Right now. In your facility.

Your dashboards
are green.
Keep reading.

You're operating one of the most capital-intensive machines ever built. Your operations team is excellent. Your monitoring systems are best-in-class. None of them were designed to see what happened in your facility today.

Your facility. Today.

The record shows flawless operations.

Your team has built something genuinely exceptional. Tier IV. Best-in-class monitoring. Sophisticated operations. Every metric you can measure confirms it.

The question no metric can answer is the one that determines your NOI.

Technical Availability

Tier IV. 26 minutes downtime per year.

99.995%

Power Usage Effectiveness

Flat for five consecutive years.

1.45 PUE

DCIM Dashboard

Rack power, thermal maps, capacity utilization.

Nominal

BMS Status

Cooling, temperature, airflow — all within bounds.

No Alerts

All systems operational. Facility performing as designed.

What did your facility produce today
relative to what it was capable of producing?

That question has no answer in any of those systems. Not because the data isn't there — but because no tool you own was designed to ask it. The gap between those two numbers — what you produced and what you could have produced — is real, it is large, and it does not appear in any report.

A training run in your facility. Today.

This happened while your dashboards showed green.

Composite of documented production scenarios. Brookhaven / LBNL research, H100-class infrastructure. Timing is representative, not exact.

09:47

A training run queued across GPU infrastructure in halls 3 and 4.

Batch size: 512 images per step. Scheduler allocates compute. Parameters within normal ranges. Nothing unusual.

10:02

Cooling demand in hall 3 rises 23%. BMS responds.

CRAH units increase output. Chiller load adjusts. System maintains thermal envelope. BMS status: nominal. Nothing logged as unusual.

10:02–14:23

Four hours, thirty-six minutes. All dashboards: green.

SCADA nominal. BMS nominal. DCIM nominal. Uptime: 100%. PUE within target. No alerts. No anomalies. Exactly as designed.

14:23

Run completes. Model outcome: successful.

Training objective achieved. All infrastructure returned to baseline. No incident report. No review triggered. Operations nominal.

The same run. A different decision.

Alternative configuration — identical model outcome

4,096

Batch size (vs. 512)

Cooling demand increase (vs. 23%)

74%

Lower energy and cooling cost

Identical

Model accuracy and outcome

Your workload scheduler had no way to know. Your BMS had no way to tell it. Every system did exactly what it was designed to do. Not one of them was designed to do this.

That event happened in your facility today.
Your dashboards showed nothing unusual.

Because nothing unusual showed up. The loss is structural. It registers in no alert, no report, and no metric you currently own.

Find your facility.

What the coordination gap costs per day.

Recoverable losses from cross-domain coordination failures. H100-class baseline. No hardware change required to recover this value.

Facility Size	Per Day	Per Month	Per Year
100 MW	$315K	$9.6M	$115M
250 MW	$790K	$24M	$288M
500 MW	$1.6M	$48M	$576M
1 GW	$3.2M	$96M	$1.15B
2 GW	$6.4M	$192M	$2.3B

Derived from peer-reviewed production research. Sources: Synestra EA White Paper §5.2 · LBNL-2001637 (2024) · IEEE Access 2025 (Latif et al.) · ASPLOS 2020 (Sakalkar et al.) · Google DeepMind (2016). Full citations: Research · Downloads

The tools you trust

You didn't buy inferior tools.
You bought tools built for a different machine.

Each system is world-class. Each does exactly what it was designed to do. None of them was designed to see your facility as a single economic machine — because no AI factory existed when they were built.

SCADA

Sees power.

Grid draw, breaker state, UPS load. Manages the electrical domain with precision. Exactly as designed.

BMS

Sees cooling.

Temperature, airflow, chiller load. Keeps the thermal envelope within bounds. Exactly as designed.

DCIM

Sees infrastructure.

Rack power, server inventory, capacity utilization. Tracks what is deployed and where. Exactly as designed.

Workload Scheduler

Sees job queues.

Batch parameters, GPU allocation, training runs. Maximizes compute throughput. Exactly as designed.

