Within Cloud Oversight
Can frontier AI training hide from clouds?
Cloud oversight could miss risky frontier work if developers split training runs, use private clusters, or move to less regulated jurisdictions.
On this page
- Splitting workloads and blurring thresholds
- Private clusters and state backed infrastructure
- Why international coordination matters for enforcement
Page outline Jump by section
Introduction
Cloud monitoring is often presented as one of the most practical ways to oversee frontier AI development. The logic is straightforward: if training the most capable AI systems requires enormous amounts of specialised computing power, then major cloud providers may be able to detect, record, and sometimes restrict the largest training runs. For people concerned about AI doom or loss-of-control scenarios, this visibility could provide an early warning system before highly capable models are deployed.
However, cloud monitoring only works if the relevant activity remains visible to regulated providers. Critics and policy researchers have therefore focused on a key question: could developers deliberately or unintentionally move important frontier AI work outside the scope of cloud oversight? The answer is probably yes to some degree, although there is substantial disagreement about how difficult large-scale evasion would actually be. The debate centres not on whether evasion is possible in principle, but on how costly, detectable, and scalable it would be. [cdn.governance.ai]cdn.governance.aiTH E INTERMEDIARY ROLE OF COMPUTE PROVIDERSTHE INTERMEDIARY ROLE OF COMPUTE PROVIDERS…March 26, 2024 — by L Heim · 2024 · Cited by 19 — In the current ecosystem, frontier AI mod…
Can frontier AI training hide from clouds?
Cloud governance proposals rely on a practical reality: modern frontier models are usually trained on clusters containing thousands or tens of thousands of advanced AI accelerators housed in major data centres. Because these resources are concentrated among a relatively small number of providers, cloud companies can become a governance “choke point” where large training efforts are visible. [cdn.governance.ai]cdn.governance.aiTH E INTERMEDIARY ROLE OF COMPUTE PROVIDERSTHE INTERMEDIARY ROLE OF COMPUTE PROVIDERS…March 26, 2024 — by L Heim · 2024 · Cited by 19 — In the current ecosystem, frontier AI mod… [2robots.ox.ac.uk]robots.ox.ac.ukHeim et al. 2024 Governing Through the Cloud The Intermediary RoleINTERMEDIARY ROLE OF COMPUTE PROVIDERS…26 Mar 2024 — We argue that compute providers should have legal obligations and ethical respons…
The central concern is that developers may adapt once monitoring exists. If regulations trigger reporting requirements above specified compute thresholds, organisations could have incentives to restructure their activities to avoid crossing those thresholds in an obvious way. Researchers increasingly describe this as an adversarial problem: regulators create monitoring systems, while developers may seek ways around them. [Institute for Law & AI]law-ai.orgInstitute for Law & AIThe Role of Compute Thresholds for AI GovernanceFebruary 20, 2025 — This article discusses the role of training compute thresholds, which use training compute to determine which potenti… [RAND]rand.orgRRA3686 1RAND CorporationStrategies and Detection Gaps in a Game-Theoretic Model…16 Jun 2025 — The authors outline strategies for cloud service…
Importantly, most discussions focus on future possibilities rather than documented cases of frontier AI developers successfully concealing existentially significant training runs. The evidence base remains largely prospective. Analysts are examining vulnerabilities before comprehensive cloud-monitoring systems are widely implemented. [RAND Corporation]rand.orgRRA3686 1RAND CorporationStrategies and Detection Gaps in a Game-Theoretic Model…16 Jun 2025 — The authors outline strategies for cloud service…
Splitting workloads and blurring thresholds
One of the most frequently discussed evasion risks is sometimes called compute structuring. The basic idea is similar to financial “structuring” used to avoid reporting thresholds: instead of conducting one easily identifiable large training run, an organisation divides work into smaller components that individually appear less significant. [AAAI Publications]ojs.aaai.orgCompute structuring, a technique where AI developers split or modify compute workloads for the purpose of avoiding regulation, poses a ch…
Several broad possibilities are often discussed:
- Breaking training into multiple stages spread across separate clusters.
- Distributing workloads across multiple providers.
- Reusing previously trained models and focusing on smaller incremental improvements.
- Combining many apparently ordinary training runs into a larger overall development effort.
