Cloud Infrastructure as a Regulatory Layer for AI?

Presentation to ITREV Forum

04 April, 2025

​

Ilan Strauss with Tim O’Reilly

����

Website: https://www.ssrc.org/programs/ai-disclosures-project/�Substack: https://asimovaddendum.substack.com/

AI Disclosures Project

Risks / Vulnerabilities in an AI product

• Are the risks just at the model level (the 15%)?
• What about the system ( ~ 85%)?

​

Observability (Health): Is with the deployed application not the model

• Observability is with the infrastructure
• Observability & moderation is mostly algorithmic (LLM!)

​

​

​

​

Our deepening cloud dependence…

​

…Now: AI’s “elastic” brain���

​

​

​

​

​

​

“If you believe companies will build applications from scratch on top of the infrastructure services if the right selection [of services] existed, and we believed they would if the right selection existed, then the operating system becomes the internet” – Andy Jassy

Who holds the power in AI markets?

Power: External observability (data) + rules (standard setting)

​

​

Gatekeeper Power: Infrastructure & platform

​

• Who caught DeepSeek’s exfiltration of OpenAI’s model?
• Who decides age limits for Character.ai?
• Who enforces AI-content labeling?
• Who sets today’s global web security protocols? �

Mastercard & Visa: de facto regulators

​

Who knows the answer to this question:

​

What are the conditions of acceptable vampire sex [for online pornographic videos]?

​

– Financial Times (2022)

Hypothesis

To interrogate, debate, and ultimately reject if need be:��

The “Cloud”, as an essential internet infrastructure, can provide a foundational regulatory layer for the AI stack, enforcing agreed upon standards and operational controls.

​

​

​

​

​

Can standards exist on the cloud? (Tentative)

Monitoring & Standards already exist

• Laws (CSAM)
• ToS / AuP (Responsible AI Policy)

​

Enforce and extend

• Enforce ToS through enhanced automated moderation
• Know your customer (KYC)
• Minimum guardrails on models deployed via AWS Bedrock
• Applying automated scoring to new customer code

​

​

​

​

​

​

We would love to hear from you – feedback

If you work with cloud & LLM system integration, LLM applications and orchestration, IT controls, and more, we would love to hear from you.

Pulling on a string

End

LEFTOVERS

Gatekeepers and AI controls

• Policy, practice, and standards often rely on key gatekeepers (deployment and dissemination).

​

• Amidst competing standards and policies, the binding rules are those who own the operating system you run on.

​

• Control is greater at the cloud level: greater usage, more controls, higher standards generally.

​

​

​

Shared Responsibility for Controls

“Shared responsibility” for security & compliance:

• “of” (Amazon) vs “in” (customer)

​

​

​

​

​

How to regulate software run on internet networks?

To interrogate, debate, and ultimately invalidate if need be��

When infrastructure is code (DevOps)

​

Which code layers must become the law? (Lessig)

​

​

​

​

​

​

​

Regulating essential AI infrastructure?

Some facts

• 22% of recently announced cloud implementations had an AI element (ML & LLM)
• Amazon, Google, and Microsoft’s clouds > 60% combined market share
• “Elastic brain” is key to AI

​

“Structural power” (Berjon)

• Control data and set standards
• Private algorithmic management

����

​

​

​

​

​

​

But risks in “weaponizing” infrastructure

Example: Visa & Mastercard with Russian sanctions

• Push people towards alternatives

​

Privacy is core to cloud adoption: can we monitor usage?

• Privacy preserving: Algorithmic management (filters)

​

What standards? Which standards? Whose standards?

​

System / “stack”

Source: Menlo Ventures, 2024: https://menlovc.com/perspective/the-modern-ai-stack-design-principles-for-the-future-of-enterprise-ai-architectures/

Motivation: Did OpenAI train on non-public data?�

�

This violates Robots.txt (protocol)���Who enforces the internet’s ‘rules’?