Every system reports nominal.
None of them can see each other's cost.

The scheduling decision that determines your cooling load is made by a system that cannot see the cooling plant. The cooling response that consumes your energy budget is made by a system that cannot see the workload that caused it. The gap is not a monitoring problem. It is a coordination problem — and no tool on the market was designed to solve it.

Why this problem exists now

Why Now

Traditional data centers intentionally separated operational technology from compute infrastructure.

That separation was correct.

AI factories changed the economics.

Power, cooling, networking, workloads, and economics now influence one another continuously.

The relationships became economically significant before the industry developed a way to observe them.

Production findings

These were the best-operated facilities in the world. They were surprised too.

Every operator below had already invested heavily in operational excellence. Their monitoring showed green. The losses existed anyway — because they lived in the space between systems.

Google DeepMind

Production deployment, 2016
Evans & Gao

After years of hardware investment and expert optimization —

Cross-domain coordination cut cooling energy by 40% more.

Applied to cooling infrastructure that had already been extensively optimized. Produced a 15% facility-level PUE improvement that zero additional hardware investment could have captured. The opportunity existed entirely in the coordination layer no tool was providing.

Google

Sakalkar et al.
ASPLOS 2020, ACM

Across tens of megawatts, over multiple years —

Software coordination sustained power oversubscription of 25%+, saving hundreds of millions.

No new hardware. The entire opportunity existed in the coordination layer no existing tool was designed to provide. An operational advantage that their monitoring systems never made visible — only direct cross-domain coordination could capture it.

Brookhaven / LBNL

Latif et al.
IEEE Access, March 2025

On a single H100 HGX node, one parameter change —

Produced a 4× swing in facility power draw with identical model outcomes.

Changing batch size from 512 to 4,096 during ResNet training: fourfold difference in training energy. The workload scheduler had no visibility into the facility-level consequence of its decision. Your scheduler made that same decision this morning. Your BMS responded. Neither system knew what the other cost you.

Lawrence Berkeley
National Laboratory

Shehabi et al.
LBNL-2001637, 2024

In facilities where every metric showed operational health —

More than 30% of servers confirmed in comatose state: powered on, producing nothing.

At $200K–$400K per GPU server, comatose infrastructure represents an extraordinary concentration of idle capital — invisible to every operational report. The facility's technical availability remains 99.995%. The economic availability of that capital is zero. No tool measured the difference.

The realization

Your facility is losing
millions
you cannot see.

The tools are real

You have SCADA. You have BMS. You have DCIM. You have workload schedulers. They are all functioning correctly. They are giving you every insight they were designed to give you. That is not the problem.

The loss is real

The coordination gap is not theoretical. It has been measured at Google, at national laboratories, at production scale across multiple GPU generations. It exists in your facility at the same scale. Right now. Today.

The gap is structural

The loss lives between your tools, not inside them. Investing further in any single domain — better SCADA, newer DCIM, upgraded BMS — does not close it. You cannot buy your way out of a coordination problem with monitoring tools.

You cannot see it because no tool was designed to show you.
That is a structural condition of how AI factories are built — and it has a name.

Introducing the metric

Synestra refers to the difference between potential output and realized output as Economic Availability.

Economic Availability™

For thirty years, the industry has measured whether the facility is on. AI factories require a different question: not whether the machine is running, but whether it is producing at its potential.

The Definition

EA = Actual Annual Revenue Output ÷ Maximum Possible Annual Revenue Output

Expressed as a percentage. EA = 100% means every megawatt of provisioned power is converting to economic output at design efficiency. Industry current: 72–78%. The gap is real, it is measured, and it is recoverable.

What you've had — Technical Availability

Is the system on?

Binary. Uptime vs. downtime. Answers whether the machine can serve. Does not capture whether it is producing at potential. A facility at 99.995% TA can simultaneously operate at 72% EA — and most do.

What AI factories require — Economic Availability™

Is the system producing?

