54east / Insights / Perspectives

Perspectives · Whitepaper

Own the model. Keep the audit.

Sovereign cloud puts upstream AI in-country. It does not, by itself, give a national oil company the weights, the fine-tune, or a trail a qualified reserves evaluator can reconstruct.

54east · Abu Dhabi · Updated 2026-08-28 · Open access. Cited. No invented case studies.

Sovereign cloud puts upstream AI in-country. It does not, by itself, give a national oil company the weights, the fine-tune, or a trail a qualified reserves evaluator can reconstruct.

Winning institutions own the system. For a national oil company, that is not an argument against Azure. Abu Dhabi’s public architecture already runs upstream AI in-country, on sovereign cloud, with hyperscaler models in the stack. The remaining question is narrower, and harder: after the data stays here, who owns the tuned model, who can take it with them, and can a qualified reserves evaluator reconstruct the estimate years later?

Jurisdiction is largely settled. Ownership is not.

The architecture that is already public

ADNOC did not keep seismic AI off the cloud. It put it in the country, on a named stack.

ENERGYai — launched at ADIPEC in November 2024, then contracted in March 2025 — is the flagship. AIQ, the energy-AI company ADNOC founded with G42 and now majority-owned by Presight, signed a US$340 million three-year contract to roll ENERGYai across ADNOC’s upstream assets: more than 28 producing fields, seismic analysis and geological modelling included. WAM names the stack in one sentence: Azure, the OSDU framework, OpenAI models, built with G42 and Microsoft.12 AIQ’s own product page is explicit on residency: built in-country with UAE resources; data sovereignty is a named benefit, not a footnote.3

The drilling layer is the same pattern, not an air-gap. On 4 August 2026 SLB published the Real-Time Operations Center: developed in the UAE, hosted in ADNOC’s sovereign cloud, more than 120 onshore and offshore rigs, critical operational data and workflows kept in-country. Not disconnected. In-country, and live.4

That is the public architecture. A paper that tells an ADNOC-scale operator to reject Microsoft-class models, or to treat air-gapped seismic AI as the default, is arguing with the operator’s own press. Do not.

The useful argument starts after that stack is in place.

Upstream is where the money is. That is why ownership matters.

McKinsey’s August 2026 upstream analysis — Bill Ambrose, Giorgio Bresciani, Spandan, Priyank Singh — puts the annual value pool at about US$65 billion near-term on today’s technology, US$125 billion as proven tools spread, and US$230 billion at full potential. The concentration is production optimisation, drilling, and reservoir management. The top ten use cases drive nearly half the identified value. McKinsey also notes that less than 20 percent of deployments track the AI-generated value and share the KPI internally. The prize is large. Capture is not automatic.5

The data underneath that prize is commercially and strategically sensitive. Seismic surveys, drilling logs, reservoir models, and production history encode decades of exploration spend and a detailed picture of undeveloped resource. Aggregated across a national operator’s asset base, a leak informs a competitor’s view of the subsurface. That is enough. It is not a direct input to OPEC+ quota-setting, and it is not what rating agencies use. OPEC+ is moving to annual production-capacity audits — consultants, 2026 assessments for 2027 baselines — not to operator seismic cubes.6 Published reserves and fiscal oil dependence are the figures that travel. Keep those two facts separate.

The implication is still sharp. If the value is in production, drilling, and reservoir, the model tuned on this operator’s fields is an asset. Licensing the interface is not the same as owning that asset.

What sovereign cloud settles — and what it does not

Sovereign cloud answers location and, depending on the controls, a large part of legal process. ENERGYai and the SLB RTOC both say: the data stays in the UAE. Core42, G42’s sovereign-cloud unit, is selling the same move at national scale — off private, often air-gapped environments and onto sovereign-enabled public cloud: hyperscale Azure plus local controls.78 That is the UAE’s published direction. Treat it as the floor, not the fight.

Three things the floor does not settle:

Fine-tune rights. A model tuned on an operator’s reservoir and production history encodes that operator’s geology. The vendor’s base model is theirs. The improvements from the operator’s data are a negotiation. If the contract is silent, the operator has licensed an interface to its own subsurface. Specify, at signature, that fine-tunes, adapters, evaluation sets, and derived weights trained on operator data are the operator’s.

Exportability. “In-country” is not “portable.” A tuned model that cannot leave the vendor’s platform — no weights, no prompts, no runbooks, no evaluation harness — is a residency claim with a lock. The test is operational: can the operator’s team take the instance to another in-country host, or to their own, without the original vendor in the room? If the answer is a professional-services statement of work, it is not owned.