The concern is not necessarily that such approaches would completely hide frontier development. Rather, they might reduce the effectiveness of monitoring systems that rely heavily on simple numerical thresholds. If regulation only watches for runs above a particular compute level, developers may have incentives to remain just below it. [RAND Corporation]rand.orgRRA3686 1RAND CorporationStrategies and Detection Gaps in a Game-Theoretic Model…16 Jun 2025 — The authors outline strategies for cloud service… 2arXiv
At the same time, researchers argue that the problem may be less severe than it initially appears. Recent work on compute structuring suggests that unusual patterns of activity could themselves become detectable. Large-scale AI development tends to leave operational traces through hardware acquisition, networking requirements, storage demands, engineering coordination, and repeated use of substantial computing resources. Evasion may therefore increase costs and complexity rather than making activity completely invisible. [AAAI Publications]ojs.aaai.orgCompute structuring, a technique where AI developers split or modify compute workloads for the purpose of avoiding regulation, poses a ch…
A related concern involves algorithmic progress. If future techniques achieve the same capabilities using substantially less compute, systems that once required heavily monitored infrastructure could potentially be developed below existing thresholds. This possibility is frequently cited as a limitation of governance systems tied exclusively to compute measurements. [Institute for Law & AI]law-ai.orgInstitute for Law & AIThe Role of Compute Thresholds for AI GovernanceFebruary 20, 2025 — This article discusses the role of training compute thresholds, which use training compute to determine which potenti…
Private clusters and state-backed infrastructure
Perhaps the most obvious limitation of cloud monitoring is that not all computing occurs in commercial clouds.
Large technology companies increasingly build their own data centres and AI clusters. Governments and state-backed organisations may also operate specialised computing facilities outside commercial cloud platforms. If frontier AI development moves onto infrastructure owned and operated directly by the developer, cloud-based monitoring loses much of its visibility. [cdn.governance.ai]cdn.governance.aiTH E INTERMEDIARY ROLE OF COMPUTE PROVIDERSTHE INTERMEDIARY ROLE OF COMPUTE PROVIDERS…March 26, 2024 — by L Heim · 2024 · Cited by 19 — In the current ecosystem, frontier AI mod…
For AI doom advocates, this is one of the strongest objections to cloud-centred governance. A monitoring system that covers only commercial providers may miss some of the actors most capable of building advanced systems. State programmes, military organisations, and well-funded corporations may possess both the resources and strategic reasons to develop independent infrastructure. [arXiv]arxiv.orgarXiv Defending Compute Thresholds Against Legal LoopholesarXiv Defending Compute Thresholds Against Legal Loopholes
However, the counterargument is that constructing and operating cutting-edge AI clusters is expensive and difficult. Acquiring large quantities of advanced chips, power, cooling capacity, networking equipment, and engineering expertise creates additional opportunities for oversight. This is one reason why some governance proposals focus not only on cloud providers but also on semiconductor supply chains, hardware tracking, and compute governance more broadly. [ai-safety-atlas.com]ai-safety-atlas.comCompute GovernanceChapter 4Compute requirements directly constrain what AI systems can be built - even with cutting-edge algorithms and vast datasets, orga… [Default]Lawfareto govern ai we must govern computeDefaultTo Govern AI, We Must Govern Computeby L Heim · 2024 · Cited by 2 — Compute governance can support AI policy goals in multiple way…
The practical question is therefore not whether private infrastructure exists—it clearly does—but whether truly frontier-scale projects can remain hidden for long while acquiring the necessary hardware and support systems.
Why international coordination matters for enforcement
Even a highly effective national monitoring regime faces a geographical problem. Developers may relocate activities to jurisdictions with weaker oversight, fewer reporting requirements, or less willingness to cooperate internationally.
This concern appears repeatedly in compute-governance discussions. If one country imposes strict monitoring while others do not, developers may shift large training runs across borders. The result could be regulatory arbitrage, where governance affects location choices more than overall activity. [Default]Lawfareto govern ai we must govern computeDefaultTo Govern AI, We Must Govern Computeby L Heim · 2024 · Cited by 2 — Compute governance can support AI policy goals in multiple way…
For that reason, many proposals emphasise international coordination rather than purely domestic regulation. Suggested approaches include:
- Harmonised reporting thresholds across multiple countries.
- Shared standards for cloud-provider monitoring.
- Coordination on advanced-chip export controls.
- International verification and inspection mechanisms.
- Common definitions of frontier AI development activities.
