EDGE
Get a demo
Log In
EDGE Insights

EDGE Insights

icon
Filter
This week:
Last week:
Generative AI Applications

Generative AI Applications

OpenAI: The ChatGPT creator leading large AI commercialization

This insight was first published on 8 February 2023. It has been updated to reflect the launch of GPT-4 and recent developments around ChatGPT.
OpenAI is easily one of the most newsworthy AI companies in contemporary times. Not only does it push the boundaries of what is possible in the field of AI, but the general accessibility of its large AI models has given way to a variety of previously unfathomable commercial applications. While many comparable models are in development and use across Big Tech firms like Google, Meta, and Microsoft, they remain somewhat veiled from the public. This is due to the imperative of ensuring responsible use and safeguarding corporate reputations while prioritizing in-house product development. 
OpenAI’s GPT series, Codex, and DALL.E series models have provoked conversation and generated serious enterprise interest since their release, but it is the company’s recent conversational AI model ChatGPT that has taken the world by storm. The subsequent release of its multimodal model GPT-4 constitutes another key milestone.
This Insight dives into the company behind these sought-after models, starting with an overview of OpenAI's business, its notable AI and commercialization status, its business model and key competitors. The Insight also explores the scope of OpenAI's multi-faceted partnership with Microsoft.
For a broader introduction to the world of large AI models, also check out our Edge Insight: Supercharging AI progress: The possibilities and pitfalls of massive models.

OpenAI: No longer purely non-profit

OpenAI is an AI research and deployment company founded in December 2015 by a team of entrepreneurs, ML experts, research engineers, and scientists. This included Y Combinator alum Sam Altman, Elon Musk, former Stripe CTO Greg Brockman, ML expert Ilya Sutskever, LinkedIn Co-founder Reid Hoffman, and PayPal Co-founder Peter Thiel (among other notable founding members). 
At first, OpenAI was primarily a non-profit AI research company focused on advancing artificial general intelligence (AGI) toward a positive human impact, without the pressure of delivering a financial return. However, in March 2019, OpenAI also created a new “capped-profit” corporation called OpenAI LP (governed by the non-profit parent company) to enable it to raise capital and shoulder significant investments in computation power and talent amidst its pursuit of beneficial AGI. OpenAI LP, in particular, focuses on three core areas: 1) capabilities, which drive advances in AI systems; 2) safety, which looks at aligning those systems with human values; and 3) policy, which ensures relevant governance for AI systems.
The idea behind OpenAI LP is that investors and employees receive a capped ROI, while any returns beyond that are owned by the original OpenAI non-profit entity. Estimates indicate that the capped-profit model limits backers’ returns to 100x their investment and possibly less in the future.
To date, however, OpenAI has only generated modest returns in comparison to the significant investments fueling the development of its AI models. Recent reports indicate that the company expects to make USD 200 million in 2023, a fraction of the ~USD 1 billion invested to date.

OpenAI’s notable models

OpenAI is best known for developing the GPT-4 multimodal model, the GPT-3.5 series of large language models, the fine-tuned coding model known as Codex, and the DALL.E series of multimodal models, which were released during 2018–22. More recently, OpenAI made headlines with the November 2022 release of ChatGPT, a model fine-tuned for conversational applications. Additionally, the company is working on training and open-sourcing an automatic speech recognition system called Whisper. OpenAI is also looking to advance AI alignment research, which explores how AI systems can be made helpful, truthful, and safe. 
For definitions of the different types of large AI models, please first refer to Appendix 1.

