Artificial intelligence (AI), whether it’s working on a school paper or solving complex code problems, is taking over the world by storm.
OpenAI, with the release of the image generation model Dall-E and its text bot ChatGPT, has been at the literal forefront of that craze.
Given its complex history, corporate set-up, and AI’s material impact on the future of humanity, many observers began to wonder who’s actually the driving force behind OpenAI and its potent AI tools.
In summary, the for-profit company behind OpenAI is currently owned by Microsoft (49 percent), existing shareholders (49 percent), and the OpenAI non-profit foundation. The latter will remain independent in perpetuity.
With that being said, let’s take a closer look at how it all started, the various shifts in business model strategy OpenAI underwent, and how Microsoft came into the picture.
It All Started with a Noble Idea
While OpenAI is projected to generate billions of dollars in revenue in the near future, those lofty financial ambitions have actually been a product of the recent past.
In fact, when the company was formally unveiled on December 11th, 2015, it did not even think about making money.
“OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return,” the announcement read.
However, what was even more head-turning was the cast of absolute rockstars that OpenAI managed to assemble along the way.
The whole endeavor began just six months prior in a private room at Silicon Valley’s Rosewood Hotel, which is situated right on Sand Hill Road, the world’s VC epicenter, in Menlo Park.
Tesla and SpaceX founder Elon Musk was having dinner with Ilya Sutskever, then a research scientist at Google Brain, to discuss the impact AI could have not only on Musk’s companies but society as a whole.
The dinner itself was brokered by none other than Sam Altman, a long-time friend of Musk and the president of the startup accelerator Y Combinator.
Both Altman and Musk, in the years prior, had repeatedly warned that artificial general intelligence (AGI), when left unchecked, could lead to the demise of humanity. Furthermore, what concerned them was that almost all world-class AI talent was being scooped up by big tech firms like Facebook or Google, which would leave the fate of humans in the hands of a selected few.
Together, they managed to recruit some of the world’s leading AI scientists as founding members, including Trevor Blackwell, Vicki Cheung, Andrej Karpathy, Durk Kingma, John Schulman, Pamela Vagata, and Wojciech Zaremba.
Those research scientists would then be managed by Greg Brockman, formerly the CTO at Stripe, who would hold the same role at OpenAI. Meanwhile, Altman and Musk would serve as co-chairs of the organization.
Now, this assortment of A-players would be newsworthy in and of itself. But there was more to the story. First and foremost, OpenAI was not launched as a profit-seeking entity but as a non-profit, that put safe AI development above anything else.
Second, that non-profit organization would be supported by an eye-popping $1 billion in donations coming not only from Altman’s and Musk’s pockets but also Reid Hoffman, Jessica Livingston, Peter Thiel, Amazon Web Services (AWS), Infosys, and YC Research.
Interestingly, OpenAI wrapped up its launch announcement by stating that its “funders have committed $1 billion, although we expect to only spend a tiny fraction of this in the next few years.”
This, as you will see in the coming chapters, couldn’t have been further from the truth…
Getting the Word Out
With the capital and team in place, it was time to get to work. OpenAI soon began to release new papers multiple times per month while launching open-source tools such as Gym.
From the get-go, OpenAI specialized in one particular branch of AI, namely (deep) reinforcement learning.
Traditionally, AI products such as chatbots were developed using deterministic models, which would require a predefined set of parameters to occur in order to trigger the next point of action. As a result, AI-based tools would often break when deviating from that predetermined course of action.
Similarly, the likes of Google and others trained their image recognition models by inputting millions of the same pictures, thus requiring large amounts of human-labeled data.
Reinforcement learning flips that concept somewhat on its head by forcing the machine to learn on its own. Instead of letting the algorithm know what a cat is (by feeding it labeled cat pictures), you train it by letting it fail over and over again until it completes the task with a certain confidence level (e.g., identifying a cat).
A great visual representation of what this looks like in practice can be seen in one of the early videos that the team released.
