Ex‑OpenAI Researcher: Large‑Model Firms Burn Money; Dwarkesh Says AGI Will Find Jobs

Andrew Ho, a former OpenAI researcher, posted a long critique on X claiming that the commercial models of frontier AI labs cannot sustain the costs required to reach AGI, describing the model race as a punitive loop.

He cites OpenAI’s 2025 figures: about $13 billion in revenue against $34 billion in total costs, resulting in a loss exceeding $20 billion, and notes that for every $1 earned the company spends roughly $2.6.

Ho argues that training the next generation of models demands ever more chips, electricity, data, and talent, and that any performance lead lasts only a few months before competitors release comparable closed‑source models or open‑weight alternatives, turning the previous generation into a cheap commodity.

Recent events illustrate this pressure: OpenAI launched the GPT‑5.6 series on July 9, then reduced Luna’s API price by 80 % and Terra’s by 20 % within three weeks; shortly before, Kimi K3’s 2.8‑trillion‑parameter weights were released for third‑party deployment.

Podcast host Dwarkesh Patel counters that model capabilities continue to grow rapidly, and a three‑month lead can produce a clear product jump—e.g., from code generation to autonomous debugging and deployment—making users willing to pay multiples for the advantage.

He points out that inference cost has fallen dramatically: Stanford’s AI Index 2025 shows GPT‑3.5‑level inference cost dropped over 280× from November 2022 to October 2024, so firms must convert the lead into revenue before competitors catch up.

On the revenue side, large‑model companies enjoy a scale effect: a single trained capability can be sold to millions, spreading the huge training expense across many customers, especially in high‑value domains such as programming, drug discovery, finance, or law.

Dwarkesh also envisions AGI as a self‑propagating digital worker that can read internal documents, learn tasks, and replicate itself, likening it to a high‑skill immigrant who finds work and creates companies without waiting for society to prepare.

He concludes that if AGI can autonomously spread and achieve sufficiently large, fast capability jumps, the “price‑quality” window of a lead will remain profitable; otherwise, the current cost structure may never be covered.