1. GPT-4 

GPT-4 is a large multimodal model and the successor to GPT-3.5, a large language model comprising a set of four base models (Ada, Babbage, Curie, and Davinci). GPT-4 marks a departure from the GPT-3.5 series in that it accepts both text and image prompts, producing text outputs (natural language, code, etc.) in response. The model is capable of handling over 25,000 words of text and enables use cases including document search and analysis, long-form content creation, and extended conversations.
Several organizations have developed products atop GPT-4, including Duoloingo, Stripe, Morgan Stanley, and Khan Academy. The model is also leveraged by Microsoft’s search engine Bing and ChatGPT Plus, the paid version of OpenAI’s ChatGPT platform. GPT-4 has also been harnessed by OpenAI for internal functions including programming, support, sales, and content moderation. Meanwhile, the GPT-3.5 model series has been leveraged by over 300 applications including notable content-writing startups like CopyAI, Copysmith, and Flowrite; productivity tools like OnLoop; and AI-generated game creators like Latitude.  
In comparison to its predecessor, GPT-4 is reportedly 82% less likely to respond to requests for disallowed content, making it safer and more aligned. The multimodal model is also 40% more likely to produce factual responses than GPT-3.5. Recent reports indicate that GPT-4 excels in test-taking in particular, faring well on challenging exams like the legal Bar exam where it scored in the 90th percentile without specific training. By contrast, GPT-3.5 scored in the bottom 10% of all Bar exam takers.

2. Codex

Codex is a general-purpose programming model derived from its predecessor GPT-3 and comprising two base models (Cushman and Davinci). The 12-billion parameter model was released in 2021 and was trained and fine-tuned on natural language and billions of lines of source code from publicly available repositories (such as GitHub) to enable translations from natural language to code. The model is most savvy in Python but also proficient in over a dozen programming languages including JavaScript, Go, Perl, PHP, Ruby, Swift, TypeScript, and Shell.
Codex has been harnessed in over 70 applications, including Microsoft's AI pair-programming tool GitHub Copilot. Other key applications for Codex include Pygma (uses Codex to turn Figma designs into different front-end frameworks), Replit (uses Codex to describe, in simple terms, what a selection of code is doing), Warp (uses Codex to run a natural language command to search directly from within the terminal), and Machinet (uses Codex to help Java developers write code to generate intelligent unit test templates).

3. DALL.E and DALL.E 2

DALL.E is a 12-billion parameter multimodal model released in 2021. Like Codex, it is derived from its predecessor GPT-3, but as a multimodal model, it can generate digital imagery from text captions across multiple styles, including photorealistic imagery, paintings, and emoji. Its successor DALL.E 2 is a 3.5 billion parameter, multimodal model released in 2022, with the capability to create more realistic images and art from a text description at 4x the resolution of DALL.E. Both models have applications in image prototyping, design (fashion, interiors, products, etc.), and even commercial projects (like illustrations for children’s books, concept art for games, storyboards for movies, mood boards for design consulting, and more). 
As of November 2022, over 3 million people use DALL.E to generate over 4 million images daily. Notable applications of DALL.E include Microsoft’s new graphic design app Designer, CALA’s fashion and lifestyle operating system, and photo startup Mixtiles.  

4. ChatGPT

ChatGPT is a fine-tuned language model derived from the successor language model GPT-3.5 and trained to interact in a conversational way. Launched in November 2022, the model attracted over 1 million registered users in only five days, making it one of the fastest-growing tech platforms. By January 2023, ChatGPT reportedly had 100 million monthly active users. The model is capable of engaging in human-like dialogue based on a prompt and can answer questions, challenge incorrect assumptions, and reject inappropriate requests. The free version of ChatGPT continues to leverage GPT-3.5, while the paid subscription plan, known as ChatGPT Plus, is powered by GPT-4.
As of March 2023, ChatGPT has been leveraged by a growing list of companies including Microsoft (Bing), Instacart (Ask Instacart), and Shopify (AI assistant for the Shop app). 

5. InstructGPT

InstructGPT is a 1.3 billion-parameter fine-tuned language model derived from the predecessor GPT-3 series. It comprises four base models—Ada, Babbage, Curie, and Davinci. It is trained using reinforcement learning from human feedback—techniques developed through OpenAI’s alignment research—to better follow user intentions and instructions than GPT-3, making its responses comparatively more helpful and truthful as well as less toxic.

Commercial availability

For a fuller discussion of how OpenAI’s earliest commercial successes—GPT-3 and Codex—were leveraged in new and existing products by startups and established companies, respectively, please refer to A tale in the making: Popular commercial applications for large AI.

