BlogSite-owner offtake

AI Data Center Power Requirements: A Site Owner’s Guide

What GPU tenants check in a site’s power before they sign an offtake agreement, and how to document it so diligence moves quickly.

Picture a hypothetical landowner with a retired industrial load and an interconnection agreement still in force. A GPU tenant calls, sounds eager, and asks for a site package. The owner sends a one-page summary listing the contracted capacity and the substation name. The tenant replies with a long list of questions about firmness, ramp timing, redundancy, and who pays for demand peaks. Weeks pass while the owner tracks down answers.

Nothing about the site changed during those weeks. What changed was how much the tenant could trust the summary. Owners who understand AI data center power requirements ahead of time can answer those questions in the first package and keep the conversation moving.

1. Firm capacity versus interruptible capacity

The first thing a GPU tenant wants to know is how much of your capacity is firm. Training clusters and inference fleets run continuously, and an unplanned curtailment can interrupt jobs that took days to schedule. Tenants will price interruptible capacity very differently from firm service, and some workloads cannot use it at all.

Document the following for each block of capacity you are offering:

  • The service type stated in your utility agreement and tariff, quoted directly.
  • Any curtailment, load-shed, or demand-response obligations, with the notice periods the utility can invoke.
  • Historical curtailment events, if any, and how long they lasted.
  • Whether capacity can be converted from interruptible to firm, and what studies or upgrades that would require.

Ask the tenant which of their workloads can tolerate interruption. Some operators split a site between firm and flexible loads, which can widen the pool of interested buyers.

2. How ramp schedules work

Few tenants energize an entire site on day one. GPU deployments arrive in phases as hardware ships, racks are installed, and clusters pass acceptance testing. Your utility may also impose its own ramp, releasing capacity in steps tied to upgrades or study milestones.

The tenant needs to see both schedules side by side. Ask them for their expected deployment phases and the conditions that move each phase forward. Then show the utility’s release schedule, the dependencies behind each step, and any minimum-take or ratchet provisions that start billing before the tenant’s load arrives. A mismatch between these timelines is one of the most common reasons offtake talks stall.

3. Power quality and redundancy expectations

GPU servers are sensitive to sags, swells, harmonics, and frequency events. Large clusters can also create their own power quality problems, since synchronized training jobs swing load quickly. Tenants want evidence that the site can absorb those swings without tripping protection or upsetting the utility.

  • Utility reliability history for the serving feeder or substation.
  • Any power quality monitoring data or studies you hold.
  • The redundancy topology you can support, described in plain terms (single feed, dual feed, or alternate source).
  • Existing UPS and generator assets, their condition, and permitting limits on run hours.

Redundancy expectations vary by tenant. Some inference operators want full backup on every rack, while some training operators accept less in exchange for faster delivery. Ask early, and see the companion guide on disclosing backup power, UPS, and generator status for what to include.

4. Rack density drives the cooling conversation

Current GPU platforms pack far more power into each rack than traditional enterprise servers. At those densities, air cooling alone often stops being practical, and tenants plan for direct-to-chip liquid cooling or rear-door heat exchangers. That choice reaches back into your power plan, because pumps, coolant distribution units, and heat rejection equipment draw load of their own.

Ask the tenant which hardware generation they intend to deploy and which cooling approach they prefer. Then document your water availability, space for heat rejection equipment, and whether the building structure can carry heavier liquid-cooled racks. If you are unsure what a modern deployment looks like, reviewing current GPU compute configurations can help you frame the discussion.

5. Metering and who pays demand charges

Tenants want a clear settlement boundary. Where is the revenue meter, who owns the equipment downstream of it, and how is consumption allocated if more than one party draws from the same service? These answers decide who carries line losses, auxiliary loads, and the cost of any shared infrastructure.

Demand charges deserve special attention. GPU loads can spike during cluster bring-up or synchronized jobs, and a single peak can set charges for a long billing window. Explain your tariff structure, how demand is measured, whether ratchets apply, and how you propose to pass those costs through. The guide on utility demand charge pass-through covers the contract language tenants usually request.

6. Transformer and switchgear readiness

A site with ample interconnection capacity can still be months or years away from serving load if the transformers and switchgear are missing. Large power transformers remain a long-lead item, and tenants treat equipment readiness as a scheduling risk on par with utility studies.

  • An inventory of installed transformers and switchgear, with ownership and condition.
  • Equipment on order, with supplier confirmations and expected delivery windows.
  • Who owns the substation and who controls switching operations.
  • Outstanding protection or arc flash studies that must finish before energization.

For procurement detail, read the guide on disclosing transformer lead time.

7. Presenting power in an offtake brief

The goal of the power section in your brief is to let a tenant’s engineering team reach a preliminary view without a call. Lead with firm capacity and its source documents, follow with the ramp schedule, and then cover quality, redundancy, cooling constraints, metering, and equipment status. Flag open items honestly with an owner and an expected resolution date.

Attach source documents wherever possible: tariff excerpts, utility letters, study reports, and equipment confirmations. The full structure is laid out in the GPU offtake brief guide for site owners. For the broader path from energized land to a tenant-ready campus, see how a powered site becomes an AI factory site.

If you hold unused grid power or an active interconnection and want to explore GPU offtake with Pacific Intelligent Technologies, Inc., book a site review call. Bring your utility agreement, ramp schedule, and equipment inventory.

FAQ

Do GPU tenants ever accept interruptible power?

Some do for flexible workloads such as batch processing, but they price it differently. Most anchor tenants want a firm block as the foundation of the deal.

Should I commit to liquid cooling before finding a tenant?

Usually it is better to document what the site can support and let the tenant’s hardware choice drive the final design. Preserving space, water access, and structural capacity keeps options open.

For the space and placement questions, use our guide to data center land requirements to assess the buildable area.

What if my power numbers are still uncertain?

Say so in the brief and explain what will resolve the uncertainty and when. Tenants respond better to clearly labeled open items than to figures that change during diligence.

Continue on the mothership

This satellite stops at the playbook. Transactions, specs, and comparisons live on pacificmachines.com. If the next step is a human, book 30 minutes with Harper.

Book 30 min