Now, a few circles and lines moving toward each other may not be worth reporting on. What is, though, are our natural fears of being displaced by machines.
One of the most prominent examples of AI’s ever-increasing potency became Google’s AlphaGo, which handsomely beat the world’s then-best Go player Ke Jie. Games like Go, due to their inherently complex nature that can span trillions of possible outcomes, were often seen as one of the last resorts where humans reign superior over AI.
However, thanks to reinforcement learning, which is the method AlphaGo was trained on, this superiority quickly became a thing of the past. OpenAI, in an effort to raise awareness about its research, didn’t shy away from similar challenges.
Just three months after the widely covered AlphaGo win, OpenAI had its models compete against the world’s best Dota 2 players, which were all beaten in 1v1 battles. Interestingly, a year later, in August 2018, its OpenAI Five version lost a team match against five other Dota players.
This was grounded in the simple fact that multiplayer games add on a level of complexity that the AI, at least at the time, wasn’t able to deal with just yet.
Unfortunately, being defeated by humans wasn’t the only defeat that OpenAI faced.
A Key Figure Departs
On February 20th, 2018, OpenAI published a blog post, stating that “Elon Musk will depart the OpenAI Board but will continue to donate and advise the organization. As Tesla continues to become more focused on AI, this will eliminate a potential future conflict for Elon.”
The latter part of that statement included the reasoning for Musk’s departure, namely a conflict of interest between OpenAI and Tesla. But how exactly could Tesla be an issue for OpenAI?
First and foremost, there is a history between the two firms that does not involve Musk. Back in June 2017, seven months before Musk’s departure, Tesla announced the hiring of Andrej Karpathy, one of OpenAI’s founding members and a pioneer in the field of neural networks.
Karpathy would go on to lead Tesla’s autonomous driving division, which is now considered to be one of the leading ones across the entire AI industry. And as Tesla continued to double down on self-driving, its need for hiring additional talent would only grow.
Musk, having direct access to OpenAI’s engineers, could have potentially taken even more scientists with him to Tesla, especially considering how attractive the automaker’s stock options were due to Tesla’s rising share price.
For all that is known, Musk and the rest of the OpenAI team never had any fallout that led to bad blood between the two parties. In fact, Musk continued to stay on as an advisor to OpenAI’s team.
However, its founding member being gone wasn’t the only change OpenAI went through.
The Literal Business Transformer
In June 2017, scientists at Google Brain released a paper called Attention Is All You Need where they introduced Transformers, a novel type of neural architecture vastly superior in almost every regard.
Before I mislead you with a wrong description of what Transformers are, I’ll let the kind folks at Nvidia introduce the concept to you:
“A transformer model is a neural network that learns context and thus meaning by tracking relationships in sequential data like the words in this sentence.
Transformer models apply an evolving set of mathematical techniques, called attention or self-attention, to detect subtle ways even distant data elements in a series influence and depend on each other.”
And since Transformers are part of the reinforcement learning branch, they don’t require large amounts of labeled data since they rely on finding patterns themselves.
Additionally, Transformers rely on parallel processing, which allows them to be trained substantially faster at a fraction of the cost of previous models. For OpenAI, which wasn’t generating any revenue back then while paying its top researchers up to $2 million a year, the technology proved to be an absolute game changer.
OpenAI quickly started to utilize Transformers within its own language models and, in February 2019, released what it dubbed Generative Pre-trained Transformer-2, or GPT-2. The model, which was trained on 40GB worth of text from sources like Reddit, would be able to generate text based on a few prompts.
Meanwhile, the results of the tests that researchers ran were allegedly so disconcerting that OpenAI decided to only release a reduced version of GPT-2.
Interestingly, this became one of the first examples of OpenAI’s viral growth strategy in which it uses the media to cause a frenzy about the (adverse) potential of its various technologies. Point in case: OpenAI only gave journalists access to GPT-2 while the general public was prohibited from accessing it – at least for the time being.