QRE reconstructability. This is the argument that survives a board and a reserves committee. SPE’s Petroleum Resources Management System (2018) requires that supporting data, analytical processes, and assumptions be documented in sufficient detail for a qualified reserves evaluator or qualified reserves auditor to understand the basis for estimation, categorisation, and classification.9 SPE’s auditing standards keep the QRE accountable for the estimate; volumetric, performance, simulation, and analogy methods are reviewed, not rubber-stamped.10 SEC Regulation S-X Rule 4-10 treats “reliable technology” as including computational methods that are field-tested, consistent, and repeatable — it does not name machine learning, and it does not waive reconstructability.11

If AI begins to sit anywhere near booked reserves or production forecasts that a QRE must stand behind, a non-exportable vendor model is in tension with those rules. McKinsey also puts additional balance-sheet value at about US$35 billion from better exploration success. That figure sits outside the US$65 / US$125 / US$230 billion operating pool.5 The claim is not that AI already writes the national accounts. The claim is that a method the operator cannot reproduce is a method the QRE cannot defend.

A contractual promise that the vendor “will not use the data to improve other customers’ models” is not the same as an architecture in which they cannot, and it is not the same as the operator holding the tuned weights. Read the difference before signature.

Isolation is not air-gap. Hybrid is not a slogan.

Some OT and classified environments require network isolation. That is real, and it is not this paper’s subject. Air-gapped AI on seismic is not a published ADNOC or Aramco requirement, and it is not the UAE’s sovereign-cloud default. Core42’s IDC-backed paper states the contrast in the open: traditional sovereign cloud meant private, often air-gapped environments; sovereign-enabled public cloud is the other approach.7 Conflating SCADA isolation with air-gapped reservoir AI is a vendor talking point. If an operator’s risk posture requires air-gap, that is a custom constraint. Write it as one.

Hybrid infrastructure, where it is documented, is also not “buy commodity chat, build an isolated seismic stack.” Petrobras runs on-premises HPC for seismic and reservoir, bursts to AWS ParallelCluster for peak simulation, and hosts ChatPetrobras on Azure OpenAI — a named split of compute, not a manifesto against hyperscalers.1213 ADNOC’s published split is sovereign cloud plus Azure, OSDU, and OpenAI, with SLB’s DrillOps inside that cloud. Name those. Do not invent an industry pattern of isolated, operator-owned seismic AI sitting beside a bought chatbot.

Usage training is not model ownership

ADNOC’s talent programmes are public, and they are serious. The AI Accelerator has trained more than 40,000 employees. ADNOC Technical Academy is graduating UAE Nationals with applied AI and robotics in the curriculum — 403 in November 2025, more than 6,000 over the academy’s life — including Neuron 5 in the course.1415 ICV, nationalised by MoIAT in 2021, and the Nafis partnership (13,500 further private-sector jobs in ADNOC’s supply chain by 2028, AI among the named sectors, not ADNOC headcount) are how the operator measures in-country value, not only barrels.1617

That is usage and workforce nationalisation at scale. It is not, by itself, the operator’s reservoir engineers holding the construction of the model that sits on their fields. The ENERGYai platform remains a JV construct — Presight majority, ADNOC 49 percent, Microsoft and G42 in the build.118 Both things can be true: the staff know how to use the system, and the deepest understanding of how it was tuned still sits with the platform partners. Nafis and ICV make that fork a procurement choice, not a culture speech. Ask whether the engagement produces people who can modify, evaluate, and eventually replace the model, or people who can operate the interface.

Three questions before the contract is signed

A board does not need a lecture on cloud regions. It needs three answers in the SOW.

  1. Fine-tune. Who owns weights, adapters, and evaluation sets trained on our fields? Can we continue to train if the vendor relationship ends?
  2. Export. Can we take prompts, logs, runbooks, and the tuned instance to another in-country host? In what format, on what notice, without the original vendor’s professional-services team?
  3. Audit. If a QRE has to reconstruct a forecast or a reserves-adjacent estimate that used this system, what artefacts do they get — data, method, versioned model, human override — and do those artefacts survive in our jurisdiction, in our custody?

If any of the three is a slide instead of a clause, the institution has rented a capable system. It has not owned it.

Commodity predictive-maintenance patterns that do not encode proprietary reservoir specifics can sit on a shared platform. Reservoir modelling, field-tuned seismic interpretation, and production forecasting that a QRE might one day have to explain cannot. That split is a contract and an artefact list. It is not a reason to stay off Azure.

The stop

The value case for AI on upstream data is not in dispute. The operators moving at ADNOC’s scale have already chosen in-country hyperscaler stacks. Fighting that choice is a way to lose the room.

The decision that is still in front of a national oil company is the one after the region is signed: own the fine-tune, own the export path, own a trail a reserves evaluator can read. Winning institutions own the system. A sovereign cloud with someone else’s weights is a region with a local logo on the slide.

