Conceptual North Carolina AI inference deployment connecting users, research data, computing resources, network paths and power.
Data Center News & Insights

AI Inference in North Carolina: A Buyer’s Siting Guide

The right location for AI inference in North Carolina starts with the application, not a campus announcement. Before reserving capacity, map the users, the data they need and the service you expect to accept.

This guide helps business leaders and their technical reviewers compare Triangle and Charlotte candidates, examine network and power commitments, and build a deployment record that separates operating resources from future promises.

Direct Answer

What should determine a North Carolina inference location?

AI inference uses a trained model to produce outputs from new inputs. Choose a North Carolina deployment location by testing the application’s user paths, data controls and performance requirements against delivered compute, connectivity and power. A proposed campus is not the same evidence as an accepted operating service.

Define the application and the data owner

Write the business task before requesting a location recommendation. “An employee assistant that retrieves approved documents” gives reviewers something to evaluate. “We need local AI” leaves the important decisions open.

NVIDIA’s explanation of AI inference provides background on inference and deployment tradeoffs. Your procurement brief should translate those concepts into requirements for one application:

  • Application and users: Identify who submits requests, from which offices or devices, and during which operating periods.
  • Data ownership: Name the person responsible for approving access to prompts, retrieval sources, logs and outputs.
  • Model requirements: Specify the model and version, expected input and output sizes, and any accelerator or memory requirements identified by the technical team.
  • Acceptance criteria: Define acceptable response time, output quality, concurrent demand, failure behavior and cost boundaries before evaluating bids.

Distinguish location preferences from actual controls. If data must remain within an approved environment, document where processing, storage, backups and administrative access are permitted. Use the sovereign AI infrastructure guide to frame those control questions; a North Carolina address alone should not close that review.

Also identify which data the application must retrieve during a request. The data gravity and AI guide addresses that adjacent placement problem. Ask the application owner to map the complete retrieval path, not just the model endpoint.

Triangle and Charlotte: different questions, not automatic winners

Treat the Triangle and Charlotte as candidate starting points for different application maps—not as predetermined first and second choices.

For a hypothetical research workflow involving teams and approved datasets in Raleigh or Durham, begin with those data locations, access controls and collaboration needs. Then ask whether a Triangle candidate can deliver the required service. Percepture’s Raleigh–Durham data center guide supports the adjacent facility-shortlisting task; it is not a benchmark of inference performance.

For an enterprise application serving Charlotte offices, begin with the users, identity services, databases and other application dependencies involved in each request. The Charlotte data center guide provides facility context. Require a separate answer about the compute, connectivity and operating responsibilities included in any inference proposal.

A customer-facing application serving users across the state needs a broader test. Include representative user locations rather than letting the headquarters address decide the deployment. If neither metro satisfies the application’s acceptance criteria, reconsider the architecture or candidate set instead of relaxing the requirements to preserve a preferred location.

A facility shortlist is not an inference service shortlist

Keep these buying decisions separate. A colocation proposal may cover space, power and connectivity while leaving the customer responsible for hardware, model serving and application operations. An inference service proposal should state which of those responsibilities the supplier accepts.

Ask every bidder to mark each responsibility as included, excluded or separately contracted. If you are buying a facility environment for your own equipment, evaluate that environment and your operating plan. If you are buying a managed model endpoint, evaluate the delivered service and its dependencies. Do not assume one proposal includes the other.

Inspect the physical path and turn-up responsibility

Do not accept “connected” as a complete network answer. Ask the operator to show how the application reaches the users, data and services in your brief.

Hunter Newby’s published discussion of building neutrality provides context for evaluating interconnection choice. For an actual order, request current operator-specific documentation rather than treating a general framework as facility inventory.

The useful questions are concrete:

  • Which carrier has a physical demarcation in the building, and where is it?
  • Which connection requires a local loop or another separately ordered service?
  • What cross-connect path, ports and equipment does this order require?
  • Who places each order, pays each charge and owns escalation?
  • What written delivery commitment applies to each dependency?
  • What evidence supports any claim that primary and backup paths are diverse?

Treat internet exchange access and physical carrier presence as separate questions. Request the demarcation inventory, connection path and written terms for the service being proposed; an exchange-access description is not a substitute for that evidence.

For a cloud-dependent application, identify the actual cloud on-ramp connectivity being ordered, including its provider, location and handoff. Do not stop at a general statement that a building offers cloud access.

Then test from representative endpoints. The low-latency networking guide for AI is a useful companion to that work. Record application response time alongside network measurements so reviewers do not mistake one part of the request path for the complete user experience.

Read the power commitment as a contract

Ask what power is available to the proposed deployment, under which agreement, and on what acceptance date. Keep that answer separate from campus expansion plans or utility announcements.

In its October 7, 2026 announcement, Duke Energy described a proposed North Carolina large-load settlement as subject to regulatory approval. Treat that announcement as a dated description of a proposal—not a substitute for the tariff, regulatory order and service agreement governing your project.

