Nvidia DSX Targets AI’s Power Bottleneck with Unified Grid Solutions

The story of the day

Nvidia’s DSX platform is reframing the AI power crisis as a systems integration problem. The solution: treat the grid, data center, and compute as a single system. With $300B in backlogged equipment and 200+ GW of projects stuck in U.S. interconnection queues, the bottleneck isn’t just capacity—it’s coordination. DSX Sim and SimReady assets let builders model power flow from substation to rack, exposing inefficiencies in siloed designs. For deployers, this means fewer stranded assets and faster time-to-market. For investors, it signals a shift from raw capacity plays to optimization tech.

New tools & launches

Nvidia Cosmos — Open tools for training, evaluating, and deploying world models, including data factory blueprints for physical AI. [Nvidia AI Blog]

Agentic CleanSlop — Open-source tool to authenticate, scan, and fix app vulnerabilities via agent workflows. [Product Hunt]

Tetriz Engg Intelligence — Connects coding sessions to PRs, issues, and ROI metrics. [Product Hunt]

Models & research

KVLearn (arXiv:2610.28637) — A cost-aware KV cache retention framework for disaggregated LLM serving. Reduces network and memory waste by predicting prefix reuse probability without modifying model weights. Open-weight.

TF-IG (arXiv:2610.28693) — Parameter-efficient time-frequency learning for PPG-based heart rate estimation. Dual-branch architecture captures pulse morphology and cardiac periodicity with <1% params vs. baselines. Open-weight.

Industry pulse

Revolut’s head of AI confirms its PRAGMA foundation model scales to 80M users on Nvidia Blackwell Ultra GPUs, processing billions of transactions. The move underscores a trend: financial services are betting on open-weight models running on proprietary infra to avoid vendor lock-in. [Nvidia AI Blog]

Stat of the day

200+ GW of U.S. energy projects are stalled in interconnection queues, per Nvidia’s DSX announcement. The backlog is worth $300B in stranded equipment. [Nvidia AI Blog]

Get the full analysis — subscribe to Gradient Pro.