Case study · AI infrastructure and regulation

Know your compute customer: building export-control compliance into a Malaysian GPU cloud

A Malaysian data centre operator had added a GPU cloud to serve regional AI start-ups, enterprises and research labs. Demand was strong. So was scrutiny: Malaysia had introduced permits for moving US-origin AI chips, and in November 2026 the US moved to require licences for advanced chips going to Malaysia and four other countries, including to stop restricted parties reaching them remotely. The operator needed to prove, continuously, who was using its GPUs and for what, without becoming impossible to buy from.

Case studyCloud infrastructure · MalaysiaSumanta Mukhopadhyay9 min read
#governance#platform#kubernetes
CASE STUDY · CASE STUDY · AI INFRASTRUCTURE AND REGULATION17 Nov 2026 · 15:50 GST
At a glance
SectorCloud infrastructure · Malaysia
My roleFractional CTO and platform architect
EngagementAssessment, architecture and delivery leadership
ScopeCustomer onboarding and screening, ownership checks, policy-gated GPU admission, usage monitoring, access logging, evidence and regulator engagement
What changed
  • No GPU capacity without completed customer screening and ownership checks
  • GPU jobs admitted only when the tenant's compliance status allows it
  • A traceable record of who used which GPUs, when and from where
  • Unusual usage patterns surfaced for review rather than discovered later
  • Evidence ready for licence applications, audits and supplier questions
Onboarding with ownership checksSanctions and restricted-party screeningKubernetes with Kueue for GPU admissionTenant policies tied to screening statusIdentity-aware access loggingUsage anomaly detectionEvidence store for auditsHardware asset register

Anonymised composite. Client names, identifying details and confidential figures are withheld; public facts are cited in the references.

The situation

A Malaysian data centre operator added a GPU cloud to its colocation business, offering training and inference capacity to start-ups, enterprises and university labs across South-East Asia. Customers signed up online, were given a Kubernetes namespace and GPU quota, and could start work within a day. Onboarding checked payment details and little else.

Malaysia had already tightened its own controls: since July 2025, exporting, transshipping or moving US-origin high-performance AI chips through the country requires a permit under the Strategic Trade Act34. The operator's hardware stayed in its own halls, so that rule mostly affected its suppliers. The bigger question was who was using the hardware remotely.

Why now. In November 2026, the US Commerce Department's Bureau of Industry and Security issued a rule requiring licences for advanced AI chips going to Malaysia, Indonesia, Singapore, Thailand and Brazil, and aimed at stopping restricted parties from reaching controlled chips through data centres in third countries12. For a GPU cloud in Malaysia, that made its customer list part of its supply chain: future hardware purchases would depend on being able to show where compute went.

What had to be true

  • Known customers: every tenant screened, including who ultimately owns or controls it.
  • Gated access: no GPU capacity until checks are complete, and access paused if they lapse.
  • Known users: every person and service using a tenant's GPUs identifiable.
  • Traceable usage: who used which GPUs, when and from where, kept as evidence.
  • Watchful: unusual patterns surfaced for review.
  • Still easy to buy: legitimate customers onboard in days, not months.

Architecture

Customers
Start-upsEnterprisesResearch labs
Onboarding
Screeningrestricted parties and sanctionsOwnership checkswho controls the customerUse declarationwhat the compute is for
Platform
Policy-gated admissionquota follows compliance statusTenant namespacesisolated GPU pools
Access
Identity for every usernamed, MFAAccess logssource, time, workload
Evidence
Usage monitoringanomaly reviewAsset registerwhich chip, whereEvidence storeaudits and licences
Compliance status flows from onboarding into the scheduler: a tenant's checks decide whether its GPU jobs run.

The platform already used Kueue to queue and admit GPU jobs fairly across tenants. That admission step became the enforcement point. Each tenant's queue was linked to its compliance status: if screening was incomplete, had expired or was under review, the tenant's jobs stayed suspended rather than running. Kueue's model of jobs waiting in a queue until admitted made this a configuration change rather than a new system5.

Kueue admission architecture: a user submits a Job through the Kubernetes API, the Kueue controller creates a Workload and evaluates LocalQueue, ClusterQueue, ResourceFlavor and Cohort, then either keeps the Job suspended or resumes it for the scheduler to place Pods
Jobs wait until admitted: the operator tied that admission decision to each tenant's compliance status. Source: Kubernetes blog, Utkarsh Raj, René Dudfield and Kevin Hannon · CC BY 4.0

The data flow, end to end

  1. ApplyA customer signs up, declares intended use and provides ownership details
  2. ScreenRestricted-party and ownership checks run; edge cases go to a person
  3. ProvisionA tenant is created with quota linked to its compliance status
  4. AdmitEach GPU job is admitted only if the tenant's status allows it
  5. RecordJobs and access are logged with identity, source and hardware used
  6. ReviewPeriodic rescreening and anomaly alerts keep status current
Checks happen once at the door and continuously afterwards, because customers change.

