OpenAI continues attracting unprecedented levels of funding, but the company’s one-model-to-rule-them-all approach to generative AI comes with its own set of risks.

OpenAI is already the best-funded AI startup by far, having raised an estimated $13.5 billion. Reports say the company is now seeking $6.5 billion in additional venture capital at a valuation of $150 billion (it was valued at around $86 billion back in February), along with $5 billion in debt financing.

But that’s not to say the future is anything close to certain for OpenAI and its CEO Sam Altman.

OpenAI’s appeal

Between the summers of 2023 and 2024, OpenAI’s annualized revenues grew from $1 billion to $3.4 billion. Most of that money comes from enterprise customers who pay to access OpenAI’s specialized models. (OpenAI also earns income via ChatGPT Pro and Enterprise subscriptions.)

OpenAI has made it easy for enterprises to integrate AI into some aspects of their business. The customer gets access to some of the most advanced models without having to incur all the costs of developing their own models. (A company might, for example, use the model to glean business insights from customer-support call transcripts.) While OpenAI does let customers do a certain amount of fine-tuning of its models with their own business-specific data, most enterprises use OpenAI models as is—often with good results.

An unknowably huge market

OpenAI’s stated goal is to produce all-purpose models that possess artificial general intelligence (AGI), or the ability to complete most tasks as well as humans. The company has just taken a major step toward that goal with the release last week of its GPT-o1 models, which can tackle large, multistep problems (and show their work). Enterprises will undoubtedly pay a higher price for access to GPT-o1, which would be deployed for complex financial modeling or to power autonomous agents.

The big picture is that OpenAI’s addressable market could be unknowably huge. The company could be in a pole position to help lead enterprises and other organizations into a business revolution on the scale of the Industrial Revolution.

“Every advance in AI capabilities is going to unlock a new set of use cases that were previously impossible,” says Robert Nishihara, cofounder and CEO of the AI computing platform AnyScale. “We’re at the outset of tremendous value being generated by AI. To realize that value, much of the hard work ahead involves taking existing [and future] capabilities, making them incredibly reliable in the real world.”

Frontier economics

In the meantime, despite its revenue growth and research progress, OpenAI is likely not profitable and may not be for some time. And running an AI company, especially one that’s developing large frontier models, is an extremely costly endeavor.

While researchers are constantly finding ways to make training AI models more cost-efficient, developing ever-more-capable models requires access to expensive and power-intensive server clusters. A 2023 report from SemiAnalysis’s chief analyst Dylan Patel estimated that hosting the GPT-3-powered ChatGPT cost $700,000 per day. On top of that, OpenAI is paying publishers and other content owners for the right to use their proprietary data for model training.

If you build it, they won’t necessarily come. OpenAI remains focused on building models that are huge, general-purpose, and closed (meaning secret and proprietary). But not everybody wants to share their business data with an OpenAI model. This is especially true when it comes to highly regulated players, one source tells Fast Company, such as banks or healthcare institutions, some of which are using open-source models like Meta’s Llama, then fine-tuning them with their own data.

Some people in the industry believe those two very different models and approaches can coexist—that both would be useful, depending on the customer and the project. But others question long-term profitability of developing ever-smarter frontier models. “They haven’t proven, in my opinion, that they have a successful business strategy,” says Andrew Jardine, who managed go-to-market at Hugging Face and now leads GTM at the fine-tuning platform Adaptive ML. “Because if you have to keep continually training larger and larger models, which is exponentially more expensive, then you have to make sure that you can validate that cost to the consumer, which they haven’t done yet.”

Looming questions

There’s also the question of OpenAI’s corporate structure. The company started life (in 2015) as a nonprofit, but over the past five years has grown, sometimes painfully, into a profit-driven company. In 2019, after realizing the immense cost of developing GPT-3 and the need to raise money, OpenAI formed a “capped-profit” subsidiary within its nonprofit structure that same year.

OpenAI’s current bylaws still place limits on how much money investors can make off the company. Reuters reports that the completion of the new $6.5 billion at a $150-billion-valuation round may be conditioned on the removal of that profit cap. Per Reuters, OpenAI is now considering changing its structure to become a for-profit benefit corporation more like Anthropic and xAI.

OpenAI, and the rest of the AI industry, also faces a potential regulatory regime in California. A new law containing a number of safety and reporting requirements for frontier-model developers has passed both houses in the California Legislature and now awaits Governor Gavin Newsom’s signature, or veto. The tech industry has been exerting maximum pressure both online and via lobbyists to derail the bill.

Even though OpenAI has so far built the best models, attracted the most money, and become the poster child for generative AI, there’s no guarantee that the company’s investors will ultimately see the payday they’re hoping for.