Abstract
ABSTRACT
When an AI workload moves between GPUs, the new worker can restore a perfect copy of the state, and the old worker can still be alive. A stale inference worker may send a queued output chunk after it has been declared failed. A restored training group may replay results that were already committed. Copying a checkpoint or acquiring accelerator capacity does not decide which worker may publish the next externally visible result. Who owns the next AI compute effect?
Based on Indian Provisional Filing: 202631136555, filed 15 November 2026, and disclosed with the Mothership (PCT/IB2026/055615) on 16 July 2026.
This patent-pending technical concept proposes a complementary mechanism for transferring the right to commit consequential effects across accelerator workload migration, inference-worker replacement, distributed training recovery, partition reassignment, and failure-driven promotion. Its core principle separates transferring computational state from transferring authority to produce the next effect.
The mechanism works in five steps:
- A persistent task identity survives while the source and destination workers change.
- An ownership generation (epoch) records which worker may commit covered effects.
- A source fence retires the old worker’s authority at the covered output points.
- The destination is activated only after its identity, checkpoint lineage, policy, and approved decision logic are verified.
- A commit watermark records progress that is already committed, so restored state cannot republish it.
The actual output or model-update boundary then checks the task identity, worker, current epoch, effect class, watermark, and next sequence number. For example, if chunks 1 to 120 are committed and the destination restores at chunk 118, a stale-source chunk 121 is rejected for its old epoch, chunks 119 and 120 are rejected as already committed, and the new worker’s chunk 121 is admitted.
The disclosure covers cryptographically authenticated cutover capabilities, non-cryptographic hardware fencing, and hybrid designs with bounded hot-path verification. Ordinary tensor math, KV-cache copying, and checkpoint transfer stay outside the governed path. If the outcome of a target-side write is uncertain, the system holds and reconciles rather than assuming it failed.
Prospective industry relevance (AI-generated; see notice below). The mapping shows where the final commit check could sit alongside documented products. Every placement is an engineering hypothesis, and none of these vendors is represented as incorporating this concept.
- NVIDIA Dynamo (disaggregated serving): the check sits at the response-gateway path that admits output chunks. It verifies the task, worker, current epoch, and sequence, and rejects chunks from retired workers or already-committed chunks. This matters because transferred KV-cache state is not output authority.
- NVIDIA GPU Operator and MIG: the check sits at a protected workload output gateway. Reassigning a partition or restarting a container must not revive the retired task’s right to commit a result.
- NVIDIA BlueField DPU and DOCA: the check filters worker, epoch, and effect class at covered traffic egress. It applies only if the actual DPU program can enforce it, and alternate routes are out of scope.
- AWS SageMaker HyperPod: the check sits at the application-owned checkpoint-commit or model-response gateway. When workers recover or GPU tasks are preempted, it rejects stale or duplicate model updates and inference results.
- Google Cloud AI Hypercomputer and GKE Inference Gateway: the check sits at a participating serving-gateway or storage-commit boundary. It keeps request continuity across inference replicas and rejects output from a superseded replica.
- Microsoft AKS GPU nodes and Azure CycleCloud: the check sits at a customer-controlled API response or protected result store. It separates scheduling by Kubernetes or Slurm from current authority to publish output or commit model state.
- AMD Instinct and ROCm: the check sits at the serving or model-state commit interface, not in kernel code. It rejects commits from restarted or repartitioned workers that lack current authority.
- Intel Gaudi 3, Red Hat OpenShift AI with KServe, Snowflake Snowpark Container Services, Ray Serve, and CoreWeave Kubernetes Service: each uses a controlled output or commit gateway after node, pod, or replica replacement. It rejects stale-worker output and duplicate model-update commits. Pod health alone is never treated as commit authority.
These are possible complementary integration points, not claims of deficiency in any named product. The practical test is to trace the real data and commit path, list the bypasses, confirm the epoch and watermark are available at the boundary, and benchmark fault recovery and p99 commit latency.
Limits. The mechanism covers only actual protected effect paths. It does not guarantee universal exactly-once delivery, does not prevent compromise of the enforcement point itself, and does not require remote validation of every accelerator instruction.
Where to read the full filing. Search “Sangam Das” on WIPO PATENTSCOPE, or search PCT/IB2026/055615 (pctib202655615). The full PCT file is also available in Technical Disclosure Commons: https://www.tdcommons.org/dpubs_series/12542/
Basis and notice. The underlying invention is human-originated and based on the Mothership application PCT/IB2026/055615, filed 4 June 2026 and publicly disclosed 16 July 2026. This abstract was prepared with AI-assisted tools, including the industry-relevance section, and may contain errors, omissions, or prospective interpretations. Readers must cross-check it against the original PCT filings and the companies’ own public documentation before drawing any conclusion. Company and product references are for explanatory, complementary architectural alignment only. They are not an endorsement, validation, or infringement statement, and imply no adoption, collaboration, licensing, or finding on patent validity or third-party rights. The original specifications and claims are authoritative.
Part 5 of 39. This is Part 5 of a 39-part technical disclosure series: 38 numbered embodiments plus one separately designated Embodiment 12A. Each part presents one embodiment, its problem, the proposed solution, practical implementation mechanisms, and prospective complementary industry relevance.
License: CC BY-NC 4.0.
KEYWORDS
AI Safety; AI Infrastructure Security; Accelerator Security; GPU Workload Migration; Distributed AI Training; Stateful Inference Serving; Checkpoint and Restore; Fencing Tokens; Epoch-Based Ownership; Split-Brain Prevention; Exactly-Once Delivery; Idempotent Commit; Disaggregated Serving; KV Cache; Multi-Instance GPU; DPU Security; Kubernetes GPU Scheduling; Fault-Tolerant Computing; Runtime Authorization; Consequence-Bound Authorization; Hardware-Rooted Authorization; Confidential Computing; Multi-Tenant Computing; Model Update Integrity; Agentic AI Security; AI Factory Security; High-Performance Computing
Creative Commons License

This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Das, Sangam Kumar, "AI Safety: The GPU Worker Was Replaced, Yet the Old One Still Published. Who Owns the Next AI Compute Effect? (Part 5 of 39)", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/13208