February 03, 2026

Global data centres & AI infrastructure briefing (UTC 2026-02-03)

Audience: Institutional asset managers and infrastructure fund managers focused on data centres, power, and grid infrastructure.

Top news (3)

1. Hardening the enterprise AI supply chain + enabling offline deployment: Dell Enterprise Hub strengthens AI supply chain security and deployment highlights multi-layer image integrity and scanning (daily Amazon Inspector scans, Cosign-signed images, SHA-384 manifest hashes, and model malware/pickle scanning) alongside support for fully offline, on‑prem deployments.

2. Inference energy and cost can swing by orders of magnitude on the same GPUs: LLM Inference Efficiency: Quantization, Batching, and Serving Strategies reports that system-level serving choices (quantization, batching, configuration) materially change energy use on NVIDIA H100s; it also reports arrival shaping can reduce per-request energy by up to 100×.

3. Data centres as grid assets (frequency regulation + emissions lens): Data center load flexibility for grid frequency regulation and emissions introduces an “Exogenous Carbon” metric and the “EcoCenter” framework to coordinate GPU data centre load flexibility for frequency regulation, with the authors reporting potential savings that can outweigh operational emissions.

Key deals and projects

No M&A, financing, lease, land, or build project announcements were included in today’s story set.

Technology and operations (what could matter for capex, utilisation, and performance)

AI platform security, verification, and enterprise deployment

• Dell Enterprise Hub strengthens AI supply chain security and deployment
  • Security controls described: daily Amazon Inspector scans; Cosign-signed images; SHA‑384 manifest hashes; model malware/pickle scanning.
  • Enterprise verification: integrity and scan results are published with images (noted as being available to verify).
  • Deployment angle: enables fully offline on‑prem deployments, which can be relevant where air‑gapped or regulated environments constrain cloud usage.

Serving efficiency, autoscaling, and concurrency control (GPU utilisation and tail latency)

• CONCUR: Congestion-Based Concurrency Control for Agentic LLM Inference

  • Introduces an admission/concurrency control layer to avoid GPU KV-cache “middle-phase thrashing” in agentic batch inference.
  • Reported throughput improvements: up to 4.09× (Qwen3‑32B) and 1.9× (DeepSeek‑V3), while remaining compatible with existing serving systems.

• SAIR: Cost-Efficient Multi-Stage ML Pipeline Autoscaling via LLM

  • Uses an LLM as an in-context RL controller for autoscaling multi-stage inference pipelines.
  • Reported outcomes (under stated assumptions): up to 50% P99 latency improvement and up to 97% effective cost reduction; 86% bottleneck detection accuracy.

• Serverless GPU Architecture for Enterprise HR Analytics BDaaS Implementation

  • Presents a Helm-packaged BDaaS blueprint integrating a single-node serverless GPU runtime with TabNet for compliance-aware HR analytics.
  • Benchmarks reported vs Spark baselines: up to 4.5× higher throughput, 98× lower latency, and ~90% lower cost per 1K inferences; compliance overhead ~5.7 ms (p99 < 22 ms).

Memory and interconnect efficiency (CXL and model serving at long context)

• TRACE: Lossless Compression and Precision Scaling for CXL Bandwidth
  • Proposes device-internal transforms enabling lossless compression and precision-proportional fetch for CXL memory.
  • Reported footprint reductions: BF16 weights 25.2%, BF16 KV 46.9%.
  • System modeling claim: throughput for GPT‑OSS‑120B‑MXFP4 at 128k tokens increases from 16.28 to 68.99 tok/s (4.24×).
  • Overheads reported: area 7.2%, power 4.7%, latency 6.0%.

Heterogeneous GPU clusters (multi-vendor flexibility)

• HetCCL Accelerates LLM Training with Heterogeneous GPU Clusters
  • Presents a collective communication library unifying vendor-specific backends to enable RDMA-based cross-vendor GPU communication without driver changes.
  • Evaluation summary: matches NVIDIA NCCL and AMD RCCL performance in homogeneous setups and scales in heterogeneous environments.

Networking controls and custom congestion control

• NECC: Enabling Non-Expert Customized Congestion Control with LLMs
  • Uses LLMs plus the Berkeley Packet Filter (BPF) interface to model, implement, and deploy custom congestion control algorithms.
  • Paper noted as an accepted manuscript for IEEE ICC 2025 with “promising” results using real-world CCAs.

Power, grid, and interconnection highlights

• Data center load flexibility for grid frequency regulation and emissions
  • Positions GPU data centres as flexible loads to help provide grid frequency regulation.
  • Introduces the Exogenous Carbon metric and EcoCenter coordination framework.
  • Reported finding: exogenous carbon savings can often outweigh operational carbon emissions (per the preprint’s results summary).

Policy and regulation

No policy, permitting, tariff, interconnection-queue, or regulatory changes were included in today’s story set.

2-line wrap

Today’s flow was dominated by AI infrastructure enablement: securing model supply chains, raising GPU utilisation, and reducing memory/interconnect bottlenecks.
Separately, load-flexibility research continues to frame GPU data centres as controllable grid resources for frequency regulation and emissions outcomes.