Supporters argue that compute governance becomes more robust as participation expands. A developer can evade one jurisdiction more easily than a coordinated group of major computing powers. Critics respond that achieving durable international agreement on strategically important AI technologies may be politically difficult, especially amid competition between major states. [arXiv]arxiv.orgarXiv Defending Compute Thresholds Against Legal LoopholesarXiv Defending Compute Thresholds Against Legal Loopholes
This creates a familiar tension within AI doom debates. Many proposed safeguards become stronger with international cooperation, yet geopolitical rivalry may simultaneously make that cooperation harder to achieve.
How serious are the evasion risks?
The strongest critique of cloud monitoring is that determined actors may eventually find ways around it. Private infrastructure, distributed workloads, algorithmic efficiency gains, and international relocation all represent plausible routes through which important AI development could become less visible. [arXiv]arxiv.orgarXiv Defending Compute Thresholds Against Legal LoopholesarXiv Defending Compute Thresholds Against Legal Loopholes
The strongest response is that perfect visibility is not the relevant standard. Monitoring systems are often valuable even when they are incomplete. Financial regulation does not eliminate money laundering, and nuclear safeguards do not guarantee perfect detection, yet both can substantially increase the difficulty and cost of prohibited activities. Cloud monitoring advocates argue that the same logic may apply to frontier AI development. [cdn.governance.ai]cdn.governance.aiTH E INTERMEDIARY ROLE OF COMPUTE PROVIDERSTHE INTERMEDIARY ROLE OF COMPUTE PROVIDERS…March 26, 2024 — by L Heim · 2024 · Cited by 19 — In the current ecosystem, frontier AI mod…
From an AI doom perspective, the key issue is whether monitoring can meaningfully improve the odds of detecting dangerous capability development before a loss-of-control scenario emerges. The debate is therefore less about achieving perfect enforcement and more about whether cloud oversight can shift frontier AI development from being largely opaque to at least partially observable. Evasion risks are one of the main reasons many researchers view cloud monitoring as only one component of a broader package that may also include chip governance, evaluations, incident reporting, and international coordination. [ai-safety-atlas.com]ai-safety-atlas.comCompute GovernanceChapter 4Compute requirements directly constrain what AI systems can be built - even with cutting-edge algorithms and vast datasets, orga… [Default]Lawfareto govern ai we must govern computeDefaultTo Govern AI, We Must Govern Computeby L Heim · 2024 · Cited by 2 — Compute governance can support AI policy goals in multiple way…
Amazon book picks
Further Reading
Books and field guides related to Can frontier AI training hide from clouds?. Use these as the next step if you want deeper reading beyond the article.
The Age of Surveillance Capitalism
Provides context on monitoring systems and information visibility.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: cdn.governance.ai
Title: TH E INTERMEDIARY ROLE OF COMPUTE PROVIDERS
Link: https://cdn.governance.ai/Governing-Through-the-Cloud_The-Intermediary-Role-of-Compute-Providers-in-AI-Regulation.pdfSource snippet
THE INTERMEDIARY ROLE OF COMPUTE PROVIDERS...March 26, 2024 — by L Heim · 2024 · Cited by 19 — In the current ecosystem, frontier AI mod...
Published: March 26, 2024
-
Source: robots.ox.ac.uk
Title: Heim et al. 2024 Governing Through the Cloud The Intermediary Role
Link: https://www.robots.ox.ac.uk/~mosb/public/pdf/3343/Heim%20et%20al.%20-%202024%20-%20Governing%20Through%20the%20Cloud%20The%20Intermediary%20Role.pdfSource snippet
INTERMEDIARY ROLE OF COMPUTE PROVIDERS...26 Mar 2024 — We argue that compute providers should have legal obligations and ethical respons...
-
Source: ai-safety-atlas.com
Title: Compute Governance
Link: https://ai-safety-atlas.com/chapters/v1/governance/compute-governanceSource snippet
Chapter 4Compute requirements directly constrain what AI systems can be built - even with cutting-edge algorithms and vast datasets, orga...
-
Source: law-ai.org
Title: Institute for Law & AIThe Role of Compute Thresholds for AI Governance
Link: https://law-ai.org/the-role-of-compute-thresholds-for-ai-governance/Source snippet
February 20, 2025 — This article discusses the role of training compute thresholds, which use training compute to determine which potenti...
Published: February 20, 2025
-
Source: rand.org
Title: RRA3686 1
Link: https://www.rand.org/pubs/research_reports/RRA3686-1.htmlSource snippet
RAND CorporationStrategies and Detection Gaps in a Game-Theoretic Model...16 Jun 2025 — The authors outline strategies for cloud service...
