Apple’s Sherlock killed the third-party Watson search utility for the Mac. Microsoft’s bundling of Internet Explorer into Windows steamrolled Netscape Navigator. And whatever OpenAI does next could upend today’s startup ideas just as quickly.
How startups can avoid that reality was one of the question at the Madrona IA Summit in Seattle on Wednesday. Jason Kwon, OpenAI’s chief strategy officer, shared advice for founders trying to innovate in the shadow of a company racing toward artificial general intelligence.
The gist of Kwon’s advice: look for specialized applications, vertical domains, and product experiences where OpenAI is unlikely to go deep.
For example, he pointed to areas such as manufacturing processes — problems that are highly specific, technical, and far from OpenAI’s quest for artificial general intelligence, or AGI, the still-theoretical goal of creating AI systems with human-level abilities across different tasks.
He warned against fine-tuning models or collecting data just to patch over the shortcomings of current AI models. A smarter bet is to assume the models will get better, and build products that deliver real value in more specialized niches.
The ‘critical path’ to AGI
It also helps to understand what OpenAI is doing in the larger scheme of things.
Speaking on stage with Madrona partner Vivek Ramaswami, and later answering questions from the audience, Kwon made it clear that OpenAI’s pursuit of AGI shapes every decision it makes, from what it chooses to build to where it decides to partner.
“If it’s on the critical path to general intelligence, that’s something we’re going to be interested in,” he explained. “If it’s not, by definition, we’re going to be less interested.”
That perspective also explains moves that might seem puzzling on the surface.
One example: OpenAI’s Sora video-generation model, which just hit V2. Kwon said Sora is more than a product in its own right — it’s a step toward AGI, helping AI systems learn about the physical world through video, not just text. Representing the world as a moving simulation could be key to teaching AI to reason more like humans.
ChatGPT is profitable, by one measure
In addition to strategy and research, Kwon also touched on OpenAI’s finances and business model, offering his thoughts for startups on how to think about costs and margins.
He noted, for example, that ChatGPT is already profitable in most markets when measured on a compute-margin basis — the difference between revenue from the chatbot and the direct cost of running the models.
OpenAI as a whole is far from profitable, as detailed most recently in a report by The Information this week based on internal financial data obtained by the news site. Kwon said this is because OpenAI is spending so much on compute capacity and research.
For startups, he cited the importance of understanding margins at a core level.
“If you’re doing well on the basis of how much you pay for compute, and how much you get per unit of delivery, and you’re positive on that, then you’re actually deriving more value out of the core input than you’re paying for,” he said.
Over time, he added, the expectation is that the cost of compute will decline.
ChatGPT’s role in e-commerce
Apart from startups, how worried should Amazon be about what OpenAI is doing? That question wasn’t asked directly, but Kwon gave some insights into OpenAI’s launch this week of Stripe integration to enable e-commerce.
He suggested the move isn’t about competing head-on with existing players, but about facilitating transactions and letting other companies build on top of OpenAI’s models. When a user asks about products, ChatGPT can show results and, in some cases, let them buy directly in the chat. Payment and shipping are handled by the merchant.
Kwon described it as an early example of how reasoning capabilities can turn AI systems into agents that don’t just suggest answers but take actions — such as making purchases — on a user’s behalf.
OpenAI’s own startup culture
Kwon also reflected on the company’s rapid growth since the launch of ChatGPT.
OpenAI had about 200 employees before the chatbot’s debut in November 2022. Since then, the company has tripled its headcount each year. Managing that kind of expansion, he said, requires constant focus on what matters most.
Kwon pointed to CEO Sam Altman’s approach at all-hands meetings, where he consistently emphasizes two priorities above all others: research and compute. Kwon said that helps provide clarity even as the organization grows quickly and faces new pressures.
That focus, he said, “centers the company.”
