Back
GovAI - Accessing Controlled AI Chips via Infrastructure-as-a-Service
governmentCredibility Rating
4/5
High(4)High quality. Established institution or organization with editorial oversight and accountability.
Rating inherited from publication venue: Centre for the Governance of AI
A December 2023 policy comment submitted to the Bureau of Industry and Security (BIS-2022-0025), co-authored by GovAI Research Fellow Lennart Heim, advocating for IaaS-based compute governance as a complement or alternative to hardware export controls.
Metadata
Importance: 62/100policy briefanalysis
Summary
This policy comment by GovAI researchers argues that IaaS cloud computing offers superior regulatory tools compared to physical chip export controls for managing advanced AI compute access. The authors recommend establishing a Know-Your-Customer (KYC) framework for above-threshold compute usage to prevent adversarial access while preserving US technological leadership. Key advantages of IaaS controls include dynamic geographic restrictions, continuous end-user verification, and real-time monitoring of compute usage patterns.
Key Points
- •IaaS compute controls offer four key advantages over physical chip export controls: country-level access restrictions, end-user verification, end-use monitoring, and flexibility to adjust in real-time.
- •Authors recommend a KYC scheme for above-threshold compute usage, building on Executive Order 14110, to identify and verify high-risk frontier AI development.
- •Physical chip export controls lack post-sale oversight—once chips are exported, their use cannot be monitored or restricted, unlike cloud-based controls.
- •Targeted IaaS controls can minimize regulatory burden on legitimate AI developers while focusing oversight on high-risk, large-scale compute usage.
- •The framework aims to prevent adversarial use cases including military modernization, CCP domestic surveillance, and intelligence operations against US interests.
Cited by 1 page
| Page | Type | Quality |
|---|---|---|
| Compute Governance | Concept | 58.0 |
Cached Content Preview
HTTP 200Fetched Mar 15, 202646 KB
Accessing Controlled AI Chips via
Infrastructure-as-a-Service (IaaS): Implications for
Export Controls
Comment on BIS–2022–0025 (RIN 0694–AI94) — Question 1
December 15, 2023
Lennart Heim
Research Fellow
Centre for the Governance of AI
lennart.heim@governance.ai
Janet Egan
MPP Candidate
Harvard Kennedy School
janetegan@hks.harvard.edu
Executive Summary and Key Recommendations
The provision of computing resources via Infrastructure-as-a-Service (IaaS) – also referred to as
cloud computing or compute provision – offers greater flexibility and precision in regulatory
controls than export controls of physical chips. We recommend using this advantage to prevent
adversaries from abusing US capabilities to undermine US national security interests, including
military modernization, CCP domestic surveillance, and intelligence operations, while avoiding
negative effects on US technology leadership.
IaaS compute access is commonly associated with the development and deployment of frontier
AI capabilities, representing the largest risk to national security. Further, IaaS compute access
enables controls that can be targeted toward the most intensive compute use:
A. IaaS providers only enable access to point-in-time computing power, offering a flexible
tool for governance, as it can be restricted or shut off at any stage.
B. IaaS providers are capable to monitor and control how much compute is being used per
customer (based on standard billing practices).
C. End-user awareness can be enabled via our recommended Know-Your-Customer (KYC)
scheme, building upon the requirements mandated by Executive Order (EO) No 14110.
—
This comment is authored by Lennart Heim and Janet Egan. It represents the view of the authors, rather
than the views of their organizations. For questions about the submission, reach out to Lennart Heim
(lennart.heim@governance.ai).
-- 1 of 20 --
Accessing Controlled AI Chips via Infrastructure-as-a-Service: Implications for Export Controls
Type of Control Physical Chips IaaS/Cloud Compute
Country-level access
— Control over which
countries can access
the technology.
Yes. Control over initial export
destinations, but no direct
control over subsequent
redistribution.
Yes. Cloud providers can restrict
access based on geographic or
jurisdictional boundaries, provided
sufficient end user checks.
End user — Control
over who is using the
technology.
Limited. Control only extends
to the initial point of sale with
no oversight over subsequent
transfers.
Possible. Continuous verification of
users' identity and credentials can be
implemented via KYC scheme.
Access can be restricted if KYC is not
passed.
End use — Control
over how the
technology is being
used.
None at this stage. Once
chips are sold, their usage for
different applications cannot
be monitored or controlled.
Possible. IaaS providers have
visibility into the volume of compute
usage, which can indicate the scale,
and potential risk, of the AI project.
Flexibility to adjust
control — Abilit
... (truncated, 46 KB total)Resource ID:
5907b6750da90f60 | Stable ID: MjY5MTNlYj