Sovereign AI

AI Infrastructure Handover and BOTT Technology Transfer: Runbooks, Access, Software Rights, Operating Skills and Acceptance Evidence

An AI infrastructure handover is complete when five areas close on evidence the receiving team has performed itself:

  1. Documentation: runbooks the receiving team executes unassisted.
  2. Operating skills: live incidents resolved by the receiving team.
  3. Software rights: licences and build materials that outlast the operator.
  4. Access cutover: operator accounts, credentials and sessions revoked.
  5. Acceptance package: per-area sign-offs and a post-cutover access recertification.

The order matters because revoking the operator's access cannot be undone while the receiving team still depends on the operator. The receiving team defines the sign-off test for each area before the transfer phase.

Before Step 1: Name an Owner per Area and Start the Access Inventory

The transfer phase begins from a known start state: the facility and GPU cluster are built, stabilized and operating, and the local team is in training. We sell and deliver the build-operate-transfer model behind this handover, which runs in four phases:

  1. Build: design and deliver the AI data center and GPU environment.
  2. Operate: stabilize and run the facility and cluster.
  3. Train: develop local technicians, engineers and operating teams.
  4. Transfer: move operational control to the sovereign, telecom operator or enterprise taking ownership.

Two preparations precede closeout work:

  1. Owner per area: the receiving organization names one owner for each of the five areas, so each area keeps an owner when responsibility changes hands.
  2. Access inventory: both sides hold a written access record of shared credentials, personal API tokens, third-party management tools, and physical keys or codes before the first revocation.

Step 1: Close Documentation Against the Live Environment

The receiving team proves the documentation by executing it. Our transfer commitments cover enablement sessions, documentation including program governance and standard operating procedures, and support onboarding that trains the local team to take over the assets and operations. Our capability set includes operating models, procedures, runbooks and training, commissioning and performance-readiness plans, and monitoring and telemetry architecture. Engagements divide these deliverables differently, so the contract's deliverable list sets the scope.

One vendor lists ten data categories for closing out physical infrastructure assets. Five of them apply to a GPU site:

  1. Commissioning test results: functional performance outcomes linked to the verified system.
  2. Warranty terms and expiration dates: coverage periods, conditions and claim contacts.
  3. O&M manuals: operating procedures and maintenance guidance tied to each asset.
  4. Outstanding punch list: unresolved defect items tracked until closed and verified.
  5. Spare parts lists: recommended stocking levels and part numbers.

Six checks close documentation:

  1. Runbook accuracy: Runbooks match the live environment. A runbook that names servers, software versions or procedures that no longer exist is a defect, closed by reconciling the document against the running cluster.
  2. Unassisted execution: The receiving team executes each runbook unassisted. One engineer from the receiving side performs the procedure end to end under operator observation, and a written record captures the result.
  3. SOPs and governance documents: Standard operating procedures and program governance documents are delivered and versioned, with dates and owners.
  4. Commissioning results and punch list: Commissioning results are in hand and the punch list is closed. Pre-handover acceptance tests and burn-in feed this evidence package directly; unresolved defect items carry an owner and a date.
  5. Warranty records: Warranty terms and expiry dates are recorded per asset. Coverage periods, conditions and claim contacts are listed against serial numbers.
  6. Dependency maps: Network and power dependency maps are current. A poor digital asset handover leaves the maintenance crew unable to tell which breaker feeds which rack; the map is verified by tracing at least one feed per room to its labeled source.

Step 2: Prove Operating Skills on Live Incidents

Live incidents test operating skills, with the receiving team running the operation under operator observation. This proof precedes access cutover, because once the operator's accounts are disabled, the receiving team can no longer fall back on operator intervention.

Skills pass on four checks:

  1. Named-role attendance: Attendance records tie each enablement session to the engineers and technicians who will hold the on-call rotation.
  2. Live incident handling: The receiving team resolves at least one scheduled maintenance action and at least one unplanned incident before cutover, with the ticket trail as evidence.
  3. Hands-off operator: The operator observes without intervening. The observation rule is the point of the test; an operator who steps in has converted an evidence run into a support call.
  4. External escalation path: The receiving team knows which vendor, contractor or internal group handles each failure class once the operator is gone.