-
Source: ojs.aaai.org
Link: https://ojs.aaai.org/index.php/AAAI/article/view/41127/45088Source snippet
Compute structuring, a technique where AI developers split or modify compute workloads for the purpose of avoiding regulation, poses a ch...
-
Source: arxiv.org
Title: arXiv Defending Compute Thresholds Against Legal [Loopholes]({{ ‘loopholes/’ | relative_url }})
Link: https://arxiv.org/abs/2502.00003 -
Source: arxiv.org
Link: https://arxiv.org/html/2604.04712v1Source snippet
arXivHardware-Level Governance of AI Compute: A Feasibility...6 Apr 2026 — The governance of frontier AI increasingly relies on controll...
-
Source: arxiv.org
Link: https://arxiv.org/abs/2507.06379Source snippet
arXivDomestic frontier AI regulation, an IAEA for AI, an NPT for AI, and a US-led Allied Public-Private Partnership for AI: Four institut...
-
Source: Lawfare
Title: to govern ai we must govern compute
Link: https://www.lawfaremedia.org/article/to-govern-ai-we-must-govern-computeSource snippet
DefaultTo Govern AI, We Must Govern Computeby L Heim · 2024 · Cited by 2 — Compute governance can support AI policy goals in multiple way...
-
Source: blog.heim.xyz
Title: xyz Compute and the Governance of AI
Link: https://blog.heim.xyz/compute-and-the-governance-of-ai-talk/Source snippet
and the Governance of AI - Talk5 Nov 2023 — This talk explores the role of computational resources, or "compute", as a node for AI govern...
Additional References
-
Source: frontier.co.uk
Link: https://www.frontier.co.uk/gamesSource snippet
Our GamesOur games each offer a unique experience, from space exploration of our very own galaxy, to intricate park management or intense...
-
Source: youtube.com
Title: AI safety guardrails stripped fast & Chrome brings AI on-device
Link: https://www.youtube.com/watch?v=Su3dw_29EagSource snippet
Evade AI monitoring compute governance cloud Agent security bypasses in practice & Governance gaps for enterprise agents - AI News (Apr 2...
Published: May 27, 2026
-
Source: flyfrontier.com
Link: https://www.flyfrontier.com/Source snippet
Frontier Airlines: Low Fares Done RightAs Home of Low Fares Done Right, find great deals and cheap flights to destinations all over North...
-
Source: GOV.UK
Title: a pro innovation approach to ai regulation government response
Link: https://www.gov.uk/government/consultations/ai-regulation-a-pro-innovation-approach-policy-proposals/outcome/a-pro-innovation-approach-to-ai-regulation-government-responseSource snippet
pro-innovation approach to AI regulation: government...6 Feb 2024 — We have already started developing the central function to support e...
-
Source: law-ai.org
Link: https://law-ai.org/wp-content/uploads/2024/11/The-Role-of-Compute-Thresholds-for-AI-Governance.pdfSource snippet
f harm under tort law, as a proxy for threat to national or public security in risk.Read more...
-
Source: aisi.gov.uk
Link: https://www.aisi.gov.uk/frontier-ai-trends-reportSource snippet
duct simplified versions of self-replication tasks and [AI research]({{ 'ai-research-loop/' | relative_url }}) and development...
-
Source: youtube.com
Title: Agent security bypasses in practice & Governance gaps for enterprise agents
Link: https://www.youtube.com/watch?v=OrWdNdRSJzUSource snippet
AI safety guardrails stripped fast & Chrome brings AI on-device - Tech News (May 27, 2026)...
Published: May 27, 2026
-
Source: youtube.com
Title: Enterprise AI Infrastructure Security Series
Link: https://www.youtube.com/watch?v=Wa0ULIyychsSource snippet
Agent security bypasses in practice & Governance gaps for enterprise agents - AI News (Apr 22, 2026)...
-
Source: ibanet.org
Title: dual oversight of cloud markets
Link: https://www.ibanet.org/dual-oversight-of-cloud-marketsSource snippet
the growing interplay...26 Mar 2026 — France is experiencing a sustained and increasing involvement of its national authorities in the r...
-
Source: oecd.ai
Title: risk thresholds consultation
Link: https://oecd.ai/en/site/ai-futures/discussions/risk-thresholds-consultationSource snippet
Public consultation on AI risk thresholds10 Sept 2024 — The EU AI Act requires providers of GPAI models that cross a compute threshold of...
Topic Tree