Skip to main content
Loading research indexW
POSITIONS

What Could Change the View

LONG · 9698 HK · GDS UQ

GDS Holdings

Mixed
COMPANY EVIDENCEEfficiency can raise workload density but may reduce near-term need for incremental physical capacity.Reviewed paper · FlashInfer cut token latency by as much as 69 percent
LATEST RESEARCH SIGNALServing software can improve cloud gross economics while also releasing effective compute supply, creating opposite read-throughs for cloud platforms and hardware or data-center demand.Reviewed paper · FlashInfer cut token latency by as much as 69 percent
WHAT TO VERIFYProduction tokens per dollar and watt, utilization before and after adoption, cloud pricing, and whether server or colocation orders are deferred.PaperAlpha checklist · Based on reviewed evidence
SignalWhy it mattersEvidenceImpactAction
01
FlashInfer cut token latency by as much as 69 percent2025 · MLSys Proceedings

Efficiency can raise workload density but may reduce near-term need for incremental physical capacity.

External researchWatch
Read →
02
AI training’s idle time may be worth thousands of unused GPUs2025 · MLSys Proceedings

Higher cluster productivity can support tenants but delay incremental capacity.

External researchWatch
Read →
03
Moving AI memory to CPUs could unlock bigger batches without new GPUs2025 · MLSys Proceedings

Higher throughput improves tenant economics but adds CPU power use and can change density requirements.

External researchWatch
Read →
04
Shared prompts more than doubled LLM throughput2025 · MLSys Proceedings

Efficiency may raise workload density while reducing incremental capacity per request.

External researchWatch
Read →