee loyalty, partner leverage and operational reality. The episode did not settle the argument. It revealed the forces any answer must withstand.
OpenAI has already changed what people expect from software: not a menu of commands, but a conversation that can draft, see, hear, code and increasingly act. Its next act is less whimsical. It must make that intelligence reliable enough for important work, cheap enough for daily use and governed well enough to deserve its reach. The spreadsheet is still running.