Revenue-referenced. Workload-sensitive. Captures coordination losses across all five domains simultaneously. The metric that makes the gap visible, quantifies it in dollars, and enables the decisions that close it.

99.995%

Technical Availability
Your facility today

≠

72–78%

Economic Availability
Your facility today

→

~95%+

Economic Availability
With coordination

"Technical availability tells you the facility is running.
Economic Availability tells you whether it is producing."

The platform

How Synestra Works

Synestra sits above existing systems rather than replacing them.

It receives telemetry from power generation, electrical distribution, cooling, networking, compute infrastructure, workloads, and operational systems.

By continuously learning how these systems influence one another, Synestra builds operational memory across the AI factory.

It exposes relationships individual systems cannot see and identifies recoverable economic value that would otherwise remain hidden.

Initial deployments operate with humans in the loop.

The instrument

Synestra

The cross-domain coordination architecture that closes the Economic Availability gap. Integrated directly into the infrastructure you already own. No rip-and-replace. No new hardware. No workload disruption.

Resource Intelligence

Maintains a real-time cross-domain campus model integrating BMS, SCADA, DCIM, EPMS, workload schedulers, and network monitoring. Continuously models the causal relationships between workload decisions and infrastructure outcomes.

↑ Stranded capacity · Coordination loss

Thermal Intelligence

Treats your chiller plant, cooling towers, CRAHs, and liquid cooling circuits as a single coordinated optimization surface. Anticipates workload-induced thermal demand — delivering pre-cooling signals before the load arrives.

↑ Cooling inefficiency loss

Compute Intelligence

Gives your ML engineering teams real-time visibility into the facility-level infrastructure cost of their scheduling decisions. Eliminates the 4× energy penalty that scheduling in isolation creates — without touching their workloads.

↑ Compute utilization · Scheduling loss

Operations Management

Establishes your EA baseline during commissioning — before the first production workload runs. Characterizes your specific campus infrastructure under controlled load conditions, identifying every recoverable opportunity before revenue operations begin.

↑ All five loss categories from day one

Lifecycle Analytics

Accumulates facility-specific behavioral data with every operating day, refining the campus model continuously. The operating intelligence becomes more precise over time — and it belongs to your facility, not to a vendor.

↑ Compounding, non-transferable advantage

Edge Data Intelligence

OEM-agnostic sensor ingestion across all infrastructure types. Translates infrastructure metrics into asset-level EA scores, hall-by-hall NOI impact, and tenant-level economic contribution — legible simultaneously to operators, investors, and asset owners.

↑ Full-stack economic visibility

Recovered value · 1GW reference campus · 40–55% of EA gap closed

Conservative

$456M / yr

+3 ppt EA · 5.6 day payback

Base Case

$543M / yr

+4 ppt EA · 4.7 day payback

Optimistic

$630M / yr

+5 ppt EA · 4.1 day payback

The Founder

John Chavner

Founder & CEO · Synestra

I spent five years as the architect of the AI orchestration platform at VoltaGrid, on the generation side of the fence watching the data center side. That vantage point built Synestra. Before VoltaGrid, large industrial projects with GE Oil and Gas, and data center consolidation and cloud migrations for Fortune 100 companies at Capgemini. Both ends of the stack. A front-row seat to everything connecting them.

What I observed: the physical infrastructure between power generation and compute — cooling systems, power distribution inside the hall, fiber networks, GPU clusters — has never been intelligently connected in real time. Every system world-class at its own job. None of them sharing causal intelligence with the others. The root cause of most failures is in the gap between systems, not inside any single one of them.

Nobody has built the platform to close it. Synestra closes it. Not threshold alerts. Not rule-based automation. A platform designed from the ground up for industrial telemetry at hyperscale. The architecture is complete. I am assembling the founding team. If you are building hyperscale AI infrastructure or investing in this space, I want to hear from you.

Synestra is being built with one founding operator. If the gap described here is real in your facility, this conversation is about whether that's you.

Request a Conversation Review the Research

A 30-minute call establishes your EA baseline. No obligation.

Your dashboards are green. Keep reading.