Will Leading AI Labs Remain Profitable? Ex‑OpenAI Researcher vs. Dwarkesh on AGI Finding Jobs

Less than a day after leaving OpenAI, researcher Andrew Ho posted a long‑form critique on X, questioning the commercial viability of frontier AI labs. He highlighted OpenAI’s projected 2025 figures: $13 billion in revenue against $34 billion in total costs, resulting in a loss of over $20 billion, which translates to roughly $2.6 spent for every $1 earned.

Ho described the current model race as a “punitive loop,” where labs must continuously pour more chips, electricity, data, and talent into training the next generation of models just to stay ahead, a lead that can only be maintained for a few months.

He illustrated the pressure with two recent events: the GPT‑5.6 series launch on July 9 followed three weeks later by an 80 % price cut for Luna’s API and a 20 % cut for Terra; and the release of the full 2.8‑trillion‑parameter Kimi K3 weights, enabling third‑party deployment.

Ho warned that even if a lab maintains a capability lead, that advantage does not automatically translate into revenue. New technologies typically require long‑term experimentation before products can be market‑ready and profitable.

Dwarkesh Patel, a well‑known AI podcast host, responded by emphasizing that modeling capability is still growing rapidly and that being ahead itself generates huge commercial value. He argued that once true AGI appears, it will actively spread itself, seeking work much like a high‑skill immigrant who finds employment, learns rules, and even starts a company without waiting for society to provide a user manual.

Dwarkesh noted that model inference costs have been falling dramatically—Stanford’s AI Index 2025 reports a >280× reduction in the cost of reaching GPT‑3.5‑level inference from November 2022 to October 2024—so a three‑month capability lead can already yield noticeable product function jumps that users are willing to pay a premium for.

Both sides agree that the “price shelf‑life” of a leading capability is shrinking as competitors catch up. Ho worries this window is too short to cover R&D expenses, while Dwarkesh believes that if the next capability leap is large and fast enough, the leader can continually open new pricing space.

Elon Musk quickly weighed in, supporting the optimistic view that AGI will create its own economic opportunities.

The discussion also touched on two possible revenue paths for frontier labs: moving upstream in the chip supply chain to lower unit‑intelligence costs, and leveraging the massive scale effect of large‑model services, where a single trained capability can be sold to millions of users across high‑value domains such as programming, drug discovery, finance, and law.

Commenters suggested that even without immediate AGI, labs could survive by either reducing production costs through custom silicon and self‑built infrastructure, or by exploiting the scale‑economy of model licensing.

Overall, the debate frames the core tension: whether the escalating R&D spend required to stay ahead can be recouped before the advantage erodes, and whether AGI’s ability to autonomously find work will fundamentally reshape the economics of AI deployment.