The Power Quality Problem

Datacenters are terrible electrical loads.

Utility engineers know this. Equipment vendors know this. Insurance underwriters know this.

AI datacenters are worse.

What makes a "bad load"

Good industrial load (factory, chemical plant):

• Steady draw, predictable ramp
• High power factor (>0.95)
• Minimal harmonics
• Scheduled maintenance windows

Bad load (poorly designed datacenter):

• Unpredictable demand spikes (training job starts, 50MW surge in 10 seconds)
• Power factor issues from switching power supplies
• Harmonic distortion from rectifiers
• Regenerative braking from UPS systems during shutdowns
• Noise propagates upstream into distribution network

AI datacenter (worst case):

• Millisecond-scale load variance (GPU workload shifts, power draw swings 20%)
• Massive in-rush current when clusters wake from idle
• Poor coordination between compute load and cooling load
• No advance signal to utility when demand spike incoming

TBF

The switchgear problem

Standard datacenter switchgear is sized for peak load + 20% margin.

Problem: AI workloads don't have "peak load" — they have chaotic load.

• Training job starts: 80MW → 95MW in 30 seconds
• Inference batch clears: 95MW → 65MW in 10 seconds
• Cooling lags by 2 minutes (thermal mass delay)
• Switchgear sees 30MW swings every few minutes

What this does to equipment:

• Circuit breaker thermal cycling (reduces lifespan)
• Transformer core saturation during spikes
• Voltage sag ripples back through distribution network
• Neighboring customers see flicker

TBF

Demand spikes no one sees coming

Utility perspective:

You get 15-minute advance notice (standard SCADA interval) that a hyperscaler is ramping load.

By then:

• GPU cluster already at 90% load
• Cooling systems playing catch-up
• Voltage regulator hunting
• Substation transformer operating outside optimal range

You can't dispatch reserves that fast.

So you keep spinning reserves online 24/7 just in case the datacenter decides to launch a training run.

Cost: ~$2M/year in reserve generation capacity per 100MW datacenter

TBF

Heat wave coordination problem

Summer afternoon, 102°F, grid at 98% capacity.

Utility needs to shed 50MW for 2 hours to avoid rolling blackouts.

Option 1: Call the datacenter, ask them to curtail load

• Problem: They don't know which workloads can pause without breaking
• They shed 10MW (non-critical batch jobs), but 40MW is "critical inference"
• Grid still overloaded

Option 2: Datacenter has onsite generators (20MW diesel backup)

• Problem: Utility doesn't know datacenter has generators available
• Datacenter doesn't know utility needs load relief
• Generators sit idle while grid fails

Both sides have capacity. Neither can coordinate.

TBF

Noise backstream (the problem no one admits)

When GPU workloads shift rapidly:

• Power draw oscillates at millisecond scale
• Datacenter UPS and rectifiers inject harmonic noise
• Noise propagates upstream through distribution transformer
• Adjacent customers on same feeder see voltage distortion
• Utility gets complaints about "flickering lights" from industrial customers 2 miles away

Root cause: Datacenter scheduler has no idea GPU workload is about to shift, so power systems can't pre-adjust.

Utility solution: Install expensive harmonic filters, oversize transformers

Actual solution: If the scheduler knew workload was shifting 10 seconds in advance, power systems could ramp smoothly

TBF

Don't you wish your datacenter was a better load?

What if your 100MW AI datacenter could:

• Signal load changes 60 seconds in advance (scheduler knows job is queued)
• Coordinate with utility telemetry (receive peak demand alerts, shed non-critical workloads)
• Dispatch onsite generation on your signal (you send "heat wave imminent" → datacenter shifts to generators for 2 hours)
• Smooth power factor (workload placement optimized for electrical characteristics, not just compute availability)

This isn't science fiction. It's logistics.

The scheduler knows what's queued. The orchestrator knows what's running. The telemetry knows which CUs are about to wake up.

No one connects this to the electrical system.

TBF

Want to know how?

We'll send you a 2-page white paper:

• How ACGEOS telemetry creates advance load signals
• How we integrate utility SCADA feeds for peak demand coordination
• How we identify shed opportunities (non-critical workloads that can pause)
• How we dispatch onsite generation on your signal (heat waves, grid emergencies)

If you're reading this page, you already know the problem.