In 2020, building a product around a large language model required scarce model-building and serving expertise, compute and experimentation. Most buyers wanted a useful capability inside an application rather than responsibility for training a frontier model. A hosted interface could reduce that infrastructure burden, but the buyer still needed to discover a task the model could perform reliably and build the surrounding product. The available technology remained uncertain: a fluent answer was not necessarily correct, and generality expanded both possible uses and possible failure modes. An application-facing service therefore had to establish useful behavior as well as technical access. OA5
OpenAI
An API made models usable by developers; ChatGPT gave OpenAI a direct relationship with the people using them.
Arena: Market Conditions Before OpenAI
What shaped the outcome
Step 1 of 4 · 2020–22
Remove the model-building task while leaving room for other applications
The 2020 API offered a general text interface that developers could integrate into their own products. OpenAI handled the model service and described commercial access as a way to support further development. The later Microsoft partnership supplied additional infrastructure and routes for model use. OA5 OA8
The customer did not need to become a frontier-model laboratory before testing an application. That changed who could try to build with language capability and made one underlying model investment available across many tasks. It also left meaningful work outside the API: prompt design, evaluation, data access and the surrounding user experience. A general-purpose interface therefore creates an option for developers, not proof that every use case is economically valuable. The commercial opportunity depends on repeatedly supplying capability they prefer to alternative models or internal development.
Rivals Internal model development and competing hosted models are alternatives. Compare delivered performance, integration effort, latency and total cost for a specific task.
Step 2 of 4 · 2022–23
Put the model in a product people could try without building anything
ChatGPT launched in November 2022 as a conversational research preview. The dialogue format supported follow-up requests and correction, while the launch explicitly acknowledged unreliable answers. By November 2025, OpenAI reported more than 800 million weekly users of ChatGPT. These are dated product and usage observations, not a claim that all use was paid or productive. OA6 OA1
The interface removed the need to write an application before exploring what a model could do. A user could ask, inspect the result and revise the request within the same exchange. That made capabilities visible through use and gave OpenAI a direct customer relationship it did not have when another developer owned the interface. Distribution and learning about use could now happen inside its own product. The recurring utility still depended on model quality and the task: conversational fluency helped access, but did not eliminate errors or the need to check consequential outputs.
Rivals An API console serves a developer; search and other assistants serve end users. The competitive test is whether dialogue helps complete the user's task more effectively than those alternatives.
Step 3 of 4 · 2023–25
Make successful individual use acceptable to an organization
ChatGPT Enterprise added organizational controls and commitments about business data. OpenAI later reported five million paying business users in August 2025 and more than one million business customers in November, using different units and product scopes. A 2026 a16z survey also found substantial enterprise use, while explicitly reporting gains by competing model suppliers. OA7 OA2 OA1 OA9
A business buyer needs more than an impressive answer. Administration, data handling and reliable access help turn individual experimentation into an approved workplace purchase. OpenAI could pursue that buyer directly while continuing to sell APIs to developers. The distinct measures matter: a business can have many seats, an API customer need not buy ChatGPT seats, and a survey respondent may use multiple providers. Together they support paid adoption, but they cannot be added into a single customer total or used to establish an uncontested enterprise lead.
Rivals Anthropic, Google and Microsoft offer overlapping organizational routes. The choice depends on the task, existing cloud relationship and organizational controls as well as model quality.
Step 4 of 4 · 2023–25
Audience creates an opening for a platform, with infrastructure dependencies still attached
Thompson's 2025 analyses contrast OpenAI's direct consumer relationship with the distribution and infrastructure owned by major technology companies. After DevDay, he interprets applications inside ChatGPT as an attempt to turn that relationship into a wider platform. The Microsoft partnership is a concrete infrastructure relationship; the lasting platform outcome remains a hypothesis. OA8 OA3 OA4
An audience can make integration attractive to application suppliers, potentially improving the product and giving users more reasons to stay. But the loop is not established merely by announcing integrations. Users must complete useful tasks, partners must receive value, and competitors must face a meaningful barrier to offering similar access. Meanwhile, model service requires capital, compute and supplier relationships. The same company can have strong demand and unresolved bargaining power or economics. The application ecosystem was still being built; its announced integrations had yet to establish durable dependence on OpenAI.
Rivals Google combines models with its own infrastructure and existing distribution; Microsoft can place AI inside established enterprise products. OpenAI's direct audience is valuable but does not remove those alternatives.
Key dates
- 2020-06-11Launch the commercial API
- 2022-11-30Launch ChatGPT
- 2023-01-23Extend the Microsoft partnership
- 2023-08-28Offer ChatGPT Enterprise
- 2025-08-07Report paid workplace use
- 2025-10Expand the application-platform ambition
- 2025-11-05Report one million business customers
Sources
Oldest first.
- OA5 OpenAI API | OpenAI. Primary company disclosure
- OA6 Introducing ChatGPT | OpenAI. Primary company disclosure
- OA8 OpenAI and Microsoft extend partnership | OpenAI. Primary company disclosure
- OA7 Introducing ChatGPT Enterprise | OpenAI. Primary company disclosure
- OA3 Checking In on AI and the Big Five - Stratechery by Ben Thompson. Outside account
- OA2 GPT-5 and the new era of work | OpenAI. Primary company disclosure
- OA4 OpenAI’s Windows Play - Stratechery by Ben Thompson. Outside account
- OA1 1 million business customers: the fastest-growing business platform in history | OpenAI. Primary company disclosure
- OA9 Leaders, gainers and unexpected winners in the Enterprise AI arms race | Andreessen Horowitz. Outside account