Enterprise self-hosted / data-residency execution planes forming across AI toolchains

1. Historical note (corrected in the 2026-09-27 review): OpenAI's ZDR reaffirmation + Private Safety Processing preview (2026-08-18, primary source, ev-20260818-05): the retention-free inference + safety-tooling commitment shape; the white paper was slated for September and is still unpublished as of this run
2. Anthropic's Enterprise Frontier Safeguards (2026-09-01, shipped with Fable 5.1, ev-20260901-01): ZDR-equivalent privacy with safeguards, customer-controlled cloud storage, phased from fall 2026 — a second vendor converging on the same shape
3. Cursor's self-hosted machines (2026-09-02, official changelog, ev-20260902-07): cloud-agent execution stays entirely on the customer's own network (codebases, build artifacts, secrets internal), dynamic machine pools, existing sandboxes supported (AWS Lambda, Coder, Cloudflare, Daytona, Modal, Namespace, Vercel, E2B) plus self-hosted computer use — advancing from data commitments to a bring-your-own execution plane
4. Adjacent signals (not counted): Netskope R141 Enterprise Browser BYOLLM / on-device open-weight execution (beta); arXiv 2609.01572, an enterprise self-hosted LLM recipe (200+ internal apps, absorbing 50% of platform traffic, 116M requests/month)
5. 5. The OpenAI Agents API execution-environment menu (2026-09-10, primary source, ev-20260910-02): cloud agents run on the OpenAI-hosted sandbox, the customer's own infrastructure, or one of nine sandbox partners (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel — including VPC deployments) — 'execution-plane choice' graduates from one vendor's product feature to a procurement option at the largest platform; adjacent: Mistral's EUR 3B Series D (9/8, ev-20260908-12) bets the European supply side on the same direction via 'sovereign open-weight AI'
6. The current official guide documents customer-side storage, protected automated review and setup. This supplies product documentation, not independent production-adoption evidence. (ev-20260818-05)
7. 2026-10-06: Apache 2.0 EmbeddingGemma 2 weights move text, image, video and audio embeddings onto edge devices.
8. 2026-10-07: Liquid d1 publishes multimodal single-pass decision models for edge devices; openTPU exposes an inspectable FPGA accelerator stack with a golden-model validation path.
9. 2026-10-21: OpenAI announces models and Codex for Google Cloud; Granite Speech 5 and RFD4-Proteina publish weights and code.