Have the operator and your qualified reviewers identify:

  • The serving utility, applicable tariff and contracting parties.
  • The power committed to your specific service or equipment footprint.
  • Whether delivery depends on upgrades, construction, approvals or another customer’s schedule.
  • The proposed equipment’s electrical and cooling requirements, and who confirms compatibility.
  • Deposit, minimum-payment, delay, termination and curtailment provisions that apply to the order.
  • The documents and tests required before you accept the service.

Avoid carrying a campus-level power figure into a workload budget without an allocation and delivery explanation. Ask the seller to connect its proposal to your order, including any conditions that remain open.

If deployment depends on new construction, put permitting and community responsibilities on the same schedule as technical delivery. Percepture’s data center community engagement guide addresses the stakeholder work; keep responsibility for approvals and engineering acceptance with the relevant specialists.

Benchmark the workload and the fallback

Compare architectures before committing to a metro. Use the on-device, edge and cloud inference comparison to frame the placement choices, then test the options that fit your application.

Consider a hypothetical enterprise evaluating a document assistant for employees in Charlotte and the Triangle. Its decision team could compare a managed cloud endpoint, a locally operated edge deployment and company-owned equipment in colocation. This is a proposed evaluation exercise, not a report of measured results.

Keep the model and version, retrieval dataset, request mix, output-quality criteria and test conditions consistent. Record differences when an option cannot support the same configuration rather than presenting unlike services as a clean comparison.

Measure what the user experiences: time to the first useful response, completion time, failed requests and performance under the expected concurrent load. Include both ordinary and demanding requests. Record the test window, configuration and operating conditions so another reviewer can understand what was compared.

Price the same scope. Request an itemized estimate covering the relevant compute or equipment, storage, connectivity, support, operating labor and fallback resources. State which costs are quoted, estimated or excluded. Do not turn a hardware price or endpoint rate into a complete operating-cost comparison.

Make failure behavior part of acceptance

Ask what happens when a connection, model endpoint or retrieval dependency is unavailable. Define whether the application should queue work, switch to an approved alternative, provide a limited response or stop safely.

Test the selected behavior under an agreed plan. Include recovery and return to normal operation, not just the initial switch. Assign someone to approve the result and record any restrictions on where fallback processing may occur.

A fast demonstration should not close a separate data-control review. Require both the application owner’s acceptance and the relevant security or data owner’s sign-off before treating a candidate as ready.

Build a deployment fact record

Use the proposed North Carolina Workload-and-Commitment Register to connect the business decision to the evidence behind it. It is an editorial worksheet for organizing review, not an engineering certification or a validated scoring model.

Create one register for each candidate deployment. Keep supporting documents alongside it, and record the source, date, responsible owner and next action for every unresolved dependency.

North Carolina Workload-and-Commitment Register

Illustrative worksheet only: all entries in the final column describe a hypothetical enterprise document assistant, not an operating deployment.
FieldEvidence to recordHypothetical entry
Application and data ownerBusiness task, accountable application owner and data-access approver.Employee document assistant; operations lead owns the application; IT data owner approves retrieval access.
User endpointsRepresentative locations, devices and application dependencies.Charlotte and Triangle offices; endpoint test results not observed.
Carrier pathDemarcations, local-loop dependencies, cross-connect order and escalation owner.Candidate connection path unknown; request operator documentation.
Delivered compute and powerOrdered compute configuration, allocated power, cooling review and delivery evidence.Configuration and allocation unknown; no delivery acceptance recorded.
Contractual dependencyApplicable service terms, external dependencies and responsibility for delay.Service agreement and utility dependencies unknown; procurement owns follow-up.
Dated acceptanceTest configuration, results, acceptance date and accountable approvers.No acceptance date; workload and fallback tests not observed.
Record controlSource document, document date, reviewer and next review trigger.Document references and review date to be entered when evidence is received.

Classify each resource separately. Operating means you have evidence of the delivered resource and its acceptance for the intended service. Contracted future supply belongs with its delivery conditions and dates. Announced plans remain planning context, not an accepted deployment.

A candidate may contain all three categories. For example, an operating building can still have a future compute order or an unresolved connection. Do not let one completed component stand in for the complete service.

Before signing, require a named owner and next action for every unknown that affects the launch decision. After acceptance, update the record when the model, data location, network path, power allocation or fallback design changes.

Keep AI-assisted shortlists separate from capacity evidence

If an AI answer introduces a supplier, save the question, answer, cited sources and date. Check its capacity and service statements against the supplier’s current documents before adding them to the register.

For teams using Prime AI Visibility, keep AI-answer measurement separate from infrastructure acceptance. The AI visibility evidence guide is a companion reference for checking AI-generated brand statements, not proof of operating capacity. Percepture has a relationship with Prime AI Visibility; this is not an independent product comparison.

Which unresolved commitment could change your decision?

Bring your completed register, candidate proposals and open questions to a buyer-education conversation with Percepture. Keep engineering acceptance, utility interpretation and legal review with your qualified specialists.

Book a Strategy Conversation

About Bob Generale

Bob Generale, President of Percepture

Bob Generale is President of Percepture. His work connects public relations, SEO, generative engine optimization and integrated digital marketing with the questions business buyers need answered. He writes about technical access, useful content, credible evidence and relevant human follow-up. His editorial approach starts with the buyer’s decision and assigns an owner to the work that follows.