Amazon shifts Nova models to maintenance mode, redirects resources to single frontier effort under Abbeel

Amazon reallocated labor and GPU hours away from its Nova suite after internal benchmarks showed the models failed to reach competitive performance against GPT-4-class systems while consuming disproportionate power. The company closed an 80-person San Francisco site focused on AGI research one week prior to the reorganization. Remaining Nova instances receive only sustaining patches; no new training runs are scheduled.

Bloomberg reporting and prior Covariant acquisition filings indicate the pivot concentrates resources on embodied AI rather than scaled language models. Abbeel’s team, integrated after the 2024 Covariant deal, targets warehouse robotics where Amazon already holds deployment data from 2023-2024 fulfillment centers. This aligns with documented ROI gaps in pure LLM projects across multiple hyperscalers.

Operationally the change ends parallel model development tracks and enforces single-threaded frontier work. Headcount and power budgets previously split across text, vision, and video now flow to one program, reducing coordination overhead but narrowing output diversity. No public timeline exists for resumed broad model releases.

Next measurable signal will appear in Amazon’s Q4 2025 infrastructure capex filing and any disclosed robotics deployment metrics from Abbeel’s group.