Step 3: Secure Software Rights and Rebuild Ability

The rights test asks whether the receiving organization can rebuild and maintain its software after the operator's exit. Escrow is the working model for third-party and operator-owned components. An escrow deposit holds the full source code, the build instructions and scripts, and the libraries, frameworks and dependencies; its purpose is rebuild, maintenance and support.

Common release events are supplier insolvency or administration, failure to provide support or maintenance, material breach of contract, and product end-of-life without a viable alternative. Release gives the right to use the materials; the receiving organization must then secure and maintain the code, support it internally or through a third party, and transition users and data.

Five checks settle software rights:

  1. Software inventory: The inventory separates third-party, operator-owned and customer-owned software. Each item carries its licence path into the post-transfer period.
  2. Licence survival: Each licence survives the operator's exit. Terms that terminate on change of control or on the end of the operating agreement are flagged and renegotiated before transfer.
  3. Build materials: Source and build materials are held for critical items. For operator-owned software, the deposit holds the code, build scripts and dependencies needed for a rebuild.
  4. Deposit verification: The deposit is verified against the live environment. Verification exists because deposits drift: missing dependencies and build environments that cannot be recreated are the common failures, and a verification build from deposit materials closes the item.
  5. Release events: Release events are defined in the contract. The common events above, or the contract's equivalents, are written down before signature.

Where the customer owns the software outright, escrow does not apply; the rights transfer and the build materials are the deliverable, and the proof test reduces to items one through three.

Step 4: Cut Over Access Last, Because Revocation Cannot Be Undone While the Operator Is Still Needed

This is the step most likely to fail quietly. It goes last because the operator is still needed for the earlier tests, and revocation ends that option.

The control set comes from personnel-termination practice. NIST SP 800-53 Revision 5 control PS-4 requires an organization to disable system access within a defined period, terminate or revoke authenticators and credentials, retrieve security-related property, and retain access to the systems the departing party controlled. Applied to a departing operator, the sequence is:

  1. Inventory first. The access inventory started in the preparation step is finished before cutover, so disabling an account cannot lock operator staff out of a system still needed for the handoff.
  2. Disable operator accounts at cutover. Identity-provider accounts for operator staff are disabled within the period the transfer agreement defines.
  3. Revoke authenticators and credentials. The receiving team terminates the certificates, keys and hardware tokens issued to operator staff.
  4. Rotate shared and service credentials. Passwords, API keys and service-account secrets known to operator staff are changed, because disabling a person's account does not change a secret they know.
  5. Reset sessions and personal tokens. Axipro's offboarding checklist warns that disabling an identity-provider account does not automatically revoke active session cookies, OAuth refresh tokens or personal access tokens. Behaviour varies by platform, so sessions are reset explicitly at cutover.
  6. Retrieve physical property. The retrieval list from the same control covers hardware authentication tokens, system administration manuals, keys, identification cards and building passes.
  7. Confirm the receiving team holds control. A receiving-team member completes one login per critical system and one physical walk of the raised floor and entry points.

Step 5: Assemble the Acceptance Package and Recheck After 30 to 60 Days

The acceptance package is the five areas' evidence, bound together: executed runbooks with results, incident tickets closed by the receiving team, the verified software-rights inventory, the completed cutover record, and the signed per-area tests. One data center commissioning guide advises against starting integrated systems testing until signed paperwork shows each component finished and tested on its own.

Acceptance follows the same gate. Each of the five areas signs off on its own evidence first; acceptance of the transfer is the last signature, and the receiving team withholds it while one area remains open.

The handover does not end at signature. Axipro's offboarding checklist sets a post-departure access certification 30 to 60 days after departure, repeated at 90 days for sensitive roles. Here the certification checks for a resurfaced operator account, token or key. The recheck is the acceptance package's closing page: a dated record that the access state tested at cutover still holds weeks after the operator's departure.

For a full transfer of operational control, Zettabyte's sovereign AI data center program delivers the build-operate-transfer path.