SenseTime Unveils 'Da Zhuangzhi' at WAIC 2026: A Compute-Power-Coordinated Agent Boosts Tokens-per-kWh Output by 80%

SenseTime's 'Da Zhuangzhi': A New Infrastructure Paradigm for Massive-Scale LLM Inference

At the WAIC 2026 (World Artificial Intelligence Conference 2026) main forum, SenseTime officially unveiled 'Da Zhuangzhi'—a software-hardware co-designed infrastructure platform tailored for continuous inference services of billion-parameter LLMs. Unlike conventional data center upgrades, 'Da Zhuangzhi' adopts a four-dimensional coupling design (compute, power, cooling, and task flow), with its core innovation being the first integration of real-time power scheduling into an autonomous Agent decision loop—a programmable, verifiable, and auditable compute-power coordination engine.

Unlike static PUE optimization or coarse-grained load migration, 'Da Zhuangzhi' deploys lightweight edge Agents (based on SenseTime’s proprietary SenseAgent v2.1 framework) that continuously ingest GPU cluster power draw (NVIDIA H100 SXM5 nodes), time-of-use electricity pricing (via State Grid East China Dispatch Center API), liquid-cooling thermal ramp rates, and LLM request SLA tiers (e.g., <200ms latency queries vs. batch summarization tasks), dynamically re-planning computation graph partitioning, KV cache residency policies, and FP8/INT4 mixed-precision scheduling paths.

The Compute-Power-Coordinated Agent: From Power Awareness to Power-Aware Decision Making

This Agent is not a traditional monitoring module but a reinforcement learning agent with three-tiered decision capabilities:

• Perception layer: Fuses DCIM sensor data, NVML GPU power sampling (100 Hz), and real-time latency histograms from Qwen2.5-72B-Instruct inference;
• Modeling layer: Embeds SenseTime’s proprietary Power-Aware Mixture-of-Experts (PA-MoE) scheduler, mapping each millisecond-level request onto the optimal point of the 'power–latency–accuracy' Pareto frontier;
• Execution layer: Directly intervenes in runtime via CUDA Graph Patching and Custom NVLink Throttling Protocol (CNTP v1.3), enabling microsecond-level power redistribution.

Benchmark results show that over 72 consecutive hours of stress testing (QPS = 12,800; model: Qwen2.5-72B-Instruct + RAG enhancement), 'Da Zhuangzhi' increased tokens-per-kWh output from 1,240 to 2,230 tokens/kWh on identical H100 clusters—yielding an absolute gain of 990 tokens/kWh and a relative improvement of 80%. This result was independently certified by the China Academy of Information and Communications Technology (CAICT) Tel Lab (Report No.: CAICT-TAE-2026-0887).

Industry Impact: Rewriting the Cost Function and Business Model of AI Infrastructure

This milestone signals a paradigm shift—from 'compute-centricity' to 'energy-value-centricity' in AI infrastructure. While LLM service pricing has historically been anchored to GPU-hour rates and bandwidth, 'Da Zhuangzhi' explicitly models kWh as a first-class cost variable. SenseTime has signed joint pilot agreements with three IDC operators in the Yangtze River Delta to include 'electricity cost per thousand tokens' (¢/kToken) in SLAs and has released the SenseAgent SDK for ecosystem partners to integrate with their own schedulers.

Notably, the solution requires no new silicon or cooling media—it delivers this gain entirely atop existing H100 + immersion liquid-cooling stacks, underscoring the practical viability of Software-Defined Efficiency (SDE). The open-source component SenseAgent Core is available on GitHub (github.com/sensetime/senseagent-core v0.9.2), while the PA-MoE scheduler and CNTP protocol remain closed-source core modules.

Technical Boundaries and Open Challenges

Despite strong metrics, the current implementation faces three constraints:

• Supports only Transformer-based LLM inference—not training or fine-tuning;
• Requires minute-level electricity pricing APIs from smart grids (currently limited to Jiangsu, Zhejiang, and Shanghai);
• Under extreme ambient temperatures (>35°C), thermal headroom reduction in liquid cooling narrows the gain to 52% (still above baseline).

SenseTime confirmed that the next-generation 'Da Zhuangzhi' Beta (codename 'Fuxi') will launch in Q1 2027, extending compute-power coordination support to Gemini 2.0, Claude-4, and multimodal models (e.g., Qwen-VL-2), and introducing a Carbon Flow Tracking module for real-time Scope 2 emissions attribution.