The hard part: remote access and changing customers

Who is really using the GPUs? A screened customer can resell capacity, share credentials or be acquired. The operator required named user identities with multi-factor sign-in, prohibited credential sharing and resale without approval in its terms, and logged where each session came from. Monitoring looked for patterns worth a human look: many users appearing under one small tenant, access from unexpected locations, or workloads that did not match the declared use.

Rescreening without friction. Ownership and restricted-party lists change. Rather than rescreen everyone at once, the platform rescreened tenants on a schedule and whenever lists were updated or a tenant's details changed, pausing only those with new matches for review. Legitimate customers rarely noticed.

Not a law firm. Export-control questions often need legal judgement. The platform's job was to collect facts and enforce decisions consistently; edge cases went to compliance staff and external counsel, and their decisions were recorded alongside the evidence.

For a GPU cloud, the customer list is now part of the hardware supply chain. Build compliance into the scheduler, not a spreadsheet, so it cannot be bypassed on a busy day.

How it ran

Platform
  1. Tenant status in queues
  2. Admission gating
  3. Access logging
Compliance
  1. Screening rules
  2. Edge-case decisions
  3. Rescreening schedule
Sales and support
  1. Clear onboarding steps
  2. Customer communication
  3. Resale and sharing terms
Leadership and legal
  1. Risk appetite
  2. Supplier and regulator engagement
  3. Licence applications
Sales had the biggest stake in keeping onboarding fast, so they helped design it.
Phase 1Know
  • Screen existing tenants
  • Ownership checks
  • Asset register
Phase 2Gate
  • Status-linked queues
  • Suspend on lapse
  • Named users and MFA
Phase 3Watch
  • Access logging
  • Anomaly review
  • Scheduled rescreening
Phase 4Evidence
  • Evidence store
  • Audit-ready reports
  • Licence support
Screening existing tenants came first; it found a few that needed closer review.

Key decisions and trade-offs

DecisionAlternativeWhy
Enforce in the schedulerManual checks before provisioningCannot be skipped under pressure
Suspend jobs, do not delete tenantsTerminate access on any doubtReview without losing customers' work
Named users with MFAShared tenant credentialsKnow who actually uses the GPUs
Event-driven rescreeningAnnual review onlyCatches changes when they happen
Humans decide edge casesFully automated decisionsLegal judgement stays with people
Evidence store from day oneAssemble evidence when askedSupports licences and audits quickly

What I'd do again, and differently

Again: put enforcement in the admission step of the scheduler; it made compliance consistent without slowing engineers. Differently: involve the hardware suppliers earlier, since their own export obligations shaped what evidence they needed from the operator. Related: the 17 November tech news, AI capacity contracts, the shared GPU case study and where I work.

If you run or buy GPU capacity in a market facing new export rules, see my AI architecture work, or analyse your challenge and bring the result to a call.

References
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5
FAQ

Common questions

What changed for GPU clouds in Malaysia in November 2026?

Reporting says the US Commerce Department issued a rule requiring licences for advanced AI chips going to Malaysia and four other countries and targeting remote access to controlled chips through data centres in third countries. This adds to Malaysia's 2025 requirement for permits to export, transship or transit US-origin high-performance AI chips.

What is know-your-customer for compute?

Screening and verifying who uses compute capacity, including who ultimately owns or controls the customer, and keeping that knowledge current as customers, users and workloads change.

How can a Kubernetes platform enforce compliance for GPU access?

By tying GPU quota and job admission to each tenant's compliance status, so jobs from tenants whose checks are incomplete or have lapsed wait or are refused, and by logging every job and access with the identity behind it.

Keep reading

Related

Open to advisory, fractional CTO & select builds

Let's talk about what you're building.

The fastest route is a free 30-minute call. Prefer to write? Send a message and I'll reply personally within two business days.

Start a conversation

How would you like to reach me?

Book a 30-min callFastest. Pick any open slot on my calendar.→ Write in GmailOpens a pre-filled draft in your browser.→ Use my email appOutlook, Apple Mail or your default client.→

I reply personally within two business days. Time zones from the Americas and Europe to the Middle East and Asia-Pacific.