With the benefit of hindsight, it thus wasn’t that much of a surprise when OpenAI, a month (03/2019) after the GPT-2 craze, announced that it would shift from being a non-profit to a limited partnership (LP).
The change of its corporate structure would soon attract a tech giant that would propel OpenAI to a completely different level.
A Partnership Made in Heaven (or rather the Cloud)
When OpenAI announced the change toward a for-profit LP structure, it not only got attention due to the shift itself but also because of the unique setup it created along the way.
More precisely, OpenAI’s profit-driven unit, which is separately overseen by the non-profit foundation, would cap the returns its investors could achieve.
Anyone that would invest in OpenAI could only cash out 100x their initial investment, with excess profits going straight back to OpenAI’s foundation.
“There is no way of staying at the cutting edge of AI research, let alone building AGI, without us massively increasing our compute investment,” said co-founder Sutskever in a subreddit just hours after the release.
Furthermore, the LP structure enabled OpenAI, apart from easing fundraising efforts, to compensate employees not just in cash but stock options. As a result, it would be able to offer more competitive salary packages to keep up with the likes of Apple, Facebook, and so forth.
Meanwhile, in order to grow OpenAI into an actual business, founder Sam Altman departed from Y Combinator and became the firm’s CEO. Its board would then be comprised of other founding members, including Brockman and Sutskever as well as previous donors like Reid Hoffman.
And one company, in particular, jumped at the chance of becoming OpenAI’s first major backer. On the 22nd of July, 2019, Microsoft and OpenAI collectively announced that they would form a multiyear “exclusive computing partnership.”
On top of that, Microsoft ended up investing a cool $1 billion to help fund the development of new AI technology. And for that type of money, it certainly wanted to see a return.
So, in November of the same year, OpenAI finally opened the Pandora’s box it swore to never release. That month, it officially unveiled GPT-2 to the general public.
Simultaneously, OpenAI began to already profit from its partnership with Microsoft. The Seattle-based software giant created a 285,000-core machine with 10,000 GPUs that are collectively capable of processing 400 gigabytes per second. The machine even ranked as one of the five most potent supercomputers on planet earth.
Said supercomputer was used to accelerate the development of GPT-3, which was first discussed in a paper dating back to May 2020. One of the mindboggling aspects of Gen3 GPT was the fact it was trained on 175 billion parameters versus 1.5 billion parameters for GPT-2.
Interestingly, when the commercial version of GPT-3 finally dropped in July 2020, CEO Altman tried his best to avoid another media frenzy.
Meanwhile, OpenAI later exclusively licensed its GPT-3 tech to Microsoft, meaning the investor had access to the model’s code, which, unlikely many of its other projects, is not made available via open-source platforms.
However, not everyone was fond of Microsoft’s ever-increasing grip on OpenAI’s models and proprietary tech.
With that being said, even Microsoft could not hold back the hype train that OpenAI eventually unleashed upon the world.
Firing On All Cylinders
Being one of the dominating tech stories in a year where record numbers of tech people lost their jobs while others wasted away billions by purchasing social media companies certainly has to mean something.
2022 undoubtedly became the year that catapulted OpenAI from Silicon Valley’s best-kept secret to worldwide awareness.
Just to set the stage: OpenAI used 2021 to further entrench itself into the Microsoft ecosystem, for example by being incorporated into Github’s Copilot tool or Office365. In the meantime, it also cut some fat by dissolving its robotics unit, which worked on cool stuff like this:
Instead, it doubled down on its natural language processing and image generation research. The first version of OpenAI’s image generation tech, called Dall-E, was released back in January 2021.
OpenAI also started to up its revenue-generating offerings, for instance by enabling customers to use its API to finetune models to their individual use cases.
All of this paled in comparison to what was about to unfold throughout 2022. The first glimpse into that new reality was the release of DALL-E 2 in April.