Review log

1. 2026-09-05 · Opened as candidate / Low: three vendors moving the same direction within three weeks — 8/18 (OpenAI ZDR/PSP), 9/1 (Anthropic EFS), 9/2 (Cursor self-hosted machines) — plus Netskope BYOLLM and the enterprise self-hosting paper as adjacent signals. Honest note: this is the weakest of the trend lines opened this run; if no second dev-tool vendor follows next cycle, or EFS/PSP slip, it should be downgraded or retired.
2. 2026-09-08 · Criterion audit (weekend window): no second dev-tool vendor shipped self-hosted agent execution; OpenAI's Private Safety Processing white paper is still unpublished (the OpenAI news index shows only two 9/6 posts after 9/3, neither PSP); Anthropic EFS remains a phased plan from fall 2026 with nothing landed. The self-set downgrade trigger fired — status moves candidate -> weakening; since September (the PSP white paper's slated window) has not ended, it is not retired yet — retire next cycle absent a second vendor or substantive EFS/PSP delivery.
3. 2026-09-12 · Downgrade trigger overturned; weakening revived to candidate (evidence item 5): the OpenAI Agents API makes the execution environment an explicit menu (own infrastructure and VPC deployments included) — a second vendor, and the largest platform, lands execution-plane choice; Cursor self-hosted machines was the first. PSP white paper still unpublished, Anthropic EFS not yet landed, independent adoption data still absent — confidence stays Low; if September ends without PSP/EFS landing and without adoption data, downgrade again.
4. 2026-09-12 · 02:08Z follow-up scan: no new delivery of OpenAI Private Safety Processing or Anthropic EFS, and no new self-hosted execution plane or independent adoption case. Stays candidate / Low.
5. 2026-09-13 · OpenAI's Private Safety Processing white paper remains unpublished (past mid-September) and Anthropic EFS has still not landed; no new self-hosted execution-plane form and no independent-org adoption case. Stays candidate / Low - downgrade if September ends without PSP/EFS delivery and adoption data (self-set criterion unchanged).
6. 2026-09-13 · 01:02Z follow-up scan: no new OpenAI PSP white paper, Anthropic EFS delivery, self-hosted execution plane or independent adoption data. Stays candidate / Low.
7. 2026-09-14 · The LiteLLM RC customer-managed KMS and gateway key are control-plane hardening, not a new self-hosted execution plane or independent adoption case. No new OpenAI PSP white paper or Anthropic EFS delivery. Stays candidate / Low.
8. 2026-09-15 · No new self-hosted execution plane or independent adoption case; the OpenAI PSP white paper and Anthropic EFS remain undelivered. The Bash-interface study reinforces the importance of strong isolation but is not execution-plane product evidence. Stays candidate / Low.
9. 2026-09-16 · No independent in-window evidence changed the confirmation criteria. Status and confidence stay unchanged.
10. 2026-09-17 · No new self-hosted execution plane or independent adoption case appeared, and OpenAI PSP plus Anthropic EFS remain unshipped. JustFit studies local inference capacity rather than an enterprise execution plane. Stays candidate / Low.
11. 2026-09-18 · No new self-hosted execution plane or independent production adoption case appeared; OpenAI PSP and Anthropic EFS remain unshipped. Stays candidate / Low.
12. 2026-09-19 · No new self-hosted execution plane or independent production adoption appeared; OpenAI PSP and Anthropic EFS remain undelivered. Stays candidate / Low.
13. 2026-09-20 · OpenVINO 2026.4 expands local Intel inference but is not an enterprise agent execution plane or independent production adoption case. OpenAI PSP and Anthropic EFS remain undelivered. Stays candidate / Low.
14. 2026-09-21 · Qwen-Image-2.1 provides local weights and broad runtime support, but its research-only license and 26-40 GiB serving footprint do not establish an enterprise self-hosted execution plane. Stays candidate / Low.
15. 2026-09-22 · No new self-hosted execution plane, independent production adoption, OpenAI PSP paper or Anthropic EFS delivery appeared. Stays candidate / Low.
16. 2026-09-23 · vLLM 0.30 expands self-hosted inference but is not a new enterprise agent execution plane or independent production-adoption case. The OpenAI PSP paper and Anthropic EFS delivery are still missing. Stays candidate / Low.
17. 2026-09-25 · No new enterprise self-hosted execution plane, independent production adoption, OpenAI PSP paper or Anthropic EFS delivery appeared. Stays candidate / Low.
18. 2026-09-26 · No new self-hosted execution plane, independent production adoption, OpenAI PSP paper or Anthropic EFS delivery appeared; Microsoft Foundry is a managed-platform expansion. Stays candidate / Low.
19. 2026-09-27 · Microsoft Copilot Managed Runtime can host execution inside the customer tenant but remains Microsoft-managed rather than customer-operated. It is adjacent data-residency evidence, not enough for an upgrade. Stays candidate / Low.
20. 2026-09-27 · Remains candidate with Low confidence. PSP now has concrete setup documentation, correcting the earlier blanket claim of non-delivery. Independent production adoption across organizations is still unproven.
21. 2026-09-27 · This scan found no new production evidence for customer-operated or data-resident execution. Stays candidate / Low.
22. 2026-09-28 · This scan found no new independent signal sufficient to change the assessment. Status and confidence remain unchanged.
23. 2026-09-29 · This scan found no new independent signal sufficient to change the assessment. Status and confidence remain unchanged.
24. 2026-10-02 · No new signals changed the status or confidence in this run.
25. 2026-10-03 · No new signals changed the status or confidence in this run.
26. 2026-10-04 · No new signals changed the status or confidence in this run.
27. 2026-10-05 · No new signal in this cycle was strong enough to change status or confidence.
28. 2026-10-06 · No new signal in this cycle was strong enough to change status or confidence.
29. 2026-10-07 · No new signal in this cycle was strong enough to change status or confidence.
30. 2026-10-08 · EmbeddingGemma 2, Liquid d1 and openTPU move multimodal retrieval, structured decisions and an inspectable FPGA stack onto local devices, but enterprise adoption and operations evidence remain limited. Stays candidate / Low.
31. 2026-10-09 · No new enterprise-adoption evidence in this run; status unchanged.
32. 2026-10-09 · No new enterprise-adoption evidence appeared; status unchanged.
33. 2026-10-10 · Agent Lightning runs on self-managed Kubernetes, but no new enterprise-adoption evidence appeared; status unchanged.
34. 2026-10-12 · Open WebUI and OpenClaw add self-hosted tooling surfaces, but no new enterprise-adoption evidence appeared; status unchanged.
35. 2026-10-12 · Littleguys adds a local multi-agent control surface, but no new enterprise deployment or data-residency adoption evidence appeared; status unchanged.
36. 2026-10-14 · No new enterprise self-hosting or data-residency adoption evidence appeared; status unchanged.
37. 2026-10-14 · NVIDIA open-sourced the full Run:ai platform, broadening self-hosted AI infrastructure; independent migration and production-adoption evidence remain absent, so status is unchanged.
38. 2026-10-16 · Muse Spark Open adds local and hosted deployment paths, but hardware and license barriers remain high; status unchanged.
39. 2026-10-16 · Beam, Nemotron 3.5 and Granite Switch 4.1 add downloadable weights and local paths, further supporting the candidate trend.
40. 2026-10-17 · No new signal in this run was strong enough to change status.
41. 2026-10-18 · No new signal in this run was strong enough to change status.
42. 2026-10-20 · TorchNeuron previews a native PyTorch-to-Trainium path, but source was not public at the cutoff, so it remains only a candidate signal; candidate / Low is unchanged.
43. 2026-10-21 · Voxtral 3 adds regional APIs and a self-deployment option, while DeepInfer exposes production inference internals but lacks a complete serving layer; candidate / Low is unchanged.
44. 2026-10-21 · OpenAI plans Google Cloud access while Granite Speech 5 and RFD4-Proteina ship public artifacts; candidate / Low is unchanged.
45. 2026-10-23 · No new signal in this run changed status or confidence.
46. 2026-10-24 · No signal in this run changes status or confidence.
47. 2026-10-24 · No new rating-changing signal in this run.
48. 2026-10-25 · No new evidence in this run changed the trend stage.
49. 2026-10-27 · No new evidence in this run changed the trend stage.
50. 2026-10-27 · No new independent signal in this run changes status or confidence.
51. 2026-10-28 · Run:ai memory enforcement improves quota control for shared execution planes, but it is a single-vendor update and does not raise the trend.
52. 2026-10-30 · GitHub-managed Copilot sandboxes add an execution-plane signal, but independent isolation and cost data are missing; status is unchanged.
53. 2026-10-31 · NemoClaw adds an owned agent-stack blueprint, but evidence remains vendor-supplied without cost or isolation tests; it remains a candidate.
54. 2026-10-31 · No new signal in this run changes the current trend judgment.
55. 2026-11-02 · No new signal in this run changes the current trend judgment.
56. 2026-11-03 · No new signal in this run changes the current trend judgment.