Sources

  1. WAM, “AIQ announces $340 million contract for large-scale deployment of agentic AI across ADNOC operations,” 10 March 2025. https://www.wam.ae/en/article/bilg57c-aiq-announces-340-million-contract-for-large-scale
  2. PR Newswire, “AIQ Announces $340 Million Contract for Large-Scale Deployment of Agentic AI Across ADNOC Operations,” 12 March 2025. https://www.prnewswire.com/news-releases/aiq-announces-340-million-contract-for-large-scale-deployment-of-agentic-ai-across-adnoc-operations-302400274.html
  3. AIQ, ENERGYai product page. https://aiq.ae/products/AgenticAI/energyai
  4. SLB, “ADNOC Deploys SLB Technology Across its Rig Fleet Through AI-Enabled Operations Center,” 4 August 2026. https://www.slb.com/newsroom/updates/2026/2026-0804-slb-adnoc-rtoc
  5. Oil & Gas Journal, reporting McKinsey & Co. analysis released ~25 August 2026 (Bill Ambrose, Giorgio Bresciani, Spandan, Priyank Singh), “Analysis: AI could unlock $230 billion in annual upstream oil and gas value.” https://www.ogj.com/general-interest/news/55400981/analysis-ai-could-unlock-230-billion-in-annual-upstream-oil-and-gas-value
  6. Reuters, “OPEC+ members to undergo annual oil capacity audit under new plan, sources say,” 2 December 2025. https://www.reuters.com/business/energy/opec-members-undergo-annual-oil-capacity-audit-under-new-plan-sources-say-2025-12-02/
  7. Core42 / Microsoft / IDC, “Balancing Innovation and Compliance in the AI Era.” https://www.core42.ai/resources/whitepapers/balancing-innovation-and-compliance-in-the-ai-era
  8. Microsoft, “Microsoft and Core42 present comprehensive whitepaper on the critical role of sovereign public clouds in the AI era,” 27 May 2025. https://news.microsoft.com/en-xm/2025/05/27/microsoft-and-core42-present-comprehensive-whitepaper-on-the-critical-role-of-sovereign-public-clouds-in-the-ai-era/
  9. Society of Petroleum Engineers, Petroleum Resources Management System (2018), §1.2.0.12. https://www.spe.org/media/filerpublic/0c/83/0c835db9-501f-4ce7-97f1-a1d6bb4e3331/prmgmtsystemv103.pdf
  10. SPE, “Standards Pertaining to the Estimating and Auditing of Oil and Gas Reserves Information,” June 2019. https://spe.org/industry/docs/ReservesAuditStandardsJune%202019Final.pdf
  11. 17 C.F.R. § 210.4-10 (SEC Regulation S-X, Rule 4-10), “Financial accounting and reporting for oil and gas producing activities.” https://www.ecfr.gov/current/title-17/chapter-II/part-210/section-210.4-10
  12. AWS HPC Blog, “Petrobras optimizes cost and capacity of HPC applications with Amazon EC2 Spot Instances.” https://aws.amazon.com/blogs/hpc/petrobras-optimizes-cost-and-capacity-of-hpc-applications-with-amazon-ec2-spot-instances/
  13. Microsoft Customer Stories, “Petrobras continues 70 years of innovation with Azure OpenAI Service.” https://www.microsoft.com/en/customers/story/1758695276753608190-petrobras-azure-openai-service-energy-en-brazil
  14. ADNOC, “AI for People.” https://www.adnoc.ae/en/artificial-intelligence/ai-for-people
  15. ADNOC, “ADNOC Technical Academy Celebrates Graduation of 403 UAE Nationals Equipped with AI and Robotics Skills,” 17 November 2025. https://adnoc.ae/en/news-and-media/press-releases/2025/adnoc-technical-academy-celebrates-graduation-of-403-uae-nationals
  16. ADNOC, “In-Country Value Creation.” https://adnoc.ae/en/empowering-people-and-driving-prosperity/creating-in-country-value
  17. Abu Dhabi Media Office, “ADNOC partners with Nafis to create 13,500 private sector jobs for UAE Nationals by 2028.” https://www.mediaoffice.abudhabi/en/energy/adnoc-partners-with-nafis-to-create-13500-private-sector-jobs-for-uae-nationals-by-2028/
  18. ADNOC, “ADNOC, G42 and Presight Partner to Accelerate AI Solutions for the Energy Sector,” 1 May 2024 (Presight 51% / ADNOC 49%). https://www.adnoc.ae/en/news-and-media/press-releases/2023/dnoc-g42-and-presight-partner-to-accelerate-ai-solutions-for-the-energy-sector

54east builds in-country. The institution that signed owns the instance: weights, prompts, logs, runbooks — not a licence to someone else’s platform. If the stack is already Azure-plus-sovereign-cloud, the briefing is the three questions above, written into the contract, then Train so the operator’s own engineers can hold the model. We leave. If those questions have not been asked yet, start there.

Related whitepapers

Why a cloud region isn’t a jurisdictionA region tells you where the disks spin. It does not tell you who can compel the operator, who holds the keys, who sees the logs, or what you walk away with if the contract ends.Build, buy, or waitA decision framework for UAE government entities: build it, buy it, or wait — with a reason, and a date to revisit.

Start with one decision.

If this paper describes the problem in front of you, the next step is a briefing with the architects.

Request a briefing

Start with
one decision.

Talk to the architects