Runway co-CEO says a video model request consumes about 100 times as many tokens as an LLM, and cost is a main reason Sora is hard to socialize
Runway co-CEO Anastasis Germanidis said in a podcast interview that a single request to a video model consumes about 100 times as many tokens as a request to a large language model (LLM). He pointed out that cost is one of the main reasons OpenAI faces difficulties in pushing Sora 2 toward social applications.
Germanidis said that the cost of running these models is very high. He added that this is also why Runway is investing heavily in real-time, distilled, and efficient video models. He expects many such models in the future to run locally on phones, with only part of the work done remotely.
Runway was founded eight years ago, and added more than $100 million in ARR last quarter. The company had previously reached about $200 million in ARR and plans to grow further before the end of the year. Germanidis also mentioned that the company experienced a brief crisis after Sora was released and推出ed a better model within a few months.
Original source:
Title: Inside Runway's $5.3B Playbook (podcast) Publisher: The Luba Show / Luba Yudasina Publication date: 2026-11-02 URL: https://podcasts.apple.com/us/podcast/inside-runways-%245-3b-playbook/id1858312769?i=1000799429470 Source type: podcast interview Verification status: Based on reading the podcast description and related reports for confirmation, verification date 2026-11-06 Supporting content: Supports all core quotations in the body, as well as the token consumption and cost statements
ABAB AI Insight
Runway co-CEO Anastasis Germanidis said in a podcast that a single request to a video model consumes about 100 times as many tokens as an LLM. He pointed out that this is a major reason OpenAI faces overly high costs when pushing Sora 2 toward social applications, and said that the cost of running these models is very high.
On costs and technical direction, Germanidis said Runway is investing heavily in real-time, distilled, and efficient video models to reduce inference costs. He expects many video models to run locally on phones in the future, with only part of the work done remotely. The company added more than $100 million in ARR last quarter, currently has about $200 million in ARR, and plans to grow further before the end of the year.
Compared with OpenAI’s Sora, which shut down its consumer app because of cost issues, Runway is shifting its focus toward enterprise and world models. The similarity is that both face high compute costs for video generation; the key difference is that Runway is achieving revenue growth by embedding with enterprise customers in production workflows rather than relying on consumer social applications. Germanidis also mentioned that the company had unexpectedly developed a model suitable for robots.
For Runway, efficient models and the enterprise market are enabling continued growth under cost constraints. For OpenAI, the Sora consumer app was shut down because unit economics were not viable. For the industry, the high token consumption of video models is driving investment in distillation and on-device execution to reduce the cost of large-scale deployment.
ABAB News · Cognitive Law
- A single request to a video model consumes about 100 times as many tokens as an LLM, and cost is the main reason Sora is difficult to socialize.
- Runway is investing in real-time, distilled, and efficient models, and expects many models to run locally on phones in the future.
- The company added more than $100 million in ARR last quarter, and currently has about $200 million in ARR.