The second-gen version of DALL-E was substantially more performant, meaning users were able to create almost photorealistic images in a fraction of the time. The leap was made possible by OpenAI’s switch to a diffusion model, which refines images until it reaches what it deems an optimum.
Millions of people jumped at the chance of creating all kinds of imagery, ranging from cover images for their blog posts all the way to whatever this is.
However, the creation of those images, which are oftentimes trained on work created by photographers and the like, soon caused a widespread backlash against OpenAI and other image creation models such as Stable Diffusion.
This backlash ultimately culminated in multiple petitions as well as the first lawsuit filed against Stable Diffusion maker Hugging Face in January 2023. Despite the widespread criticism, OpenAI went ahead and made Dall-E 2 available to everyone starting from July 2022.
Three months later, DALL-E 2 was even integrated into Microsoft’s Azure OpenAI Service, which the firms unveiled the year before.
Nevertheless, all of this paled in comparison to what was about to unfold. On November 30th, 2022, OpenAI quietly released ChatGPT, a refined version (v3.5) of its GPT-3 model.
To say that ChatGPT took the world by storm is probably the understatement of the year. Within five days of launching to the public for free, it already amassed over one million users.
People around the globe and from all walks of life couldn’t believe what they were seeing. Whether it’s creating bedtime stories for your children or answering complex programming questions, ChatGPT seemed to have a solid response for almost any query.
Many started to compare ChatGPT’s launch to the introduction of the iPhone in 2007, theorizing whether it may have a similar impact going forward. Google, in response to the craze, allegedly even issued a ‘Code Red’, which prompted CEO Sundar Pichai to directly lead the search giant’s AI teams.
Meanwhile, students were also using ChatGPT’s text-generation capabilities to submit exams. OpenAI, in response, began working on algorithmically watermarking the text that its GPT technology generates.
Once again, OpenAI managed to cause a widespread hysteria, which the firm used to its advantage during the fundraising process. So, without further ado, let’s finally take a closer look at who owns what of OpenAI.
So, Who Owns OpenAI?
Let me preface this section by highlighting that OpenAI’s ownership structure is currently not openly accessible to the public.
After all, OpenAI remains a private corporation and is thus not obligated to disclose revenue figures or ownership percentages to the public.
Luckily, we do have some data points that help us dissect who owns what of OpenAI. What we do know is that it’s one of the most unique deals ever done, largely due to OpenAI’s mission to distribute the benefits of AGI and its existing models to all humans and not just a single entity.
With that being said, this continues to be a developing story, so facts may change. After the worldwide ChatGPT hype, OpenAI and Microsoft decided to deepen their relationship, primarily to advance the development of GPT-4.
Microsoft, according to reporting from Semafor, was planning to invest $10 billion into OpenAI at a valuation of $29 billion.
After the deal officially closed on January 23rd, 2023, Microsoft ended up owning 49 percent of OpenAI LP. Another 49 percent would be in the hands of existing investors like Andreessen Horowitz as well as the firm’s employees. Meanwhile, the remaining 2 percent would be owned by the OpenAI non-profit foundation.
The deal would allow early employees and investors to cash out some of their equity holdings and thus take home some profits from their work.
Now, here is where the deal gets interesting. Microsoft would also receive 75 percent of all future OpenAI profits until it recoups the $10 billion it invested in.
What may or may not seem like the deal of the century for Microsoft is grounded in the fact that OpenAI continues to bleed money. CEO Altman described the compute cost of offering ChatGPT for free as “eye-watering” while OpenAI projected a loss of more than $500 million in 2022 alone.
The OpenAI non-profit arm, once the $10 billion have been paid back in full, would then receive 100 percent of all future profits. There have also been some rumblings that Microsoft’s stake in the LP would decrease as the investment is being paid back but this is yet to be confirmed.
In summary, Microsoft as well as existing OpenAI employees and shareholders would be the two majority owners of the commercial business. Simultaneously, the non-profit foundation cannot be owned by any single entity or person and is thus overseen by a group of independent board directors.