OpenAI’s other key projects

1. Whisper

Whisper is OpenAI’s automatic speech recognition (ASR) system announced in 2022. The system was trained on ~680,000 hours of multilingual and multitask supervised data from the web to enable English speech recognition with human-level robustness and accuracy. Whisper enables transcription in multiple languages and translation from those languages to English. It is currently available for on-demand access through OpenAI’s API.

2. Alignment research

Alignment research explores how best to align future machine learning systems with human interests. The ethical and social risks of harm from these powerful models are well-documented. As such, this line of inquiry stems from the imperative to train AI models to be helpful and honest
OpenAI takes an empirical and iterative approach to aligning AGI with human values and intent. This includes training AI systems to 1) use human feedback, 2) assist human evaluation, and 3) do alignment research. The InstructGPT model, for instance, is an example of OpenAI’s efforts in alignment-focused fine-tuning of large AI models. 

OpenAI’s business model

OpenAI leverages a pay-per-use business model to provide individuals and enterprises with access to its language and image models via its OpenAI API. For the language models, pricing is assigned per 1,000 tokens (tokens are pieces of words, with 1,000 tokens being around 750 words).
To understand how OpenAI’s pricing has evolved and the differences between the base models, please also see the “Cost of access to GPT-3” section in A tale in the making: Popular commercial applications of large AI.

Costs of OpenAI API access to GPT-4

Costs of OpenAI API access to language models

For its image model DALL.E, which only became available for public use in November 2022, OpenAI charges a price per image across three tiers of resolution.

Costs of OpenAI API access to image models

In addition to its API, OpenAI provides access to its language models and its coding models via Microsoft’s Azure OpenAI Service, which also leverages a pay-as-you-go consumption model. The Azure OpenAI Service sets itself apart by bringing together the OpenAI API with Azure enterprise-level security, compliance, and regional availability. Please see Appendix 2 for a full breakdown of the costs of Azure OpenAI API access.
Notably, the Azure OpenAI Service matches OpenAI’s prices for access to the base language models and offers the same price point regardless of whether users wish to access the base or fine-tuned language models. The only variations are in terms of hosting and training charges, which are priced on an hourly basis. Information on how revenue is shared between Microsoft and OpenAI is currently unavailable.
When it comes to OpenAI’s latest model, ChatGPT, the company has leveraged a freemium model in a surprising departure from its usual pay-per-use approach. Initially, access to ChatGPT was offered free for anyone who signed up via the OpenAI website. However, mounting compute costs and the pressure to turn a profit have since driven OpenAI to monetize the model. In February 2023, OpenAI announced ChatGPT Plus, a USD 20/month plan that reportedly allows access during times of peak demand, faster response times, and priority access to new features. OpenAI also offers API access to its ChatGPT model for USD 0.002 per 1,000 tokens and its audio model Whisper for USD 0.006 per minute.

OpenAI’s competitors

OpenAI's competitive landscape includes companies building AI models either in support of their in-house product development or for third-party licensing and use, with a special focus on developers of like-for-like large AI models (i.e., Google, Microsoft, and Meta). It also includes AI labs like A121 Labs and open-source AI communities like BigScience and EleutherAI.
Additionally, our competitive landscape also captures companies with offerings that constitute a challenge to one or more of OpenAI's models. For example, Stability AI and its multimodal model Stable Diffusion represent an alternative to OpenAI's DALL.E models. 
Among the players excluded are 1) companies that offer AI-based products like chatbots, writing assistants, and document processing platforms but not access to the underlying models as OpenAI does (e.g., TextCortex and Hyperscience); 2) companies purely focused on AGI safety and research (e.g., Anthropic); 3) AGI companies focused on specialized applications (e.g., Hanson Robotics and humanoid robots); 4) companies that do not build or offer access to comparable large AI models (e.g., One AI); and 5) companies with large AI models that are not exclusively focused on English language processing (e.g., Yandex, Huawei, and LG).

OpenAI and key competitors

For a breakdown of the commercialization status of these competitor models, please refer to Appendix 3.

