Creating a new market is like starting a long race. Competitors are racing for a position while spectators cheer enthusiastically. Then, just like in races, markets enter a second, more subdued phase. The field is sorted between leaders and stragglers. The crowds dwindle.
In the battle to dominate the future of artificial intelligence (AI), OpenAI, a Microsoft-backed company, gained an early advantage with the launch of ChatGPT last November. The app reached 100 million users faster than any previous app. Rivals scrambled to catch up. Google and its parent company, Alphabet, rushed to release their chatbot, Bard, and startups like Anthropic followed suit. Big funds invested over $40 billion in AI companies in the first half of 2023. Public interest in AI peaked a couple of months ago, according to Google search data. The number of visitors to the ChatGPT website dropped from 210 million in May to the current 180 million.
In terms of technology, OpenAI is still at the forefront. Its latest AI model, GPT-4, is outperforming rivals in the majority of fields (such as reading and math comprehension). In a head-to-head battle, it surpasses the current runner-up, Anthropic’s Claude 2, by a decent but not insurmountable margin. More importantly, OpenAI is starting to make real money. According to tech publication Information, it is generating annualized revenue of $1 billion, compared to the meager $28 million it made in the year prior to the launch of ChatGPT.
Can OpenAI translate its early advantage into lasting success and join the ranks of the tech giants? To do so, it must avoid the fate of early tech pioneers from Netscape to MySpace, who were overtaken by rivals who learned from their early successes and missteps. And as a pioneer, the decisions it makes will also say a lot about the broader direction of the nascent industry.
OpenAI is an intriguing company. It was founded in 2015 by a group of entrepreneurs, including current CEO Sam Altman and Tesla CEO Elon Musk, as a nonprofit organization. Its goal was to build artificial general intelligence (AGI), which would match or surpass human capability in all intellectual tasks. An intermediate goal was an AI that could dominate a video game called Dota. While working on that problem, OpenAI experts focused on a simple approach that involved leveraging massive computing power. In 2017, when Google researchers published a paper describing a groundbreaking machine learning technique called “transformer,” OpenAI engineers realized they could expand on it by combining vast amounts of data from the internet with processing power. The result was the pre-trained generative transformer, or GPT for short.
Securing the necessary resources required some financial engineering on OpenAI’s part. In 2019, the company created a “benefit-capped corporation” within its nonprofit structure. Initially, investors in this business could earn up to 100 times their initial investment, but no more. Instead of distributing capital, the company distributes rights to a portion of future profits without ownership rights (“profit-sharing units”). OpenAI emphasizes that this is a “high-risk investment” that should be viewed more like a “donation.” “We’re not for everyone,” says Brad Lightcap, OpenAI’s COO.
Musk stepped down in 2018, and some potential investors were hesitant about OpenAI’s recent funding round due to its complex structure. However, Altman and Lightcap managed to win them over. To become more appealing, the company loosened its profit limit to be based on an annual rate of return (though it won’t confirm the maximum rate). And aside from academic debates about the meaning of the action, the profit-sharing units themselves can be traded on the market like standard stocks. The company has already provided several opportunities for early employees to sell their units. Investors who chose to buy in seem to trust that they can achieve risk-scaled returns if the company continues to grow.
SoftBank is believed to be the latest investor willing to make a big bet on OpenAI. So far, the startup has raised around $14 billion. The majority, perhaps $13 billion, came from Microsoft, whose Azure cloud division is also providing OpenAI with the necessary computing power. In return, the software giant will receive the majority of OpenAI’s profits, if they ever materialize. In the short term, it will be able to license OpenAI’s technology and offer it to its own clients, including the majority of the world’s largest companies.
Altman has said that OpenAI could very well end up being “the most capital-intensive startup in Silicon Valley history.” Training OpenAI’s latest model, GPT-4, is estimated to have cost around $100 million, several times more than GPT-3.
For now, investors seem happy to pour more money into the business, but they ultimately expect a return.
GPT-4 already shows some cost-consciousness. For example, as noted by Dylan Patel of research firm SemiAnalysis, it was divided into 16 parts specializing in different types of tasks. This makes its design more complicated than a monolithic model. However, it becomes cheaper to use the model once it has been trained because not all specialists are required to answer questions. Cost is also a significant reason why OpenAI is not training its next big model, GPT-5. Instead, according to sources familiar with the company, they are building GPT-4.5, which would have “similar quality” to GPT-4 but cost “much less to run.”
But it is on the revenue generation side of the business where OpenAI has undergone the most transformation and where it has been most energetic lately. “AI can create a lot of value long before acting. Brains are as versatile as humans,” says Lightcap. OpenAI’s models are generalists, based on a large amount of data, and capable of performing a variety of tasks. The popularity of ChatGPT has made OpenAI the default choice for consumers, developers, and companies eager to adopt the technology. Despite the recent drop, ChatGPT still receives 60% of the traffic to the top 50 generative AI websites, according to a study by venture fund Andreessen Horowitz, which has invested in OpenAI.
However, OpenAI is no longer just about ChatGPT, or even primarily about it. It is increasingly becoming a B2B platform. It is creating tailor-made products for large corporate clients, including Morgan Stanley bank. It also offers tools for developers to build products using its models.
Additionally, the company has a $175 million fund to invest in smaller AI startups that create applications on its platform, simultaneously promoting its models and capturing value if app creators strike gold.
Being the first mover certainly works in OpenAI’s favor. The high fixed costs of GPT-type models create significant barriers to entry for competitors. This, in turn, can make it easier for OpenAI to attract corporate clients. If they are going to share internal company data to customize the model to their needs, many clients may prefer not to do it more than once, for cybersecurity reasons or simply because it is costly to move data from one provider to another, as is already the case between cloud providers. Teaching large models to think also requires a lot of tacit engineering knowledge, from recognizing relevant data to understanding context.
In the race to dominate the future of AI, OpenAI has taken an early lead, but the real test will be whether it can maintain that lead and secure a lasting position in the market.