Conference Presentation, Keynote
Agents are Here. Compute has to Change | Sachin Katti, OpenAI | RAISE Summit 2026
- Industry transition from chatbots to reasoning capabilities is expected to begin in late 2022, evolve into the O-series model phase in 2024, and currently resides in an era of agents where AI interns are deployed to define, execute, and aid next-generation research to enable full recursion.
- Internal transformation utilizing Codex for nearly all output tokens and agentic interfaces for daily work is planned as a predictor for broader enterprise adoption, with rapid integration into non-engineering disciplines like customer support, legal, and sales intended to demonstrate fundamental impact on knowledge work.
- The iteration speed for training and research is accelerating due to agent deployment, shifting the product release cadence from a previous six to nine-month cycle to effectively one new model every month.
- AI-performed research will enable individual researchers to run significantly more experiments in parallel, driving an explosion in compute volume required for larger training clusters and increasingly complex individual experiments.
- Scaling laws across pre-training, synthetic data, post-training, and test time compute are expected to place significant stress on research compute infrastructure as more power is needed to push the intelligence frontier.
- Product compute requirements are anticipated to explode in volume and heterogeneity, necessitating a shift beyond simple GPU clusters to include CPUs, storage, and complex networking to support long-horizon tasks and tool interleaving.
- Compute optimization efforts aim to reduce task completion time and latency, targeting near-instantaneous responses for short turns and loops to maintain user flow, while a diverse fleet with multiple latency profiles will support everything from instant responses to deep research tasks.
- Accelerating agentic workflows for both AI research and knowledge work are predicted to sustain a continuous increase in the pace of model releases, delivering more capable models faster than ever before.