OpenAI and Microsoft: A multi-faceted partnership

OpenAI and Microsoft
Source: SPEEDA Edge based on multiple sources

1. Financial investments

The scope of OpenAI’s partnership with Microsoft has been the subject of intense speculation ever since the tech giant’s USD 1 billion infusion in July 2019 to support OpenAI’s mission to advance beneficial AGI. Initially, the partnership entailed Microsoft becoming OpenAI's exclusive cloud provider and preferred partner for commercializing new AI as well as the two companies jointly developing Azure AI supercomputing technologies. 
Since then, Microsoft’s financial stake in OpenAI has only scaled in line with an expanding partnership, culminating in a reported multiyear, multibillion-dollar investment announced in January 2023. Some reports estimate a USD 10 billion infusion that would value OpenAI at USD 29 billion, with Microsoft gaining a potential 75% stake in OpenAI’s profits until it recoups its investment. After this threshold is reached, it is alleged that Microsoft would reportedly scale back to a 49% stake, with other investors taking another 49% and OpenAI’s non-profit parent receiving 2%. 

2. Technology and infrastructure

Aside from OpenAI’s Azure-hosted API, perhaps the most noteworthy technological development between the two players was the May 2020 launch of Microsoft’s Azure-hosted supercomputer developed exclusively for training OpenAI’s large models. With over 285,000 CPU cores, 10,000 GPUs, and 400 gigabytes per second of network connectivity for each GPU server, the machine reportedly ranks in the top five of the Top 500 supercomputers in the world. Even OpenAI CEO Sam Altman emphasized the importance of Microsoft and the Azure platform for the infrastructure underlying its large AI models.
Microsoft credit tweet
Source: Twitter

3. Commercialization and enterprise access

A critical milestone in this mutually beneficial partnership was Microsoft’s launch of the Azure OpenAI Service in November 2021. This enabled enterprise customers to also tap into OpenAI’s large models while backed by the security, reliability, compliance, and data privacy of the Azure cloud and computing infrastructure. For instance, Al Jazeera Digital leverages the Azure OpenAI Service to support its journalism, including content production tasks like summarization and translation, topic selection, content extraction, and style guide rule application. 
Initially, the Azure OpenAI Service was only available to Microsoft-managed customers and partners to ensure that early collaborations were backed by responsible AI safeguards. However, Microsoft announced the general availability of the service in January 2023, which means that more businesses can now apply for access to OpenAI’s models. 

4. New product development and integrations

In September 2020, Microsoft announced that it would exclusively license GPT-3, enabling the company to leverage OpenAI’s language model for developing and delivering AI solutions for its own customers as well as creating new products altogether. There remain some questions around the exclusivity of this license, but the following years witnessed Microsoft launch new solutions and product integrations powered by OpenAI’s models. In May 2021, for instance, Microsoft unveiled the integration of GPT-3 in its low-code development platform Power Apps. Through the integration, developers can now use basic conversational language to describe programming goals, with the fine-tuned GPT-3 model offering options for transforming these commands into relevant formulae for the Power platform.
In June 2021, Microsoft also announced GitHub Copilot, a brand new AI pair programming tool developed using OpenAI’s Codex. The solution was launched just two years after Microsoft acquired the code-repository service GitHub in a bid to increase its focus on open-source development and expand the reach of its developer tools and services. In October 2022, the company launched Microsoft Designer, a new DALL.E-2 powered graphic design app in Microsoft 365. Microsoft also has plans to integrate the app into Microsoft Edge to enable AI-powered design suggestions to enhance social media posts and other visual content within the convenience of a browser window. 
In February 2023, Microsoft announced a GPT-3.5-powered version of its CRM companion application Viva Sales, enabling users to generate customized email content for a range of scenarios. The company also launched Teams Premium with features powered by GPT-3.5. The latest version includes an intelligent recap feature to automatically generate notes, tasks, and meeting highlights, alongside custom meeting templates and watermarking to protect meeting contents. In March 2023, Microsoft also confirmed that Bing would be powered by GPT-4.

Appendix 1: Large AI model types

Appendix 2: Costs of Azure OpenAI API access to language and coding models

Appendix 3: Commercialization stage of OpenAI competitors

Contact us

Gain access to all industry hubs, market maps, research tools, and more
Get a demo
arrow
menuarrow

By using this site, you agree to allow SPEEDA Edge and our partners to use cookies for analytics and personalization. Visit our privacy policy for more information about our data collection practices.