OpenAI reallocates GPUs to coding and enterprise
SOTA model nightclub hype cycle
Shutting down Sora shows OpenAI no longer treats frontier video as a prestige side quest, it treats GPUs as scarce inventory that should be pointed at the products with the clearest path to revenue and lock in. Video was consuming enormous compute, about $15M per day by estimate, while coding agents and enterprise deployments sit much closer to OpenAI’s fastest growing monetization lanes and to the competitive fight with Anthropic.
-
Sora was expensive in a way text products are not. The video workflow used about 40 minutes of GPU time for each 10 second clip, and video inference was framed as roughly 100x more compute intensive than text, making it hard to justify while compute remained the bottleneck.
-
The strategic pressure came from coding. Anthropic had already become the backbone model for Cursor, Bolt.new, and v0, with about 80% of its revenue coming from API usage, then moved further up stack with Claude Code, a terminal agent that edits, tests, and debugs code directly.
-
OpenAI also lacked the downstream workflow edge in video. Strong video businesses pair generation with editing, distribution, or a large native content surface, like YouTube, TikTok, CapCut, or creator tools such as Runway, Synthesia, and HeyGen. Sora had model quality, but not the full product loop that turns clips into a durable business.
From here, OpenAI looks more like a compute allocator than a broad AI lab. New capacity is likely to keep flowing into coding agents, enterprise copilots, and cloud distribution, where usage compounds into recurring revenue and customer dependence, while AI video increasingly belongs to companies that already own either the audience, the editing workflow, or both.