Utilization gaps are real.
Server and GPU infrastructure can consume substantial power while producing less than proportional economic output.
Synestria Knowledge Center
Research Foundation
The industry has already documented the problem. Synestria measures, attributes, and helps recover its economic impact across the AI factory chain.
Research explains what Synestria believes and why. The references page shows the underlying source material. The downloads page provides public Synestria papers.
Evidence frame
Hidden losses are documented. The coordination layer is missing.
The research points to utilization gaps, stranded capacity, metric blind spots, and proven value from software coordination.
Research findings
Four evidence pillars support the Synestria thesis.
Capacity can be stranded.
Provisioned power, cooling, and compute capacity can exist physically without being safely converted into productive workload output.
Existing metrics have blind spots.
PUE, uptime, and availability measure whether systems are running. They do not measure whether the AI factory is producing at economic potential.
Coordination creates value.
Published work from hyperscale production environments shows software coordination can recover value without replacing the underlying infrastructure.
Physical fabric creates hidden loss.
Production data from the largest AI training operations in the world documents that fiber and networking failures create silent, cross-domain economic loss that does not appear in conventional availability metrics.
Evidence Strength Matrix
Where the evidence is strongest.
Research AreaEvidence StrengthComment
Utilization GapsStrongDocumented repeatedly in industry, national lab, and academic research.
Stranded CapacityStrongSupported by infrastructure and capacity planning research.
Metrics Blind SpotsStrongPUE and uptime limitations are well understood.
Software CoordinationStrongProduction-scale optimization results have been published.
Economic AvailabilityEmergingSynestria extends existing evidence into a new operating metric.
Operational MemoryEmergingExpected to become a differentiated knowledge asset over time.
AI Factory Coordination LayerEmergingCore Synestria thesis and category opportunity.
"We believe the next-generation winner here will combine live telemetry, predictive simulation of thermal and power flows, and AI-native anomaly detection into a system that operators actually trust to make autonomous decisions, not just dashboards that consultants interpret." — Bessemer Venture Partners, May 2026. The category winner does not yet exist. Synestria is designed to be that system.
Physical Fabric and Networking Layer
The fiber connecting compute to the network is a documented source of hidden loss at commissioning and at scale.
Fiber cabling and network interconnects are not background infrastructure. They are the physical layer that converts provisioned GPU capacity into delivered AI output. When they degrade — silently, without triggering alarms — the consequence chain runs from physical link to compute output to economic availability.
The Ghost Problem
Every fiber link flap corrupts the network's self-knowledge. A link reports "up" but drops traffic. A GPU appears healthy but its training job stalls. No alarm fires. The industry has a name for this: a ghost. At 2025 cluster scale — over 10 million optical links — a ghost occurs somewhere in the fabric every 48 seconds.
The Commissioning Gap
Polarity errors, end-face contamination, and microbending from improper cable routing are introduced at commissioning and persist for the life of the facility. A single miswired trunk connecting 96 GPU ports can silently degrade performance across an entire training cluster. Most commissioning processes test pass/fail, not degradation trajectory.
The Checkpoint Tax
The industry's primary response to networking failures is checkpointing — saving training state so jobs can restart after interruptions. Meta's LLaMA 3 training experienced 419 interruptions in 54 days. ByteDance logged over 38,000 failures in three months. Checkpointing consumes 12–43% of training time. That is not a failure recovery mechanism. It is a normalized economic loss.
The Cross-Domain Blind Spot
Network monitoring tools see link state. Compute monitoring tools see GPU utilization. No existing domain tool correlates a degraded NVLink connection with the reduction in training throughput it causes, or traces that throughput reduction to its economic consequence. The domain boundary is where the loss hides.
Fiber Link DegradesBelow alarm threshold. Reports healthy.
Network Ghost CreatedTraffic routed to unreachable path.
GPU Throughput DropsTraining job slows. No fault declared.
Checkpoint Tax Applied12–43% of training time consumed.
Economic Availability GapRevenue-producing compute never delivered.
The physical fabric consequence chain spans four domain boundaries. Synestria is designed to observe all four simultaneously — the only intelligence layer that can trace from fiber to economic output.
Independent Research Highlights
The industry already documented the problem. Here is where.
Synestria did not invent utilization gaps, stranded capacity, or the value of software coordination. Peer-reviewed research and government studies established all of it. These are the key sources.
Google DeepMind / NeurIPS, 2022
Controlling Commercial Cooling Systems Using Reinforcement Learning
Reinforcement learning achieved 9% and 13% energy savings at two live production data center sites. Peer-reviewed, production-scale proof that AI can autonomously control infrastructure systems and deliver measurable economic results — not just recommendations.
View on arXiv →Google / IEEE Computer, 2007
The Case for Energy-Proportional Computing
Server power consumption does not scale proportionally with useful work. A facility can run at 99.99% technical availability while producing far less than proportional economic output. The foundational academic argument for why Technical Availability and Economic Availability are not the same metric.
View on Google Research →Google / ISCA, 2007
Power Provisioning for a Warehouse-Sized Computer
Large-scale facilities can safely recover value from the gap between provisioned peak power and actual peak consumption — but only when compute, power, and workload behavior are coordinated simultaneously. Stranded capacity is documented, real, and recoverable.
View on Google Research →Google DeepMind, 2016
DeepMind AI Reduces Google Data Centre Cooling Bill by 40%
Machine learning reduced cooling energy at Google production data centers. One of the clearest public examples that software intelligence can materially improve infrastructure economics without replacing any hardware — the coordination layer does the work.
View DeepMind Article →LBNL / U.S. Department of Energy, 2024
United States Data Center Energy Usage Report
U.S. data center electricity use grew from 58 TWh in 2014 to 176 TWh in 2023 and is projected to reach 325–580 TWh by 2028. AI infrastructure is becoming a national-scale energy and economic system. The scale of the problem makes optimization economically consequential in ways it never was before.
View LBNL Report →Bessemer Venture Partners, May 2026
Legacy DCIM Platforms Are Architected for an Earlier Era
"We believe the next-generation winner here will combine live telemetry, predictive simulation of thermal and power flows, and AI-native anomaly detection into a system that operators actually trust to make autonomous decisions, not just dashboards that consultants interpret."
The category winner does not yet exist. Bessemer named the architecture. Synestria is building it.
Read BVP: The AI Data Center Stack →Research Update Policy
A living evidence base.
The Synestria Research Foundation is updated as new peer reviewed studies, hyperscale operating data, and AI factory research become available.
Perspectives — Analysis & Commentary
Going deeper on specific consequence chains.
Each piece below examines one specific loss pattern in detail — the physics, the economics, and why existing systems miss it.
Thermal
Thermal Derating: The Hidden Tax on Your GPU Cluster
How thermal drift quietly reduces compute throughput while every system reports normal operation — and how to measure the loss.
Commissioning
What Commissioning Defects Cost You Over Ten Years
Small calibration errors at commissioning compound for years. The cumulative cost is almost never measured — until Synestria traces it.
Power Economics
The Power Price Window: Why Real-Time EA Decisions Are Worth More
Real-time power price signals create economic windows for AI workload scheduling. Most operators cannot see them. Synestria can.
Capacity
The Stranded Megawatt Problem
Committed power that sits unused while compute demand goes unserved. The stranded megawatt is the most expensive outcome in AI factory economics.
Next step
The research is the foundation. The pilot is where it becomes real.
Synestria's 35-day pilot produces a quantified EA baseline for your specific infrastructure. Everything on this page describes what the pilot is built to find.